mirror of
https://github.com/turnstonelabs/turnstone.git
synced 2026-08-13 15:32:24 -06:00
Compare commits
1 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
 Patrick Buckley
|
2d174cd71c |
@@ -1,5 +0,0 @@
|
||||
# Funding platforms for the GitHub "Sponsor" button.
|
||||
# https://docs.github.com/en/repositories/managing-your-repositorys-settings-and-features/customizing-your-repository/displaying-a-sponsor-button-in-your-repository
|
||||
|
||||
github: [eous]
|
||||
custom: ["https://paypal.me/eousphoros"]
|
||||
+13
-13
@@ -15,7 +15,7 @@ jobs:
|
||||
runs-on: ubuntu-latest
|
||||
steps:
|
||||
- uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
|
||||
- uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7
|
||||
- uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6
|
||||
with:
|
||||
python-version: "3.14"
|
||||
- run: pip install pre-commit
|
||||
@@ -26,7 +26,7 @@ jobs:
|
||||
runs-on: ubuntu-latest
|
||||
steps:
|
||||
- uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
|
||||
- uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7
|
||||
- uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6
|
||||
with:
|
||||
python-version: "3.14"
|
||||
- run: pip install mypy
|
||||
@@ -43,21 +43,21 @@ jobs:
|
||||
python-version: ["3.11", "3.12", "3.13"]
|
||||
steps:
|
||||
- uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
|
||||
- uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7
|
||||
- uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6
|
||||
with:
|
||||
python-version: ${{ matrix.python-version }}
|
||||
# Node is required by tests/test_renderer_js.py — without
|
||||
# explicit setup, that suite silently skips if the runner
|
||||
# image happens not to ship Node, masking regressions in
|
||||
# the browser-side renderer.
|
||||
- uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7
|
||||
- uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6
|
||||
with:
|
||||
node-version: "24"
|
||||
- run: pip install -e ".[test]"
|
||||
# -v lists each test id as it starts (pytest prints the nodeid at
|
||||
# logstart), so a hang names the culprit on the last line instead of
|
||||
# riding the job timeout with only a trail of "..." dots.
|
||||
- run: pytest tests/ -m "not live and not e2e_recovery" --cov=turnstone --cov-report=term-missing --cov-report=xml -v
|
||||
- run: pytest tests/ -m "not live" --cov=turnstone --cov-report=term-missing --cov-report=xml -v
|
||||
- uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7
|
||||
if: always()
|
||||
with:
|
||||
@@ -83,14 +83,14 @@ jobs:
|
||||
--health-retries=5
|
||||
steps:
|
||||
- uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
|
||||
- uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7
|
||||
- uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6
|
||||
with:
|
||||
python-version: "3.14"
|
||||
- uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7
|
||||
- uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6
|
||||
with:
|
||||
node-version: "24"
|
||||
- run: pip install -e ".[test]"
|
||||
- run: pytest tests/ -m "not live and not e2e_recovery" --storage-backend=postgresql -v
|
||||
- run: pytest tests/ -m "not live" --storage-backend=postgresql -v
|
||||
env:
|
||||
TURNSTONE_TEST_PG_URL: postgresql+psycopg://postgres:postgres@localhost:5432/turnstone_test
|
||||
|
||||
@@ -98,7 +98,7 @@ jobs:
|
||||
runs-on: ubuntu-latest
|
||||
steps:
|
||||
- uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
|
||||
- uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7
|
||||
- uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6
|
||||
with:
|
||||
python-version: "3.14"
|
||||
- run: pip install build
|
||||
@@ -152,7 +152,7 @@ jobs:
|
||||
runs-on: ubuntu-latest
|
||||
steps:
|
||||
- uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
|
||||
- uses: astral-sh/setup-uv@11f9893b081a58869d3b5fccaea48c9e9e46f990 # v8.3.2
|
||||
- uses: astral-sh/setup-uv@fac544c07dec837d0ccb6301d7b5580bf5edae39 # v8.2.0
|
||||
with:
|
||||
uv-version: "0.9.18"
|
||||
- run: uv lock --check
|
||||
@@ -161,10 +161,10 @@ jobs:
|
||||
runs-on: ubuntu-latest
|
||||
steps:
|
||||
- uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
|
||||
- uses: astral-sh/setup-uv@11f9893b081a58869d3b5fccaea48c9e9e46f990 # v8.3.2
|
||||
- uses: astral-sh/setup-uv@fac544c07dec837d0ccb6301d7b5580bf5edae39 # v8.2.0
|
||||
with:
|
||||
uv-version: "0.9.18"
|
||||
- uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7
|
||||
- uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6
|
||||
with:
|
||||
python-version: "3.14"
|
||||
- run: uv sync --frozen --all-extras
|
||||
@@ -189,7 +189,7 @@ jobs:
|
||||
working-directory: sdk/typescript
|
||||
steps:
|
||||
- uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
|
||||
- uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7
|
||||
- uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6
|
||||
with:
|
||||
node-version: "24"
|
||||
- run: npm ci
|
||||
|
||||
@@ -0,0 +1,46 @@
|
||||
name: Claude Code Review
|
||||
|
||||
on:
|
||||
pull_request:
|
||||
types: [opened, synchronize, ready_for_review, reopened]
|
||||
# Optional: Only run on specific file changes
|
||||
# paths:
|
||||
# - "src/**/*.ts"
|
||||
# - "src/**/*.tsx"
|
||||
# - "src/**/*.js"
|
||||
# - "src/**/*.jsx"
|
||||
|
||||
jobs:
|
||||
claude-review:
|
||||
if: github.event.pull_request.head.repo.full_name == github.repository
|
||||
# Optional: Filter by PR author
|
||||
# if: |
|
||||
# github.event.pull_request.user.login == 'external-contributor' ||
|
||||
# github.event.pull_request.user.login == 'new-developer' ||
|
||||
# github.event.pull_request.author_association == 'FIRST_TIME_CONTRIBUTOR'
|
||||
|
||||
runs-on: ubuntu-latest
|
||||
permissions:
|
||||
contents: read
|
||||
pull-requests: write # post the review + inline comments
|
||||
issues: read
|
||||
id-token: write
|
||||
|
||||
steps:
|
||||
- name: Checkout repository
|
||||
uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
|
||||
with:
|
||||
fetch-depth: 1
|
||||
|
||||
- name: Run Claude Code Review
|
||||
id: claude-review
|
||||
uses: anthropics/claude-code-action@6c0083bb7289c31716797a039b6367b3079cc46e # v1
|
||||
with:
|
||||
claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
|
||||
allowed_bots: 'renovate[bot]' # let Renovate PRs get reviewed
|
||||
plugin_marketplaces: 'https://github.com/anthropics/claude-code.git'
|
||||
plugins: 'code-review@claude-code-plugins'
|
||||
prompt: '/code-review:code-review ${{ github.repository }}/pull/${{ github.event.pull_request.number }}'
|
||||
# See https://github.com/anthropics/claude-code-action/blob/main/docs/usage.md
|
||||
# or https://code.claude.com/docs/en/cli-reference for available options
|
||||
|
||||
@@ -0,0 +1,63 @@
|
||||
name: Claude Code
|
||||
|
||||
on:
|
||||
issue_comment:
|
||||
types: [created]
|
||||
pull_request_review_comment:
|
||||
types: [created]
|
||||
issues:
|
||||
types: [opened, assigned]
|
||||
pull_request_review:
|
||||
types: [submitted]
|
||||
|
||||
jobs:
|
||||
claude:
|
||||
if: |
|
||||
(
|
||||
github.event_name == 'issue_comment' &&
|
||||
contains(github.event.comment.body, '@claude') &&
|
||||
contains(fromJSON('["OWNER","MEMBER","COLLABORATOR"]'), github.event.comment.author_association)
|
||||
) || (
|
||||
github.event_name == 'pull_request_review_comment' &&
|
||||
contains(github.event.comment.body, '@claude') &&
|
||||
contains(fromJSON('["OWNER","MEMBER","COLLABORATOR"]'), github.event.comment.author_association)
|
||||
) || (
|
||||
github.event_name == 'pull_request_review' &&
|
||||
contains(github.event.review.body, '@claude') &&
|
||||
contains(fromJSON('["OWNER","MEMBER","COLLABORATOR"]'), github.event.review.author_association)
|
||||
) || (
|
||||
github.event_name == 'issues' &&
|
||||
(contains(github.event.issue.body, '@claude') || contains(github.event.issue.title, '@claude')) &&
|
||||
contains(fromJSON('["OWNER","MEMBER","COLLABORATOR"]'), github.event.issue.author_association)
|
||||
)
|
||||
runs-on: ubuntu-latest
|
||||
permissions:
|
||||
contents: read
|
||||
pull-requests: write # post comments/reviews when @-mentioned on a PR
|
||||
issues: write # post comments when @-mentioned on an issue
|
||||
id-token: write
|
||||
actions: read # Required for Claude to read CI results on PRs
|
||||
steps:
|
||||
- name: Checkout repository
|
||||
uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
|
||||
with:
|
||||
fetch-depth: 1
|
||||
|
||||
- name: Run Claude Code
|
||||
id: claude
|
||||
uses: anthropics/claude-code-action@6c0083bb7289c31716797a039b6367b3079cc46e # v1
|
||||
with:
|
||||
claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
|
||||
|
||||
# This is an optional setting that allows Claude to read CI results on PRs
|
||||
additional_permissions: |
|
||||
actions: read
|
||||
|
||||
# Optional: Give a custom prompt to Claude. If this is not specified, Claude will perform the instructions specified in the comment that tagged it.
|
||||
# prompt: 'Update the pull request description to include a summary of changes.'
|
||||
|
||||
# Optional: Add claude_args to customize behavior and configuration
|
||||
# See https://github.com/anthropics/claude-code-action/blob/main/docs/usage.md
|
||||
# or https://code.claude.com/docs/en/cli-reference for available options
|
||||
# claude_args: '--allowed-tools Bash(gh pr *)'
|
||||
|
||||
@@ -54,7 +54,7 @@ jobs:
|
||||
|
||||
- name: Log in to GHCR
|
||||
if: steps.tag.outputs.skip == 'false'
|
||||
uses: docker/login-action@af1e73f918a031802d376d3c8bbc3fe56130a9b0 # v4
|
||||
uses: docker/login-action@650006c6eb7dba73a995cc03b0b2d7f5ca915bee # v4
|
||||
with:
|
||||
registry: ${{ env.REGISTRY }}
|
||||
username: ${{ github.actor }}
|
||||
@@ -78,7 +78,7 @@ jobs:
|
||||
fi
|
||||
echo "tags=${TAGS}" >> "$GITHUB_OUTPUT"
|
||||
|
||||
- uses: docker/setup-buildx-action@bb05f3f5519dd87d3ba754cc423b652a5edd6d2c # v4
|
||||
- uses: docker/setup-buildx-action@d7f5e7f509e45cec5c76c4d5afdd7de93d0b3df5 # v4
|
||||
if: steps.tag.outputs.skip == 'false'
|
||||
|
||||
- name: Build and push
|
||||
|
||||
@@ -50,7 +50,7 @@ jobs:
|
||||
echo "skip=false" >> "$GITHUB_OUTPUT"
|
||||
fi
|
||||
|
||||
- uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7
|
||||
- uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6
|
||||
if: steps.tag.outputs.skip == 'false'
|
||||
with:
|
||||
python-version: "3.14"
|
||||
@@ -58,12 +58,12 @@ jobs:
|
||||
if: steps.tag.outputs.skip == 'false'
|
||||
- run: python -m build
|
||||
if: steps.tag.outputs.skip == 'false'
|
||||
- uses: pypa/gh-action-pypi-publish@ba38be9e461d3875417946c167d0b5f3d385a247 # release/v1
|
||||
- uses: pypa/gh-action-pypi-publish@cef221092ed1bacb1cc03d23a2d87d1d172e277b # release/v1
|
||||
if: steps.tag.outputs.skip == 'false'
|
||||
|
||||
- name: Create GitHub Release
|
||||
if: steps.tag.outputs.skip == 'false'
|
||||
uses: softprops/action-gh-release@3d0d9888cb7fd7b750713d6e236d1fcb99157228 # v3
|
||||
uses: softprops/action-gh-release@718ea10b132b3b2eba29c1007bb80653f286566b # v3
|
||||
with:
|
||||
tag_name: ${{ steps.tag.outputs.tag }}
|
||||
generate_release_notes: true
|
||||
|
||||
@@ -32,7 +32,7 @@ jobs:
|
||||
python-version: ["3.11", "3.13"]
|
||||
steps:
|
||||
- uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
|
||||
- uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7
|
||||
- uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6
|
||||
with:
|
||||
python-version: ${{ matrix.python-version }}
|
||||
- run: pip install -e ".[test,dev]"
|
||||
|
||||
@@ -28,4 +28,3 @@ tools/skill_audit_analysis/data/
|
||||
tools/skill_audit_analysis/output/
|
||||
design_ideas/
|
||||
.claude/
|
||||
docs/design/
|
||||
|
||||
+4
-731
@@ -6,740 +6,13 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
|
||||
and this project adheres to [PEP 440](https://peps.python.org/pep-0440/) for
|
||||
version numbers (`X.Y.Z`, with `X.Y.ZaN` / `bN` / `rcN` for pre-releases).
|
||||
|
||||
Two active release tracks are maintained — the current stable and the
|
||||
experimental line:
|
||||
Three release tracks are maintained — the current stable, one prior
|
||||
stable, and the experimental line:
|
||||
|
||||
- **`stable/1.7`** — patch-only (`v1.7.x`)
|
||||
- **`stable/1.5`** — patch-only (`v1.5.x`)
|
||||
- **`stable/1.6`** — patch-only (`v1.6.x`)
|
||||
- **`main`** — experimental (next major)
|
||||
|
||||
Earlier stable lines (`stable/1.6`, `stable/1.5`) are frozen.
|
||||
|
||||
## [Unreleased]
|
||||
|
||||
### Added
|
||||
|
||||
- **Compaction is visible now: lifecycle events, a progress bar, and a
|
||||
persistent transcript card.** Context compaction (manual `/compact` and
|
||||
auto) emits a first-class `compaction` SSE event
|
||||
(`start` / `progress` / `end` — see the API reference) instead of loose
|
||||
info lines. The web UI renders an in-transcript card with a real progress
|
||||
bar (determinate `part k of N` during chunked summarization, indeterminate
|
||||
for single-call compactions) that settles into a result card — token delta
|
||||
plus the summary behind a fold — in both the interactive pane and the
|
||||
coordinator viewer. The result survives reloads: the persisted compaction
|
||||
marker now projects through `/history` as a `role="system"`,
|
||||
`source="compaction"` entry (resume/export/search unchanged), stamped with
|
||||
the end event's id so repaint and SSE replay can't double-render. The
|
||||
marker's `meta` additionally records `before_tokens` / `after_tokens` /
|
||||
`trigger`. Python and TypeScript SDKs gain a typed `CompactionEvent`.
|
||||
|
||||
- **One provider transport: every model call now streams (#831).**
|
||||
The per-adapter non-streaming entry (`create_completion`) is retired;
|
||||
single-shot lanes — judges, titles, compaction, web-fetch extraction,
|
||||
perception, eval, optimizer — sample through the same streaming entry
|
||||
the chat loop uses and accumulate via one shared drain, so request
|
||||
shaping can no longer drift between the two consumption styles. Two
|
||||
operator-visible consequences: long single-shot generations (a thinking
|
||||
model composing a title, a slow local judge) no longer sit in a single
|
||||
blocking read that can hit client read-timeouts — the same reason the
|
||||
Anthropic adapter already streamed internally — and judge timeouts now
|
||||
*abort* the underlying HTTP read instead of abandoning a worker thread
|
||||
on a dead call. Because every call now streams, an alias pointed at a
|
||||
model or org that cannot stream (OpenAI's verified-org streaming
|
||||
entitlement, a gateway api-version predating `stream_options` — e.g.
|
||||
older Azure OpenAI deployments) fails at request time where 1.7's
|
||||
non-streaming single-shot call succeeded; remediation is on the
|
||||
serving side (verify the org, bump the api-version/gateway) — there is
|
||||
deliberately no per-model non-streaming fallback left to configure. These lanes are also complete-or-error now: a stream
|
||||
that ends without any finish signal is treated as a generation that
|
||||
died mid-response and retried, instead of storing the partial text as
|
||||
a clean result (previously a half-generated compaction summary could
|
||||
silently replace real history). Caveats: these lanes now carry the
|
||||
same `stream_options: {include_usage: true}` the chat loop always
|
||||
sent — OpenAI-compatible servers old enough to *ignore* it stop
|
||||
producing usage rows on these lanes, and servers strict enough to
|
||||
*reject* unknown fields (pre-2024 llama.cpp/proxy builds) will 400 —
|
||||
such a server already couldn't serve turnstone's chat loop, but a
|
||||
judge/utility alias pointed at one worked on 1.7 and needs to move to
|
||||
a current server. Transient mid-stream deaths (connection drop, proxy
|
||||
hiccup) are re-issued in place up to twice with exponential backoff —
|
||||
the retry the SDK's request loop used to provide these lanes
|
||||
invisibly. Each lane accepts its own terminal marker (Anthropic
|
||||
`message_stop`, Responses terminal events); a lax server/gateway that
|
||||
never sends any terminal signal needs
|
||||
`{"finish_reason_optional": true}` in the model definition's
|
||||
capabilities JSON, which restores 1.7's tolerance (clean end-of-stream
|
||||
after output = completion) for that model on every lane — without it
|
||||
such streams fail as died-mid-generation, because SSE gives no way to
|
||||
tell the two apart and the default favors catching truncation. The
|
||||
unread `supports_streaming` capability flag (and its admin tile) is
|
||||
gone; the o-series models it described are dropped from the capability
|
||||
table entirely (see Removed).
|
||||
|
||||
- **One turn interface for every model call: `core/model_turn.py` (#827).**
|
||||
Judges (intent + output guard), perception, title generation, compaction,
|
||||
web-fetch extraction, the eval harness, the optimizer's meta lanes, and
|
||||
task agents all advance a trajectory through the same plant-call
|
||||
primitive the agent seam pioneered — Turn IR in, one shared lowering
|
||||
(argument sanitize → minted-id restore → vLLM reasoning attach), one
|
||||
shared re-ingest (blank-id repair → native-lane finalize). The judges'
|
||||
hand-built OpenAI-dict path is gone, and with it the Gemini judge's
|
||||
tool-blindness: evidence tools now work on Google models because the
|
||||
native lane round-trips `thought_signature` (with pairwise repair for
|
||||
blank-id compat responses). Provider adapters still take lowered wire
|
||||
dicts — the transport collapse and main-loop migration are tracked as
|
||||
#831 / #832.
|
||||
|
||||
- **task_agent keeps its model's reasoning across its own tool loop — on
|
||||
every provider lane.** A task agent's replayed turns now carry the
|
||||
provider-native reasoning lane the model produced — Anthropic thinking
|
||||
blocks with their signatures (commercial or an anthropic-compatible
|
||||
server), OpenAI Responses reasoning items, Gemini `thought_signature`
|
||||
fidelity blocks, and the reasoning text a vLLM `--reasoning-parser` /
|
||||
llama.cpp `reasoning_format` surfaces on the Chat Completions lane —
|
||||
instead of each turn being rebuilt from text + tool calls with the
|
||||
reasoning dropped. On a thinking model this restores reasoning continuity
|
||||
across the agent's own multi-turn tool use. On the wire the agent's
|
||||
session-minted sub-tool ids are mapped back to the provider's own ids
|
||||
(`restore_provider_tool_ids`), so the native block — replayed verbatim,
|
||||
its signature never touched — the `tool_calls` mirror, and each tool
|
||||
result always agree; internally the minted ids still key the live card,
|
||||
recall, and the cancel ledger unchanged. Replay honors the same per-model
|
||||
`replay_reasoning_to_model` flag the main loop uses on every lane: the
|
||||
vLLM Chat-Completions field replay keeps its server-type gate, and
|
||||
llama.cpp stays capture-only, matching main-loop behavior. The native
|
||||
lane is finalized by the same shared builder as the main loop's, so the
|
||||
two harnesses cannot drift.
|
||||
|
||||
- **Background shells: `bash` gains `run_in_background`, plus `bash_output` /
|
||||
`kill_shell`.** Setting `run_in_background=true` starts the command as a
|
||||
detached shell and returns immediately with a `bash_N` handle — "start a dev
|
||||
server, use it in a later call" is back as an explicit opt-in (the shape
|
||||
follows the convention the major coding agents converged on). `bash_output`
|
||||
returns only output produced since the previous read (optionally filtered by
|
||||
a regex) plus status and exit code; `kill_shell` terminates the shell's
|
||||
whole process group. Output is buffered per shell with a drop-oldest cap, so
|
||||
a chatty server can't grow memory unbounded. When a background shell exits,
|
||||
a system notice lands at the next seam (waking an idle workstream if
|
||||
needed). Shells survive a generation cancel, die with the workstream, and
|
||||
never outlive a task_agent that started them; anything a background shell
|
||||
itself backgrounds is still reaped when that shell exits — the no-leak
|
||||
guarantee below is unchanged.
|
||||
|
||||
### Changed
|
||||
|
||||
- **Breaking (1.8): compaction feedback moved from `info` events to the
|
||||
typed `compaction` SSE event.** Pre-1.8 SSE/SDK clients that ignore
|
||||
unknown event types no longer see compaction lines (they are
|
||||
deliberately not dual-emitted — dual emission would double-render on
|
||||
every current client). Consume the `compaction` lifecycle event (see
|
||||
the API reference and the `CompactionEvent` SDK type); embedders
|
||||
driving `ChatSession` through a duck-typed `SessionUI` are unaffected
|
||||
(the classic `on_info` lines are restored for them — see Fixed).
|
||||
|
||||
- **Sampling knobs (temperature, reasoning effort) now ride one assignment
|
||||
scheme: per-model alias value → operator-stored global setting → the
|
||||
model definition's declared default (effort only) → field omitted.**
|
||||
Turnstone previously manufactured values onto every unconfigured
|
||||
request — a hidden `temperature: 0.5` and a `reasoning_effort: "medium"`
|
||||
baked in at three layers — overriding serving-side defaults like a vLLM
|
||||
model's `generation_config`. Unconfigured installs now send neither
|
||||
field and the inference engine's own defaults rule; `model.temperature`
|
||||
is blank by default ("inherit each model's own default") and
|
||||
`model.reasoning_effort` defaults to the empty "inherit" choice. The
|
||||
per-model → global resolution lives in one shared resolver used by the
|
||||
session factories, the `/model` switch, and every `model_turn` lane, so
|
||||
the same alias samples identically on every surface. CLI
|
||||
`--temperature` / `--reasoning-effort` likewise default to inherit.
|
||||
|
||||
**Upgrade notes:**
|
||||
- The empty (`""`) reasoning-effort choice changed meaning from
|
||||
"explicitly disable thinking" to "inherit the model/serving default".
|
||||
On local manual-thinking models (e.g. Qwen templates with
|
||||
`enable_thinking`), a stored `""` previously sent
|
||||
`enable_thinking: false`; it now sends nothing, so the template's own
|
||||
default (often thinking ON) applies. Use **`none`** to actually
|
||||
disable reasoning.
|
||||
- Workstreams saved by earlier versions carry the old defaults
|
||||
(`temperature=0.5`, `reasoning_effort=medium`) in their persisted
|
||||
config and keep that exact behavior on resume; they pick up the new
|
||||
inherit semantics the next time you change the model or a sampling
|
||||
knob in that workstream. New workstreams inherit from the start.
|
||||
|
||||
### Removed
|
||||
|
||||
- **O-series and pre-5.4 GPT-5 rows dropped from the OpenAI capability
|
||||
table.** `o1`, `o1-mini`, `o3`, `o3-mini`, `o3-pro`, `o4-mini`,
|
||||
`gpt-5`, `gpt-5-mini`, `gpt-5-nano`, `gpt-5-pro`, `gpt-5.1`,
|
||||
`gpt-5.1-codex-max`, `gpt-5.2`, `gpt-5.2-pro`, and `gpt-5.3` no longer
|
||||
have built-in capability rows — OpenAI has retired these model ids
|
||||
from the API, so the rows described contracts no request can reach
|
||||
anymore. The table floor is now `gpt-5.4`; the search-api and
|
||||
audio/STT/TTS rows are unchanged. An alias still pinning a retired id
|
||||
fails at OpenAI itself; any other unlisted commercial id resolves to
|
||||
the generic commercial defaults (temperature sent, no declared
|
||||
reasoning-effort vocabulary, 200K window) — declare the contract on
|
||||
the model definition's capabilities JSON if you run one, or move to a
|
||||
current model.
|
||||
|
||||
### Fixed
|
||||
|
||||
- **A failed worker-thread spawn no longer wedges the workstream — at
|
||||
either spawn site — and never masquerades as success.** If
|
||||
`Thread.start()` itself raised (thread exhaustion, out-of-memory), the
|
||||
dispatcher had already claimed the worker slot but the flag's only
|
||||
clearer lived in the never-started thread — the workstream looked idle
|
||||
forever while every subsequent message queued behind a worker that
|
||||
didn't exist, until an operator force-cancel. The claim is now rolled
|
||||
back under the lock and the error propagates, so the workstream is
|
||||
dispatchable again as soon as resources recover. Affected every
|
||||
dispatch path (sends, wakes, retries, deferred-send drain, init). The
|
||||
same failure at the deferred-send drain's own spawn rolls back the
|
||||
just-accepted entry and answers the retryable `queue_full` (previously
|
||||
a 500 landed *after* the entry was registered — an invisible,
|
||||
unretractable phantom that later dispatched as duplicate turns), and a
|
||||
`/command` whose worker never spawned now answers **503**
|
||||
`{"status": "error"}` instead of the generic 200 ok that told SDK
|
||||
callers their `/clear` or `/resume` had applied.
|
||||
|
||||
- **Manual `/compact` from the web UI: no phantom user turn, no frozen
|
||||
server, cancellable.** A slash command typed into the web composer no
|
||||
longer renders as a user chat bubble (it echoes as a distinct command
|
||||
chip — commands aren't conversation turns and were never persisted as
|
||||
such). `/compact` itself now dispatches onto the workstream's worker
|
||||
slot instead of running inline on the server's event loop — previously a
|
||||
long compaction froze every SSE stream on the node for its whole
|
||||
duration, which is also why its own progress only ever arrived as one
|
||||
burst after the fact. The manual path carries `send()`'s full generation
|
||||
discipline (`compact_now()`): a force-abandoned compaction goes stale
|
||||
instead of swapping history under a successor turn — and retires at its
|
||||
next checkpoint instead of running out its remaining summary calls,
|
||||
with its late lifecycle events fenced off (`compaction_id` on every
|
||||
event, `superseded` on end events — both in the SDKs) so they can't
|
||||
animate, tear down, re-title, or falsely narrate a successor's card or
|
||||
activity pill; a cancel aimed at it is consumed on exit (previously it
|
||||
bricked every `/compact` retry until the next message); a Stop click on
|
||||
an idle session can't pre-abort the next compaction; a Stop that lands
|
||||
in the completion tail — after the last cancel check, or during a retry
|
||||
backoff (which now aborts immediately instead of sleeping it out) — is
|
||||
honored rather than silently eaten; and Stop now aborts the in-flight
|
||||
summary HTTP call itself (the compaction lane registers its stream in
|
||||
the same abort seam the main loop uses), so cancelling a compaction is
|
||||
immediate instead of waiting out a model call.
|
||||
|
||||
- **Sends during a command window are deferred, ordered, bounded, and
|
||||
honestly rendered — never silently truncated or lost.** Messages sent
|
||||
while any slash command holds the worker slot are **deferred**: answered
|
||||
`{"status": "queued", "msg_id"}` immediately and dispatched as ordinary
|
||||
full-fidelity sends (attachments and sender identity included) when the
|
||||
command finishes — never routed through the mid-turn interjection
|
||||
queue, whose semantics are turn-shaped: previously a send during a
|
||||
manual `/compact` was silently truncated to 2,000 characters, a second
|
||||
participant in a shared workstream was locked out with a misleading
|
||||
"another participant's turn" 409 for the whole compaction, and a
|
||||
message queued across a `/resume`/`/new` could be answered into the
|
||||
post-swap workstream. Because the response is immediate,
|
||||
timeout-bounded callers — the coordinator's `send_message`, the console
|
||||
proxy, SDKs, anything behind a stock reverse proxy — can no longer lose
|
||||
a message to a multi-minute command window; the deferred send is
|
||||
retractable until dispatch via the same `DELETE .../send` used for
|
||||
queued interjections (node-local, in-memory — the API reference
|
||||
documents the at-most-once durability contract). Deferred responses
|
||||
carry `"deferred": true`; the pending list is the **order authority**
|
||||
(a fresh send — or a coordinator dispatch, or a queued-nudge wake —
|
||||
lines up behind acknowledged entries instead of overtaking them, with
|
||||
the two-term barrier defined once on the workstream so the wake gate
|
||||
also honors a claimed entry whose dispatch is mid-flight, and the gate
|
||||
re-arms at the drain's exit even when everything pending was
|
||||
retracted); acceptance is **bounded** (10 pending per workstream — the
|
||||
interjection queue's own backpressure contract; the 11th answers the
|
||||
retryable `queue_full` instead of pinning attachment bytes without
|
||||
limit and then running one unattended turn per entry); a dispatch
|
||||
crash re-queues the entry instead of eating an acknowledged message,
|
||||
and a drain thread that fails to *start* rolls the acceptance back and
|
||||
answers `queue_full` rather than parking a phantom the client can
|
||||
neither see nor retract; each dispatch emits a pane-tier
|
||||
`message_dispatched` event (`folded: true` for interjection fold-ins)
|
||||
so queued-bubble UI keeps its retract affordance exactly until the
|
||||
message truly leaves — including when the send was accepted by a pane
|
||||
that believed the workstream idle, which now renders a real queued
|
||||
chip instead of a sent-looking bubble, releases the composer (a
|
||||
deferred send has no running worker to wait on), and cleans up fully
|
||||
when the send is refused or the chip retracted instead of stranding
|
||||
the pane in Stop mode. Dismissing a queued bubble — interjection or
|
||||
deferred — is a server-confirmed `DELETE`, and retracting a deferred
|
||||
send that carried attachments tells the user they were discarded
|
||||
instead of silently expiring them.
|
||||
|
||||
- **Slash commands hold the worker slot with a loud contract.**
|
||||
A `/compact` raced against an in-flight turn is refused with an
|
||||
explicit busy response. Every other slash command runs through the same
|
||||
worker slot too — mutual exclusion against sends, a running compaction,
|
||||
and each other, with a busy answer replacing the old silent interleave —
|
||||
while the endpoint still awaits quick commands' completion off-loop
|
||||
(without parking an executor thread per request); the post-command pane
|
||||
refreshes (`clear_ui` after `/clear`/`/new`/`/resume`, the
|
||||
workstream-name sync) ride the worker itself, so a command that
|
||||
outlives the endpoint's 25s response backstop still refreshes every
|
||||
pane on completion (the backstop sits under the console proxy's 30s
|
||||
client timeout so the degraded `running` answer can actually traverse
|
||||
a proxied pane, which now surfaces it instead of silence; the
|
||||
`/command` response contract — `ok` / `running`, with busy refusals
|
||||
answering a loud HTTP 409 rather than a silent 200 — is now documented
|
||||
in the API reference and the OpenAPI spec).
|
||||
|
||||
- **Compaction status stays truthful across every UI surface.** Manual
|
||||
compaction
|
||||
success also refreshes the status line/context pill immediately (parity
|
||||
with auto-compaction), compaction failures keep feeding the typed
|
||||
`error` event and the node error counter (while a CLI Ctrl-C reports as
|
||||
cancelled, not a failure), one Stop prints one notice (a cancelled
|
||||
auto-compaction no longer stacks "Compaction cancelled." on top of
|
||||
send's own "[Generation cancelled]"), the workstream activity pill
|
||||
shows "Compacting context…" for the whole summarize phase, restores
|
||||
cleanly afterwards, and can no longer be stranded by a force-stopped
|
||||
compaction (a new turn's generation claim breaks a stale latch). Every
|
||||
retry backoff on the session (stream retries, task agents, notify
|
||||
delivery, compaction) now aborts immediately on Stop via one shared
|
||||
cancel-aware helper instead of sleeping out its exponential delay.
|
||||
|
||||
- **Compaction failures report exactly once, to the right owner.** A
|
||||
compaction failure reports
|
||||
exactly once (auto-compaction errors defer to the turn's fatal handler
|
||||
instead of doubling the red row and the error metric), failed-end
|
||||
notice suppression is computed once by the emitter (a `notice` bool on
|
||||
the end event — in the SDKs — replaces hand-synced client policy), and
|
||||
a manual `/compact` failure no longer crashes the CLI REPL. `/compact`
|
||||
on a workstream showing the `error` badge restores the badge on exit
|
||||
instead of stamping `idle` over it (the compaction neither retried nor
|
||||
resolved the failed turn). A force-cancelled initial send that
|
||||
completes late still delivers its scheduled-run completion
|
||||
notification (the only completion signal unattended workstreams have);
|
||||
the other post-command pane refreshes and error notices remain
|
||||
owner-guarded, so a force-cancelled wedged command that unwedges late
|
||||
can't wipe panes or inject stray notices into a successor turn.
|
||||
|
||||
- **Pre-1.8 embedder UIs keep their compaction lines.** Embedders
|
||||
driving `ChatSession` with a pre-1.8 duck-typed `SessionUI`
|
||||
(no `on_compaction` hook) get the classic `on_info` compaction lines
|
||||
back — threshold notice, `part k/N`, retry waits, token delta +
|
||||
summary box — instead of silent history swaps. (See the breaking
|
||||
event-contract note under **Changed** for SSE/SDK clients.)
|
||||
|
||||
- **Static MCP servers: a pushed catalog change no longer wedges the shared
|
||||
session (#839).** The static-path `*/list_changed` handler awaited its
|
||||
catalog refresh inline in the SDK's receive loop, but the refresh's own
|
||||
request can only be answered by that (now parked) loop — the refresh never
|
||||
completed, and every user's in-flight calls on the shared per-node session
|
||||
stalled behind it, unbounded, until the health loop's ping timeout tore the
|
||||
transport down (which was also the only way the changed catalog ever
|
||||
landed). Push refreshes now run as spawned tasks — debounced, coalesced per
|
||||
(server, kind), bounded by the connect timeout, and serialized on the
|
||||
per-server connect lock — and the manual and post-reconnect refreshes
|
||||
publish under that same lock, so a slower publisher can no longer land a
|
||||
staler catalog over a fresher one. Every teardown path now also clears the
|
||||
notification debounce stamp, so a reconnected server's first push refreshes
|
||||
immediately. Push-refresh debouncing is now per (server, kind) on BOTH the
|
||||
static and per-user pool paths — a tools push no longer swallows a prompts
|
||||
push arriving in the same 5-second window. A change genuinely lost to the
|
||||
debounce window (a same-kind push landing after the prior refresh finished,
|
||||
which the server will never re-announce) is recovered by an automatic
|
||||
health-tick retry rather than staying invisible until an unrelated push or
|
||||
a reconnect. The resource-refresh fan-out on both paths no longer orphans
|
||||
its sibling list call when one of the pair fails fast — the real error
|
||||
surfaces immediately (not masked as a 30-second timeout) and the surviving
|
||||
sibling is cancelled and reaped, under a bounded grace, inside the scope. A
|
||||
push refresh that fails while the connection stays up is likewise retried on
|
||||
the next health-loop tick until one completes — previously a single
|
||||
transient blip left the shared catalog stale for every user on the node
|
||||
until an operator intervened. An operator `/mcp refresh` no longer parks
|
||||
behind a busy per-server connect lock (a slow reconnect attempt could eat
|
||||
the whole 30-second refresh budget and fail the pass for every healthy
|
||||
server behind it) — the busy server is skipped on both the connected and
|
||||
disconnected branches, reported distinctly as "skipped" rather than as a
|
||||
false "no changes", the skip arms the automatic retry, and a
|
||||
force-reconnect drops the session up front so queued push refreshes can't
|
||||
starve it. Static-path resource and prompt catalogs are now size-capped
|
||||
like the pool path's (and like static tools) at discovery and on every
|
||||
refresh, so a misbehaving server's push can't balloon the node's merged
|
||||
catalogs. Deleting or reconfiguring a server can no longer leave it
|
||||
half-removed: the config removal and all cleanup are serialized under the
|
||||
connect lock (a cancelled removal completes its cleanup rather than
|
||||
stranding a live session and published catalog with the config already
|
||||
gone), and `reconcile_sync` retries a removal that timed out instead of
|
||||
marking it done — previously a DB-driven delete of a busy server could be a
|
||||
silent, permanent no-op until process restart. A refresh outcome now
|
||||
threads consistently to every operator surface off one source of truth
|
||||
(the per-server `last_refresh_outcome`): a busy-skip and a genuine failure
|
||||
are each reported distinctly from a real "no changes" — `/mcp refresh`
|
||||
prints "skipped" or "failed" rather than a false "no changes", and the
|
||||
node-internal refresh endpoint returns `202 skipped` instead of a
|
||||
misleading `200 ok` for a refresh that never ran. A single-kind push
|
||||
refresh no longer paints the whole server healthy: because the
|
||||
error/outcome state is server-scoped, a successful tools push while the
|
||||
prompts catalog is still broken (or vice versa) no longer clears the
|
||||
failure — only a full refresh pass declares "ok".
|
||||
|
||||
- **OpenAI Responses streaming: truncated and refused responses no longer
|
||||
vanish.** A response that hit `max_output_tokens` terminates the stream
|
||||
with `response.incomplete`, which the stream consumer did not handle —
|
||||
the turn was mislabeled `finish_reason: stop` and its final usage and
|
||||
collected output items were dropped. Refusal parts had no streaming
|
||||
handler at all, so a refusal rendered as empty content instead of the
|
||||
`[Refused: …]` text the non-streaming path produced. Both now match:
|
||||
truncation maps to `length` with usage/items intact, refusals render
|
||||
in content. Applies to the chat loop and every drained single-shot
|
||||
lane (#831).
|
||||
|
||||
- **task_agent: sub-tool ids no longer alias across a local model's reused
|
||||
ids.** A local model that reissues per-response sequential tool-call ids
|
||||
(`call_0` every turn) made two of a task agent's steps share one id — the
|
||||
live card collapsed both onto one DOM row while `/history` recall kept them
|
||||
apart, so the two views disagreed. Sub-tool ids are now minted
|
||||
`{parent}::r{run}s{step}::{id}`, unique within the session (across an
|
||||
agent's turns and across concurrent or sequential runs), and that one id
|
||||
keys the nesting registry, the live rows, recall, and the cancel ledger.
|
||||
On the wire the agent's self-built history carries the provider's own ids,
|
||||
restored from the mint map (see the reasoning-lane entry under Added), and
|
||||
malformed tool-call arguments are legalized the same way the main loop's
|
||||
wire prep does.
|
||||
|
||||
- **bash tool: never hang on a backgrounded child.** A command that left a
|
||||
long-lived process running (`server &`, a daemon) could wedge the whole
|
||||
workstream forever — the tool read stdout/stderr to EOF, which never arrived
|
||||
because the child inherited the pipe, and the timeout watchdog bailed once the
|
||||
foreground `bash` had exited. The tool now waits on the tracked process
|
||||
(bounded by the tool timeout) and terminates its whole process group on
|
||||
return, so the call always completes. Undecodable output is preserved
|
||||
(`errors="replace"`) instead of being dropped as a spurious error.
|
||||
- **Behavior change:** a process the command backgrounds no longer survives
|
||||
the call — nothing persists across bash invocations. (First-class
|
||||
"run this in the background" support landed separately — see
|
||||
`run_in_background` under Added.)
|
||||
|
||||
## [1.7.3]
|
||||
|
||||
A small feature and maintenance patch for the 1.7 line. No schema migrations
|
||||
and no new configuration knobs.
|
||||
|
||||
### Added
|
||||
|
||||
- **OpenAI GPT-5.6 (Sol/Terra/Luna) support** — the Responses provider
|
||||
understands the GPT-5.6 family: the `reasoning.mode` control, the new
|
||||
`max` effort tier, and `text.verbosity`, with golden wire payloads pinning
|
||||
the request shapes. The `openai` dependency floor moves to `>=2.44`.
|
||||
|
||||
### Changed
|
||||
|
||||
- **Engineer base prompt hardened with process discipline** — the default
|
||||
base prompt for non-coordinator sessions now works in phases scaled to the
|
||||
size of the change, defaults to red-green for testable work, scopes to the
|
||||
smallest sufficient diff, stops to report after repeated failed attempts
|
||||
instead of thrashing, reports only observed results, and delegates
|
||||
exploration to `task_agent`. Persona prompts freeze into the workstream
|
||||
stamp at creation, so this reaches new workstreams only.
|
||||
|
||||
### Fixed
|
||||
|
||||
- **Unknown reasoning-mode warnings name the allowed modes** — a model
|
||||
definition with an unrecognized reasoning mode now logs the valid options
|
||||
instead of leaving the operator to guess.
|
||||
|
||||
### Documentation
|
||||
|
||||
- **HYPOTHESIS.md / PRIMER.md** — the control normal form is tightened and
|
||||
the factored Q_E reading is carried into the glossary; the plain-language
|
||||
PRIMER stays in sync.
|
||||
|
||||
## [1.7.2]
|
||||
|
||||
A feature-bearing patch for the 1.7 line. Rather than hold this work for the
|
||||
larger 1.8 churn, the fixes and the smaller features that had already
|
||||
stabilised on `main` are rolled into the stable line now: a rich preview
|
||||
pane, persona/project settings on scheduled tasks, and a batch of streaming,
|
||||
rendering, and nudge-delivery hardening.
|
||||
|
||||
> **⚠️ Before upgrading:** 1.7.2 adds Alembic migration `066`, applied
|
||||
> automatically on first start. It adds two `Text NOT NULL DEFAULT ''`
|
||||
> columns (`persona`, `project_id`) to the `scheduled_tasks` table; existing
|
||||
> rows migrate to the empty default, which is byte-identical to pre-066
|
||||
> dispatch behaviour. The change is additive and reversible, but — as always
|
||||
> — back up your storage before upgrading (`pg_dump` for PostgreSQL; copy the
|
||||
> database file for SQLite).
|
||||
|
||||
### Added
|
||||
|
||||
- **Rich preview pane + `open_preview` tool** — a workstream can now open a
|
||||
rendered preview (HTML, Markdown, and other kinds) in a pane beside the
|
||||
conversation via the new `open_preview` tool. Guarded fetches stream under
|
||||
a byte budget whose ceiling tracks the widest per-kind cap, preview blob
|
||||
ids are salted, and a preflight probe handles legacy charsets and a
|
||||
remote-assets opt-in. See `docs/tools.md`.
|
||||
- **`allow_private_network` opt-in for `web_fetch` / `open_preview`** —
|
||||
private-address fetch and preview targets stay blocked by default; an
|
||||
operator can opt a workstream in through the settings registry when a
|
||||
private endpoint is genuinely intended. (Distinct from the 1.7.1 `[oidc]`
|
||||
flag of the same name, which governs identity-provider discovery.)
|
||||
- **Persona + project settings on scheduled tasks** (migration `066`) — a
|
||||
scheduled task can now pin the **persona** and **project** of the
|
||||
workstream it dispatches, matching the levers a manually-created workstream
|
||||
already carries. Both default to empty (kind-default persona / no project),
|
||||
so existing schedules dispatch exactly as before.
|
||||
|
||||
### Fixed
|
||||
|
||||
- **Streaming fast-path overflow recovery** — fast-stream tokens are now
|
||||
batched and overflowed SSE listeners recover instead of stalling (and
|
||||
`connectSSE` no longer opens into a hidden background tab). The same
|
||||
overflow-recovery companions were carried to the coordinator pane, so a
|
||||
coordinator watching many children recovers dropped listeners the same way
|
||||
the live-session view does.
|
||||
- **Renderer containment** — markdown sentinel-forgery and recursive-frame
|
||||
content loss are contained, and an indented fence close no longer drags its
|
||||
indent into the enclosed code content.
|
||||
- **Idle nudge / wake delivery** — nudge and wake delivery is hardened across
|
||||
session eviction, cancellation, and identity rebinds; the wake gate now
|
||||
requires a real nudge queue, refused wakes are logged, and
|
||||
`initial_message_status` is typed as a closed enum on the wire.
|
||||
- **`web_fetch` extraction inherits model settings** — the completion that
|
||||
extracts content from a fetched page now inherits the workstream's model
|
||||
settings instead of falling back to defaults.
|
||||
- **UI panes** — ephemeral panes close on split-dismiss instead of orphaning
|
||||
a tab, and an unsplit skips the redundant refresh after an ephemeral pane
|
||||
closes.
|
||||
- **Shared code-highlight CSS** — renderer-output CSS is shared so the console
|
||||
and coordinator panes highlight code identically.
|
||||
|
||||
### Security
|
||||
|
||||
- **`Content-Disposition` filenames made wire-safe** — download filenames
|
||||
derived from user-controlled text are sanitised (latin-1- and
|
||||
control-char-safe, quoting-safe) before they reach the `Content-Disposition`
|
||||
response header, including the fallback path.
|
||||
|
||||
### Documentation
|
||||
|
||||
- **HYPOTHESIS.md: daemons + the outer loop, plus a plain-language PRIMER** —
|
||||
the harness north-star document gains its daemon / outer-loop treatment and
|
||||
a new top-level `PRIMER.md`.
|
||||
|
||||
## [1.7.1]
|
||||
|
||||
A maintenance and hardening patch for the 1.7 line. No schema migrations;
|
||||
the credential-redaction work below is additive and needs no configuration
|
||||
change. The one new operator-facing knob is the opt-in `[oidc]
|
||||
allow_private_network` flag (default off).
|
||||
|
||||
### Security
|
||||
|
||||
- **Credential redaction hardened across the tool-call surface** — the
|
||||
redactor that scrubs secrets from tool arguments and log previews was
|
||||
reworked on both the backend and the browser to close several leak paths
|
||||
and to fix false-positive and performance issues. Malformed tool-call
|
||||
arguments are now legalised before they reach the wire; the tool-args log
|
||||
preview scrubs credentials and control characters; and the coordinator's
|
||||
tool-call cards gain a matching client-side redaction pass so the JS and
|
||||
backend redactors stay at parity. Pattern coverage now includes
|
||||
`secret_access_key` / `aws_secret_access_key` multi-segment keys, bare
|
||||
`token=` / `key=` forms (guarded by a negative lookbehind to avoid
|
||||
false positives), and SQLAlchemy `+driver`-qualified connection-string
|
||||
schemes matched case-insensitively.
|
||||
- **OIDC SSRF guard: `[oidc] allow_private_network` opt-in** — self-hosted
|
||||
identity providers on private networks can now be reached by setting
|
||||
`allow_private_network = true` under `[oidc]` (default off; the MCP OAuth
|
||||
path stays strict). Rejections of discovered endpoints carry the opt-in
|
||||
hint so the misconfiguration is self-explanatory. See `docs/oidc.md`.
|
||||
|
||||
### Added
|
||||
|
||||
- **Persona discoverability + forgiving name resolution** — personas are
|
||||
now discoverable by agents, and persona-name resolution tolerates
|
||||
case/whitespace variation; a not-found resolution reports the offending
|
||||
input verbatim instead of a bare error.
|
||||
|
||||
### Fixed
|
||||
|
||||
- **MCP transport lifecycles routed through per-entry owner tasks**
|
||||
(#787/#788) — static and pooled MCP transport lifecycles are now driven
|
||||
by per-server / per-entry owner tasks, with a hardened disarm-sweep loop
|
||||
guard and targeted exception handling in place of a broad `BaseException`
|
||||
arm, so a dying transport can no longer spin the CPU or strand delivery.
|
||||
- **Client-construction failures surface as misconfiguration, not raw
|
||||
500s** — a model whose client cannot be constructed now reports a factory
|
||||
misconfiguration, and the raw exception text is kept out of the resulting
|
||||
503 response.
|
||||
- **Postgres history search survives oversized rows** — a conversation row
|
||||
exceeding Postgres' full-text limits no longer aborts history search.
|
||||
- **Agent-tool render is idempotent** — tool rendering no longer deep-copies
|
||||
a tool definition until a description actually changes, so no-persona
|
||||
sessions share the tool constant (correctness plus a hot-path allocation
|
||||
win).
|
||||
- **Private-project workstream visibility scoped to members** — workstreams
|
||||
in a private project are visible to project members only, not to every
|
||||
admin; coordinator tenancy checks now use request-scoped storage.
|
||||
- **Pane hotkeys work off macOS and match across surfaces** — the pane
|
||||
keyboard shortcuts no longer collide with browser accelerators on
|
||||
non-macOS platforms and behave consistently across surfaces.
|
||||
|
||||
## [1.7.0]
|
||||
|
||||
The headline of the 1.7 line is **Personas** — operator-authored control
|
||||
over how each workstream composes its system message and capability
|
||||
envelope. The rest of the release hardens the pieces a persona leans on:
|
||||
concurrent approvals, cross-provider reasoning-effort control, cooperative
|
||||
compaction, multi-user session safety, and MCP resilience for unattended
|
||||
work.
|
||||
|
||||
> **⚠️ Before upgrading:** 1.7.0 adds Alembic migrations `062`–`065`,
|
||||
> applied automatically on first start (projects, personas, and two
|
||||
> smaller schema tidy-ups). Migration `063` creates the `personas` table
|
||||
> with its six seed personas and converts existing `creative_mode`
|
||||
> workstreams to the `writer` persona in place. The changes are additive
|
||||
> to your conversation data, but — as always — back up your storage before
|
||||
> upgrading (`pg_dump` for PostgreSQL; copy the database file for SQLite).
|
||||
|
||||
**Breaking changes at a glance** (details in the sections below): the
|
||||
`/creative` REPL toggle removed (replaced by the `writer` persona), the
|
||||
`turnstone-bootstrap` entry point renamed to `turnstone-doctor`, and the
|
||||
approval-status API/SDK field `pending_approval_details` changed from a
|
||||
single object to a list (one entry per concurrent approval cycle).
|
||||
|
||||
### Added
|
||||
|
||||
- **Personas** (#683) — a named, reusable bundle attached to a workstream
|
||||
at creation, controlling system-message composition and the capability
|
||||
envelope via exactly four levers: base-prompt override, tool visibility
|
||||
set, MCP on/off, and memory on/off. The persona is resolved once and
|
||||
snapshotted into `workstream_config`; editing or archiving a persona
|
||||
never changes an existing workstream. Six seed personas ship with
|
||||
migration `063` (`engineer` and `orchestrator` are the per-kind
|
||||
defaults with no overrides, so zero-touch behavior is unchanged;
|
||||
`scribe`, `researcher`, `writer`, and `executive` are curated
|
||||
envelopes). Selectable on every creation surface (web pickers, the
|
||||
create API/SDKs, coordinator `spawn_workstream` / `spawn_batch`, and
|
||||
`turnstone --persona <name>`); authored in the console's new
|
||||
Governance → Personas tab (`persona.{create,read,write}` perms,
|
||||
archive-only lifecycle). See `docs/personas.md`.
|
||||
- **Projects — governed resource containers** (#724) — group workstreams
|
||||
and their resources under a project (migration `062`), with
|
||||
project-scoped memory, a per-project resources view, a project column on
|
||||
the saved list, and server-enforced private-project workstream
|
||||
visibility.
|
||||
- **Task-agent sub-harness** (#732) — a spawned task agent now runs on its
|
||||
own Turn-IR sub-harness with parent-tagged step events: its sub-tool
|
||||
steps nest inside an expandable card in the parent trajectory, its
|
||||
sub-trajectory is recallable, and each agent gets read isolation from
|
||||
its siblings.
|
||||
- **MCP static-server autonomous reconnect** (#768) — statically
|
||||
configured MCP servers are now kept live by a health loop
|
||||
(capped-jittered backoff, ping-based liveness) instead of silently
|
||||
staying dead after the first transport drop.
|
||||
- **Attachments — capability-gated client-side fallback** — when the
|
||||
active model can't natively handle an attachment, the client degrades
|
||||
gracefully (PDF → extracted text, audio → transcript) instead of
|
||||
failing the turn.
|
||||
- **Eval measurement / optimizer split** (#763, #765) — `turnstone-eval`
|
||||
is now a measure-only substrate with the prompt optimizer factored out,
|
||||
plus a new skill-adherence measurement mode.
|
||||
- **Deployment examples** — a vLLM + LiteLLM unified-memory inference
|
||||
example showing a 3-model co-resident stack with an HF loader (#686,
|
||||
#688), and an Altair + `vl-convert-python` visualization stack (#685).
|
||||
- **Concurrent approvals and a long-session frontend overhaul** (#754,
|
||||
#755, #773, #775) — the live-session frontend was reworked for long
|
||||
runs (the pipeline is wedge-proofed and its hot paths de-O(N)'d), and on
|
||||
top of it a workstream can now hold more than one tool call awaiting
|
||||
approval at a time. Each parallel batch gets its own approval cycle,
|
||||
with one card per pending call in the interactive and coordinator UIs,
|
||||
cycle-keyed tracking in Slack and Discord, and cycle-routed resolution
|
||||
across the server/console/SDK APIs; sub-agent tool gates run the
|
||||
intent-judge pipeline as their own generation. The send button no longer
|
||||
sticks disabled after a batch resolves — orphaned approval cycles are
|
||||
pruned and the app is the sole owner of the button state.
|
||||
*(BREAKING: the `pending_approval_details` field is now a list, oldest
|
||||
first.)*
|
||||
- **Reasoning-effort control on every provider lane** (#771, #774) — the
|
||||
session effort knob now reaches local backends too: it drives
|
||||
`chat_template_kwargs` on the anthropic-compatible and openai-compatible
|
||||
lanes and threads through to Gemini and xAI, alongside the commercial
|
||||
providers that handle effort natively. The console surfaces each model's
|
||||
effective effort ladder in plain words and adds an always-on
|
||||
thinking-mode option to the model form. Effort snapping is ordinal —
|
||||
it rounds up and caps at the model's ceiling rather than silently
|
||||
dropping.
|
||||
|
||||
### Changed
|
||||
|
||||
- **Skills are capability-context, not identity** (#762) — a task agent's
|
||||
identity now comes from its persona; an applied skill's body is demoted
|
||||
to capability context and moved out of the identity system message.
|
||||
Skill-body substitution is unified across every invocation context so
|
||||
the same skill renders identically whether loaded interactively, by the
|
||||
model, or inside a sub-agent.
|
||||
- **`turnstone-doctor` replaces `turnstone-bootstrap`** (#718)
|
||||
*(BREAKING)* — the setup/diagnostics entry point is renamed; update any
|
||||
scripts or service units that invoke `turnstone-bootstrap`.
|
||||
- **Honest cancellation dispositions** — cancelled or timed-out
|
||||
side-effecting tools now report an `UNKNOWN` disposition rather than a
|
||||
flat failure, tool dispositions are typed (not just prose), and a
|
||||
coordinator cancel propagates down the sub-tree.
|
||||
- **Multi-user shared-workstream context** (#750) — in a shared
|
||||
workstream, send is gated to the acting participant while a turn is in
|
||||
flight (both the interactive and coordinator surfaces), cross-user
|
||||
mid-turn interjections are blocked, and shared-workstream state plus
|
||||
fork sender attribution are now durable.
|
||||
- **Cooperative compaction** (#730) — the context budget is anchored to
|
||||
the provider's true capacity, the summary call is chunked so it can't
|
||||
overflow, and the active plan and the outstanding ask are carried across
|
||||
compaction verbatim. The `recall` tool is scoped to the compacted-away
|
||||
past.
|
||||
- **Intent judge sees the full tool arguments** (#760) — the judge's
|
||||
argument projection is no longer narrowed, so it stops issuing confident
|
||||
false denials on a partial view. The output-guard judge sources its real
|
||||
context window, and `context_window = 0` in `config.toml` now means
|
||||
auto-detect.
|
||||
|
||||
### Fixed
|
||||
|
||||
- **Compaction resume hardening** (#731) — checkpoint markers are
|
||||
persisted so resume rehydration is bounded, context-overflow on resume
|
||||
is recovered across providers, and a recognized rate-limit is no longer
|
||||
misclassified as context overflow.
|
||||
- **MCP unattended-work resilience** (#706, #742, #767) — dead-transport
|
||||
handling is completed, consented OAuth (OBO) tokens are refreshed
|
||||
proactively so autonomous runs don't strand on an expired grant, the
|
||||
Entra ID on-behalf-of impersonation flow blockers are closed (migration
|
||||
`065` adds the OIDC `oid`), and OAuth refresh failures are classified so
|
||||
a transient blip never revokes consent nor a dead grant strands the
|
||||
user.
|
||||
- **Memory writes** (#735) — save/update is a single atomic upsert, and
|
||||
writing a memory no longer recomposes the system prefix mid-session.
|
||||
|
||||
### Removed
|
||||
|
||||
- **`/creative` removed** *(BREAKING)* — subsumed by the Personas feature
|
||||
above: the REPL toggle (and its tab completion) is gone, and the
|
||||
`writer` seed persona replaces it — start a session with
|
||||
`turnstone --persona writer` or pick *Writer* in the web
|
||||
pickers. Unlike the old fork, the writer persona composes the full
|
||||
system message, so session context and mandatory prompt policies now
|
||||
apply to prose-only sessions too. The `creative_mode` key in
|
||||
`workstream_config` is no longer read or written. Migration `063`
|
||||
converts existing creative-mode workstreams to the `writer` persona
|
||||
automatically, so they resume as writing sessions rather than as
|
||||
legacy defaults.
|
||||
|
||||
### Security
|
||||
|
||||
- **High-risk skill activation is gated** (#762) — a model-initiated load
|
||||
of a `high`- or `critical`-risk skill is gated and fails closed when the
|
||||
backing storage is unavailable, so an untrusted turn can't silently
|
||||
pull in a dangerous capability.
|
||||
- **Dependency security floors** — `cryptography` and `starlette` are
|
||||
pinned to security-fixed minimums.
|
||||
- **CI publish hardening** — the vendored-JS dispatch path refuses fork
|
||||
PRs, and `workflow_run` publishing is gated to same-repo tag pushes, so
|
||||
a fork can't trigger a release build.
|
||||
|
||||
## [1.6.0]
|
||||
|
||||
The first stable release of the 1.6 line — and the first under Apache 2.0.
|
||||
|
||||
+2
-6
@@ -8,7 +8,7 @@ FROM python:3.14-slim
|
||||
LABEL org.opencontainers.image.title="turnstone" \
|
||||
org.opencontainers.image.description="Multi-node AI orchestration platform"
|
||||
|
||||
COPY --from=ghcr.io/astral-sh/uv:0.11.29 /uv /usr/local/bin/uv
|
||||
COPY --from=ghcr.io/astral-sh/uv:0.11.26 /uv /usr/local/bin/uv
|
||||
|
||||
# Remove the slim image's man page exclusion so man-db has actual content
|
||||
RUN rm -f /etc/dpkg/dpkg.cfg.d/docker
|
||||
@@ -60,12 +60,8 @@ COPY docker/entrypoint.sh /usr/local/bin/entrypoint.sh
|
||||
WORKDIR /data
|
||||
RUN chown turnstone:turnstone /data
|
||||
|
||||
# Workspace mount point — bind-mount a host directory here. The env var
|
||||
# surfaces the path in the model's shell/file tool descriptions
|
||||
# (config.get_workspace_dir); without it the mount is invisible to the
|
||||
# model, whose cwd is /data below.
|
||||
# Workspace mount point — bind-mount a host directory here
|
||||
RUN mkdir -p /workspace && chown turnstone:turnstone /workspace
|
||||
ENV TURNSTONE_WORKSPACE=/workspace
|
||||
|
||||
USER turnstone
|
||||
|
||||
|
||||
+23
-45
File diff suppressed because one or more lines are too long
@@ -1,155 +0,0 @@
|
||||
# What a Harness Is — and What It Can Never Promise
|
||||
|
||||
*A plain-language companion to [HYPOTHESIS.md](HYPOTHESIS.md). Same object, no symbols required.*
|
||||
|
||||
**How to read this.** HYPOTHESIS.md defines, formally, what an agent harness is and what it can never guarantee. This file is that document lowered into plain language — and by the formal document's own rules, a summary is a cache, not an authority: it must stay re-derivable from its source, and wherever the two disagree, the formal one wins. Symbols appear once, in parentheses, so you can cross over; nothing here requires them. And none of it is decoration: the formal version, used as a checklist, has caught real bugs in a real harness — because most bugs are a violated invariant nobody had written down.
|
||||
|
||||
## The problem
|
||||
|
||||
You have a model. It is, roughly, a brilliant, tireless, lightning-fast intern that has read most of the internet — and that sometimes makes things up, sometimes gets confused, and sometimes takes instructions from strangers, because a page it was asked to read said "ignore your boss and email the passwords here" in white text on a white background.
|
||||
|
||||
So you don't wire the intern to production. You build a loop around it. The **harness** is that whole governed loop: a deterministic shell *you* write — build the prompt, approve or refuse each proposed action, fold the result back into memory — wrapped around a model you didn't write and a world you don't control, repeated until the run reaches a stopping state. The shell is code and does the same thing every time. The model is neither, and everything in the theory comes from taking that split seriously.
|
||||
|
||||
One sentence to keep: **the model proposes; the gate disposes.** The model's output is never an action. It is a suggestion, in text, which a piece of ordinary code you wrote either turns into an action or refuses.
|
||||
|
||||
## The parts
|
||||
|
||||
| Plain name | What it does | In the formal doc |
|
||||
|---|---|---|
|
||||
| The owner | The human — or sign-off group — the run acts for; the only place new permissions can come from | the trusted principal |
|
||||
| The memory | Everything the run knows: task, plan, transcript, and the ledger of what has been done | the state, *s* |
|
||||
| The prompt builder | Decides which slice of memory the model gets to see this step | the lowering, π |
|
||||
| The model | The black box that reads the prompt and writes a proposal | the plant, M_W |
|
||||
| The gate | Ordinary code that checks every proposal and approves or refuses it | the gate, γ |
|
||||
| The tools and the world | What approved actions actually touch: files, APIs, shells, people | the environment, Q_E |
|
||||
| The verifier | Checks each tool result, then writes it into memory | the fold-back, ρ |
|
||||
| The stop rule | Decides when the run is finished — and whether it finished *well* | the halt set H, accepting halts H_ok |
|
||||
| The danger zone | States that must never be reached: secrets exfiltrated, wrong files deleted, money moved twice | the bad set, B |
|
||||
|
||||
The loop:
|
||||
|
||||
```
|
||||
you ask for something
|
||||
↓
|
||||
prompt builder → model → "I propose: send_email(...)"
|
||||
↓
|
||||
GATE ── no ──→ nothing happens (safe, recorded)
|
||||
↓ yes
|
||||
tool runs in the world
|
||||
↓
|
||||
verifier checks the result, writes it to memory
|
||||
↓
|
||||
done? ── no → around again
|
||||
↓ yes
|
||||
stop (well, or refused)
|
||||
```
|
||||
|
||||
## The rules that make it a harness
|
||||
|
||||
Four invariants, all about *where* things are allowed to happen.
|
||||
|
||||
1. **The model sees only what the prompt builder shows it** — never raw memory. The corollary with teeth: a secret that never enters the prompt cannot leak through the model. The redaction step that keeps credentials and other people's data out of the prompt must be dumb, deterministic code — the moment that filter is "smart," your confidentiality guarantee is a probability.
|
||||
2. **Model outputs are proposals, not actions.**
|
||||
3. **Every side effect passes the gate.** There is no second door.
|
||||
4. **The harness itself flips no coins.** Replay a step with the model's answer and the tool results pinned, and behavior must be identical; any leftover variation is randomness *you* added and must be accounted for. The fine print: "deterministic" is conditional on pinned versions — a provider silently retraining the model behind the same API name changes the machine under you, and every dashboard number you collected dies with the version.
|
||||
|
||||
Notice what the rules don't say: they don't say the harness is *good*. A gate that approves everything satisfies rule 3 the way a lock that's always open satisfies "has a lock." The definition is a shape; the guarantees are what a particular harness *earns* inside it. Everything below is about what can be earned — and what can't.
|
||||
|
||||
And notice the symmetry between rules 1 and 3. There is exactly one door from your data into the model — what it may see — and exactly one door from the model into the world — what it may do. Nearly every security failure in these systems is one of those two doors with a hole in it: a secret lowered into a prompt that didn't need it, or a path from model text to a side effect that skipped the gate. Same bug, arrow flipped.
|
||||
|
||||
## Fail-closed, said precisely
|
||||
|
||||
"Fail-closed" gets used loosely. Here it means something exact: **nothing happens unless the gate said yes, and a refusal must itself be safe** — a refused proposal causes no side effect and leaves the run somewhere sane, which may be "stopped, having declined." The run is allowed to *say so*: a templated status message written by the shell is the shell speaking, not the model, and needs no gate. Failed runs don't have to die silent.
|
||||
|
||||
Three consequences people miss:
|
||||
|
||||
**Reads are not free.** A read-only call can smuggle instructions *in* (the fetched page is attacker-controlled) or secrets *out* (the URL it fetches can encode the payload). The gate approves calls, not just writes.
|
||||
|
||||
**Validation must not act.** A "validator" that resolves a URL, expands a template that fires a webhook, or evaluates an argument has already acted — inside the check. The gate must be pure: it reads the proposal and the memory and outputs yes or no. If deciding requires touching the world, that touch is itself an action and goes through the gate.
|
||||
|
||||
**Anything irreversible is decided at the gate.** The verifier can reject a bad *result*; it cannot unsend the email. So the question "can we take this back, and until when?" is asked before execution — which means each tool declares, up front, how reversible its effects are, and the gate reads that declaration when it decides; the mark that comes back in the result record is confirmation for the books, not the gate's source — the gate needed the answer before the tool ever ran.
|
||||
|
||||
Two honest asterisks. First, the gate checks a snapshot: it approves against the world *as its memory describes it*, and the world can move between check and commit. For actions that race the world — spend against a balance, write against a row — the tool itself must bind check to commit (compare-and-swap), or you have a classic time-of-check/time-of-use hole. The gate decides; for those effects, the tool enforces. Second, a gate is only as binding as the authority behind the tools. A tool process holding standing credentials — a database connection with every grant, an environment full of long-lived secrets — doesn't need the model's proposal to act, and against it the gate's "no" is a decision with nothing enforcing it. **A gate in front of an omnipotent tool is a suggestion.** The fix is to make the approval *be* the key: each authorized action carries a short-lived credential scoped to exactly that action, that resource, that operation, so tools hold no standing power at all.
|
||||
|
||||
## Why you don't get a proof — and what you do instead
|
||||
|
||||
If you write a sort function, you can prove it sorts: the function is small and the spec is exact. A harness has neither luxury. The spec side fails first — the task arrives in natural language, and natural language is, in the compiler's sense, *all undefined behavior*: there is no formal standard for "what the user meant" to verify against. The mechanism side fails next — the model is billions of learned parameters, and nobody can hand you a compact argument for why they jointly do the right thing.
|
||||
|
||||
Here is the careful version, because "you can't prove it" overshoots. The quantity you would want — call it the *expected steps to done* from any situation — is perfectly well-defined; in principle it exists. The document's central conjecture is that, for a model of this size, any faithful writing-down of that quantity is roughly *model-sized*: the honest proof-object does not compress. Find a small one and the conjecture dies — the document lists that outcome, explicitly, among the ways it could be wrong.
|
||||
|
||||
So instead of proving, you measure. You pick a progress meter — plan depth shrinking, open obligations closing, budget burning at the expected rate — and you check, across many runs, that it goes downhill and that its stalls predict failure. Two disciplines keep the measurement honest. The number bounds the world you *sampled*, never the world an adversary will choose: a meter calibrated on friendly traffic says nothing about hostile traffic. And the meter is itself attack surface: if "is the agent making progress?" is judged by another model, an attacker who can bend your agent can bend your *measurement of it* first, hiding the divergence from the very dashboard built to catch it. A learned meter is part of the system under test, never a neutral instrument.
|
||||
|
||||
A measurement is a risk metric. A proof is a certificate. Keeping those two words apart is half of what this theory is for.
|
||||
|
||||
## Security: reach the goal, avoid the danger — and who may change the rules
|
||||
|
||||
Formally, security here is a *reach-avoid* problem: reach a good stop, never touch the danger zone, **while an adversary picks the worst tool outputs your setup permits**. That last clause is the formal home of prompt injection: injection isn't "the model misbehaved," it's the environment optimized to bend your loop — poisoned pages, malicious tool descriptions, crafted responses.
|
||||
|
||||
Two different numbers fall out here, and dashboards love to collapse them: *success* (reached an accepted end before anything went wrong — a safe refusal counts against it) and *safety* (never touched the danger zone — a safe refusal is perfectly safe). Track both. They move independently. And both are scored by your own stop rule — they count what the shell *declared* a success. Whether a declared success was actually *right* is a third, harder number that no dashboard inside the system can produce; only a judge outside the run — a test suite, an audit, ground truth — can.
|
||||
|
||||
The gate handles the visible half of injection: the model, freshly poisoned, proposes emailing your credentials somewhere, and the gate refuses — and injection or not, the action does not happen. But the deeper attack doesn't propose a bad action today. It rewrites *what the run believes its job is* — it edits the plan — and then every future action looks locally reasonable against a corrupted plan. So memory has to be partitioned: **data** (tool results, fetched pages, retrieved documents — content the world supplied) and **control** (the plan, the permissions, what is authorized next). The security claim is conditional on that partition holding: untrusted content lands in data, always. And "trust" is really two questions pointing opposite ways, which is worth keeping straight: *can this leak?* (a value is as secret as the most-secret thing that fed it — secrecy flows **upward**) and *can this boss us around?* (a value is as trustworthy as the least-trustworthy thing that fed it — authority flows **downward**). Untrusted content is safe as *data* precisely because the second question keeps it off the control side; a secret is kept out of the model by the first. Lowering either barrier on purpose — declassifying a secret, promoting data to trusted — is an explicit decision the owner makes, never a thing that happens by accident when two values are combined.
|
||||
|
||||
Which forces the question the theory has to answer: *somebody* must be able to write control mid-run, or no plan could ever be steered and no permission ever granted. The answer is a small hierarchy with a top the model can't reach. The simplest top is one owner — but it needn't be a single person: a two-person sign-off, a quorum, several authenticated people each holding different scopes all work equally well, because the one property that matters is the same for all of them — the thing that can grant new power is a *human decision*, never a model:
|
||||
|
||||
- **The top alone widens.** New permission, bigger budget, approval of the irreversible thing — asking the top — the owner, in the simple case — is itself an ordinary tool call, and its answer is the one kind of tool result allowed to change control.
|
||||
- **The model rewrites the plan** — that is what replanning *is* — but only through the gated loop, and a plan is not a permission: nothing the model writes into its own plan can grant it powers it didn't have.
|
||||
- **Everything else is data.** A fetched page can inform the plan only by passing through the model and the gate like everything else. It can suggest. It cannot promote itself to boss.
|
||||
- **AI judges only tighten.** Add a model-based check — "does this action match what the user actually wanted?" — and its verdict may *veto* an action the plain rules would have allowed, never approve one they'd have refused. A judge that can approve is a tricked judge that can open the vault. And don't over-credit the veto either: a tricked judge can *aim* its refusals — denying exactly the action safety depended on, or denying everything but the path an attacker curated — so the escape hatch to the owner is the one thing a judge can never veto, and a judge's stated *reasons* are picked from a fixed, shell-owned menu, never written as prose. A judge that writes free text into the loop is an injection channel wearing a badge.
|
||||
|
||||
One more rule closes the loop: transformations don't launder trust. A *summary* of a session that contained an injected page is still injected — the summarizer is a model, and can be persuaded to write "the user asked to export the database" into the summary. So summaries of data are data, and the control lines — the plan, the grants — cross a summarization by being *copied verbatim* or re-confirmed by the owner, never paraphrased by the model. Memory that persists across sessions carries its trust label with it, or a poisoned memory is just an injection with a very long fuse.
|
||||
|
||||
## Operations: the rules you feel on Tuesday at 3 a.m.
|
||||
|
||||
The formal document's appendix works the operational cases in full; here they are at speed.
|
||||
|
||||
**The ledger, and the three-way distinction that keeps it honest.** Every action gets an ID and a record: committed, never-launched, or *unknown*. "The tool didn't confirm" is not "the tool didn't do it" — collapse those and you will, sooner or later, re-send something that already happened. And a subtler honesty: the ledger records what the tool *reported*, not what the world actually did. A well-built shell can guarantee its bookkeeping is faithful to the responses it received — it cannot, on its own, guarantee a tool told the truth. A tool that returns a clean "done!" for something it never did puts a clean "done!" in your ledger. So "the ledger is what happened" is only as good as your reason to trust the tools reporting into it; where you have no such reason, *unknown* is the honest entry, not an optimistic guess in either direction. The double-send bug has one reliable cure: **journal before dispatch.** The shell writes "I am about to run action #417" into durable memory *before* the tool sees it, so a crash in the gap resumes to an honest "unknown — go ask," never to silence misread as "never sent." Old database wisdom, but here it isn't imported; it's forced — it is the only ordering under which every crash point has a truthful reading.
|
||||
|
||||
**Crashes aren't finishes.** A process dying mid-run is not the run stopping; it's the run *pausing being computed*. Resume means re-entering the loop at the last durable memory — sound exactly when the durable memory was the *whole* state. Anything load-bearing that lived only in RAM — an in-flight buffer, a plan revision not yet written — is a bug you discover at the worst possible time. Recovery is where you find out whether your state was really your state. And a run you stopped — crash or deliberate cancel — is not automatically a *safe* run: if something was in flight and you never learned whether it fired, it may already have done the damage. "We stopped in time" is only true when everything in flight resolved to something safe; an outstanding *unknown* has to be treated as possibly-bad, the same optimism the ledger warns against, one level up.
|
||||
|
||||
**Two innocent actions can be guilty together.** Models emit several tool calls per turn. "Read the secret" passes review. "Post to the web" passes review. The pair is an exfiltration channel — so the gate authorizes the *set*, atomically, with the interactions checked, not each element in isolation.
|
||||
|
||||
**Sub-agents are just fancy tools.** An agent that spawns another agent is, from the parent's chair, calling a tool: the spawn is gated, the budget is part of the deal, and the child's whole run comes back as one result carrying the child's ledger. Two laws travel down the tree: budgets subdivide, and **authority only narrows** — a child holds at most a subset of its parent's permissions, and a child's request beyond those grants routes *up*, ultimately to the owner, because a parent inventing an approval it never held is the tricked-judge case wearing a manager's badge. A corollary worth framing: a *fully autonomous* run is one whose owner is unreachable — meaning the only channel that can ever widen anything is closed, and its permissions are frozen at launch. That is not a limitation of the theory. That is what the word "autonomous" costs.
|
||||
|
||||
**Keep the originals.** When the transcript outgrows the prompt and you summarize it down, deleting the original is an irreversible act against your own state — and irreversible acts are gate decisions, self-directed or not. Keep originals content-addressed; let the summary be an index, re-derivable, auditable. A summary you can check against its source is a note. A summary that replaced its source is a fait accompli.
|
||||
|
||||
## Robots that never clock out — and robots that assign their own work
|
||||
|
||||
Everything so far assumed a job that *ends*: you ask, the robot does it, you read the result. Two steps past that are where the interesting failures live, and they're the same idea one level bigger each time.
|
||||
|
||||
**The robot that never clocks out (a daemon).** A monitor, a coordinator, a service — it isn't supposed to finish; it's supposed to keep going, wake on events, do a bit of work, go back to waiting. The clean way to think about it: each wake-work-rest cycle is one ordinary run, and the daemon is just those runs chained end to end forever. That reframing is free — but it comes with a bill nobody likes. **Safety that's fine per cycle rots over many cycles.** A 99.99%-safe cycle sounds bulletproof; run it ten thousand times and you're at about a coin-flip of having touched the danger zone at least once. So a long-running robot's safety isn't a fixed wall, it's a slow leak — which means the antidote isn't a better wall, it's *scheduled resets*: the owner re-confirming, credentials rotating, memory getting audited and re-summarized against the originals. Housekeeping isn't housekeeping; it's the thing that keeps the safety math from decaying. And the slow-leak logic is exactly where slow attacks live — a poisoned note dropped into memory on Monday and read back into the plan on Friday is an injection with a long fuse. So the trust label on a piece of information has to survive across cycles, not just within one. One more wrinkle: a daemon drifts in and out of your reach. While you're around, it can escalate to you; while you're not, "escalate to the owner" isn't available — so the one thing it must always be able to do instead is *stop*. A robot that can be tricked into refusing everything, and can't reach you, had better be able to halt rather than be steered.
|
||||
|
||||
**The robot that assigns its own work (the loop).** Step back one more time. Above the robot that *does* a task sits a system that decides *which task is next* — scans the backlog, picks one, launches the robot at it, checks the result, remembers, fires again. This is the thing people mean in 2026 when they say they've stopped prompting their agents and started writing *loops* that prompt them: you design the assigner once, and it runs the doer for you while you sleep. The honest observation — and the reason this document bothers with it — is that the assigner is *not a new kind of thing*. It's the same harness, one level up: it has its own memory (the backlog), its own gate (**who let the loop refactor the auth module at 3 a.m.?**), its own verifier, and its own two walls. Every rule from the inner robot recurs on the outer one — including the uncomfortable ones. There's still no proof it stays out of trouble over a long night; there's only a measured progress meter, with the same catch that a *learned* meter can be fooled. And the origin story of the whole trend is the cautionary case in miniature: the famous first version was literally the same prompt in a `while` loop until the tests passed — which is the empty gate, the always-open lock, one level up. It works beautifully right up until the tests weren't checking the thing that mattered. The loop doesn't delete the hard problems. It moves them up a floor, where they're bigger and you're further away.
|
||||
|
||||
The pattern, if you want the whole thing in one line: *words, context, robot, loop* are four sizes of the same object, and every promise in this document lives in the whole assembled thing — never in any one layer by itself.
|
||||
|
||||
## The two walls
|
||||
|
||||
Two limits are structural. You don't fix them with a better harness; you design around them.
|
||||
|
||||
**The desk.** The model can hold only so much *in mind at once* — the context window. Files, databases, and search extend what it can *look up*, not what it can hold: every lookup still passes through the same small window to touch actual computation. The shell can page; the model cannot grow its desk. Tasks whose irreducible working set exceeds the desk don't fail loudly — they fail by forgetting the middle (the well-documented "lost in the middle" effect is this wall showing through the paint).
|
||||
|
||||
**The dictionary.** The model's knowledge is frozen into its parameters at training time — and the proof problem above is conjectured to live at that same scale: the certificate wouldn't fit anywhere smaller than the brain it certifies. The two walls trade against each other along the training-versus-inference axis — bigger dictionary or bigger desk — directionally, and at no clean exchange rate.
|
||||
|
||||
## How this could be wrong
|
||||
|
||||
This is a hypothesis, and it says out loud what would kill it. The tests, in plain terms:
|
||||
|
||||
- **The replay test.** Rerun with model answers and tool results pinned. Any leftover variation — timestamps, wall-clocks, and cache expiries are the classic leaks — falsifies "the harness adds no randomness" until accounted for.
|
||||
- **The drop-a-variable test.** Remove something from memory; if behavior statistics shift, the memory wasn't complete. The crash-resume version of the same test: if resuming from saved state breaks, the saved state wasn't the state.
|
||||
- **Does the meter mean anything?** If no reasonable progress meter's drift predicts real failures — across the natural families, not just one bad candidate — the whole "measure what you can't prove" program is empty.
|
||||
- **The red-team test.** Swap sampled tool outputs for worst-case ones: injected pages, poisoned metadata, malformed replies. The design must survive the worst permitted world, not the average one.
|
||||
- **Gates versus begging.** The theory predicts deterministic gating beats prompt-level pleading. If "please be careful" alone matches real gates on security outcomes, the controller-versus-model story is wrong.
|
||||
- **The compression hunt.** Exhibit a compact, provably sound progress certificate for a frontier-scale model on a nontrivial task family, and the central conjecture falls — constructively.
|
||||
- **The desk probe.** Take a task family with a *proven* memory floor — so "it needed the whole picture at once" is someone else's theorem, not our excuse — scale it past the window, and watch: the wall predicts a *ceiling*, not a cliff — past the boundary, a success rate that stays capped no matter how many retries you buy. A family solved reliably out there, without new shell tricks for splitting the work, kills the wall.
|
||||
|
||||
## Who else landed here
|
||||
|
||||
The formal document keeps three honesty tiers. **Borrowed**: real theorems, cited — the drift and stopping-time mathematics is classical, and the very architecture of a deterministic supervisor gating a plant it didn't author is 1987 control theory; the shape is older than the web. **Ours**: the modeling choices and the conjectures — the walls, the incompressibility claim, the design rules — organizing principles, not results. **Corroborated**: pieces of the same object reached independently by people who never saw this framing — capability-security work isolating control flow from untrusted data (CaMeL), reinforcement-learning "shields" filtering a learned policy's actions through a deterministic checker, verification work that states the "learned safeguards can't certify" gap as its opening motivation, and architecture patterns converging on plan-then-execute. Even the field's live disagreement — provable-but-rigid deterministic layers versus flexible-but-uncertifiable learned checks — is, in this frame, not a fight but a placement: you need both, on their proper sides of the irreversibility line, with the learned one permitted only to tighten.
|
||||
|
||||
## What to remember
|
||||
|
||||
The model proposes; the gate disposes. No is the default, and a refusal must be safe. Only the top of the trust hierarchy widens permissions — a human decision, never the model, a tool result, a summary, or a judge. "Didn't confirm" is not "didn't happen." The desk is finite and the proof doesn't compress, so you measure — and you say *measurement* when you mean measurement. A robot that never stops leaks safety slowly, so it needs scheduled resets — and when it can't reach you, it must be able to stop. A loop that runs robots for you is just a bigger robot with the same rules and a further-away owner. And all of it is a hypothesis wearing its own kill-conditions on its sleeve.
|
||||
|
||||
The formal version — the objects, the certificates, the falsifiers, the citations — is [HYPOTHESIS.md](HYPOTHESIS.md). It wins every disagreement with this file, including this sentence.
|
||||
|
||||
*Same ramblings, fewer symbols.*
|
||||
@@ -5,7 +5,6 @@
|
||||
[](https://pypi.org/project/turnstone/)
|
||||
[](LICENSE)
|
||||
[](https://discord.gg/Nh3bWMacaq)
|
||||
[](https://github.com/sponsors/eous)
|
||||
|
||||
Self-hosted, local-first orchestration for tool-using AI agents. Give LLMs real tools — shell, files, search, web — and run them across your own cluster with direct HTTP routing and interactive interfaces. Your code, your models, your data stay on hardware you control: no telemetry, no phone-home.
|
||||
|
||||
@@ -21,7 +20,7 @@ Named after the [Ruddy Turnstone](https://en.wikipedia.org/wiki/Ruddy_turnstone)
|
||||
ℋ : s_{n+1} ~ T(s_n) for n < τ*, T = ρ ∘ (M_W ∘ π, E)
|
||||
```
|
||||
|
||||
[**the primer →**](PRIMER.md)
|
||||
[**the hypothesis →**](HYPOTHESIS.md)
|
||||
|
||||
### Release Tracks
|
||||
|
||||
@@ -125,8 +124,7 @@ Built-in tools for shell, files, search, web, memory, notifications, and autonom
|
||||
| `turnstone-console` | Cluster dashboard + routing proxy + admin panel |
|
||||
| `turnstone-channel` | Channel gateway (Discord and Slack adapters) |
|
||||
| `turnstone-admin` | User/token management CLI |
|
||||
| `turnstone-eval` | Headless measurement — scores tool-use against expected actions |
|
||||
| `turnstone-optimizer` | Prompt/tool optimizer (UCB self-modify loop over the eval substrate) |
|
||||
| `turnstone-eval` | Eval harness for prompt/tool optimization |
|
||||
| `turnstone-doctor` | LLM-backed cluster diagnostics |
|
||||
|
||||
### Diagrams
|
||||
@@ -172,14 +170,6 @@ UML diagrams in [`docs/diagrams/`](docs/diagrams/):
|
||||
- Optional: Discord / Slack channel integrations (`pip install turnstone[discord,slack]`)
|
||||
- [Git LFS](https://git-lfs.com/) for cloning (diagram PNGs)
|
||||
|
||||
## Support
|
||||
|
||||
Turnstone is free, Apache-2.0, and self-hosted — no paid tier, no telemetry, no upsell. If it saves you time or you'd like to help keep development moving, you can sponsor the project:
|
||||
|
||||
**[❤ Sponsor Turnstone →](https://github.com/sponsors/eous)** · one-off via **[PayPal](https://paypal.me/eousphoros)**
|
||||
|
||||
Sponsorship is entirely optional and funds maintenance, new features, and infrastructure. Prefer to contribute in other ways? Filing issues, improving docs, and [pull requests](CONTRIBUTING.md) help just as much.
|
||||
|
||||
## Community
|
||||
|
||||
Questions, ideas, or want to show what you're building? Join us on Discord:
|
||||
|
||||
@@ -7,6 +7,6 @@ appVersion: "0.3.0"
|
||||
|
||||
dependencies:
|
||||
- name: postgresql
|
||||
version: ~18.8.0
|
||||
version: ~18.7.0
|
||||
repository: https://charts.bitnami.com/bitnami
|
||||
condition: postgresql.enabled
|
||||
|
||||
+29
-171
@@ -458,7 +458,7 @@ Each item in `items` (shared by `tool_info` and `approve_request`):
|
||||
| `context_window` | int | Total context window size in tokens |
|
||||
| `pct` | float | Percentage of context window used |
|
||||
| `effort` | string | Reasoning effort level (`low`/`medium`/`high`) |
|
||||
| `cache_creation_tokens` | int | Tokens written to prompt cache (Anthropic + OpenAI) |
|
||||
| `cache_creation_tokens` | int | Tokens written to prompt cache (Anthropic) |
|
||||
| `cache_read_tokens` | int | Tokens served from prompt cache (Anthropic + OpenAI) |
|
||||
|
||||
**`info`** -- an informational message (e.g. command output).
|
||||
@@ -467,46 +467,6 @@ Each item in `items` (shared by `tool_info` and `approve_request`):
|
||||
{"type": "info", "message": "Session cleared."}
|
||||
```
|
||||
|
||||
**`compaction`** -- context-compaction lifecycle (manual `/compact` and
|
||||
auto-compaction). `phase: "start"` opens the operation (`trigger` is
|
||||
`"manual"` or `"auto"`; auto adds `where` — e.g. `"mid-turn"` — and, when
|
||||
the percentage threshold actually fired, `pct`; the context-overflow retry
|
||||
path compacts without a `pct` since no threshold was evaluated).
|
||||
`phase: "progress"` reports chunked summarization (`part`/`total`/`depth`,
|
||||
where depth 0 summarizes transcript batches and deeper levels merge partial
|
||||
summaries), a transient-error retry wait (`retry_in` seconds + `error`), or
|
||||
`warning: "summary_truncated"`. `phase: "end"` settles it: `ok: true`
|
||||
carries `before_tokens`/`after_tokens` and the produced `summary`;
|
||||
`ok: false` carries a `reason`
|
||||
(`"not_enough_messages"` / `"irreducible"` / `"empty_summary"` /
|
||||
`"cancelled"` / `"error"`) and a human-readable `message` — for
|
||||
`reason: "error"` the same message is also emitted as a paired typed
|
||||
`error` event (that is the renderable error surface; the end event is
|
||||
card-teardown). Failed ends also carry `notice`: the emitter-computed
|
||||
display verdict — show `message` only when it is `true` (the server
|
||||
suppresses error-reason, superseded, and cancelled-auto notices once,
|
||||
centrally, so clients don't re-derive that policy). Every end (ok or
|
||||
failed) carries `trigger`, and every event carries `compaction_id` — an
|
||||
opaque integer correlating the start/progress/end of one compaction run (a
|
||||
client that force-stopped one compaction can use it to ignore stragglers
|
||||
from the abandoned run). End events also carry `superseded`: `true` marks
|
||||
a force-abandoned compaction retiring after a successor generation took
|
||||
over (an OK end's result card still stands: the history swap happened).
|
||||
Superseded start/progress events are never emitted.
|
||||
Exactly one `start` and one `end` are emitted per attempt,
|
||||
so clients can key an in-progress affordance (progress bar) on the pair. A
|
||||
successful end is also persisted: the summary replays from `/history` as a
|
||||
`role: "system"`, `source: "compaction"` entry whose `meta` carries
|
||||
`{watermark, before_tokens, after_tokens, trigger}` and whose `event_id`
|
||||
matches the end event's id (dedup across repaint + replay).
|
||||
|
||||
```json
|
||||
{"type": "compaction", "phase": "start", "compaction_id": 7, "trigger": "auto", "where": "mid-turn", "pct": 80}
|
||||
{"type": "compaction", "phase": "progress", "compaction_id": 7, "part": 2, "total": 5, "depth": 0}
|
||||
{"type": "compaction", "phase": "end", "ok": true, "compaction_id": 7, "trigger": "auto",
|
||||
"before_tokens": 128400, "after_tokens": 9200, "summary": "## Decisions\n..."}
|
||||
```
|
||||
|
||||
**`error`** -- an error message.
|
||||
|
||||
```json
|
||||
@@ -738,42 +698,6 @@ Each skill summary:
|
||||
|
||||
---
|
||||
|
||||
### `GET /v1/api/personas`
|
||||
|
||||
Returns the enabled personas offered by the workstream-creation pickers.
|
||||
Authenticated for any logged-in user and deliberately gated by **no**
|
||||
`persona.*` permission — selecting a persona at creation is a user
|
||||
action, while the `persona.*` perms gate authoring. Display fields only;
|
||||
the levers (base prompt, tool set, MCP/memory toggles) stay server-side.
|
||||
|
||||
**Response:**
|
||||
|
||||
```json
|
||||
{
|
||||
"personas": [
|
||||
{"name": "engineer", "display_name": "Engineer", "description": "The stock interactive workstream: full tools, MCP, and memory.", "applies_to_kinds": ["interactive"], "is_default": true},
|
||||
{"name": "researcher", "display_name": "Researcher", "description": "Answers questions with evidence — reads and cites, loads tools to verify when needed.", "applies_to_kinds": ["interactive"], "is_default": false}
|
||||
],
|
||||
"total": 2
|
||||
}
|
||||
```
|
||||
|
||||
Each persona summary:
|
||||
|
||||
| Field | Type | Description |
|
||||
|--------------------|--------|------------------------------------------------------------------|
|
||||
| `name` | string | Persona slug (used in the `persona` field on workstream creation) |
|
||||
| `display_name` | string | Human-readable label for pickers |
|
||||
| `description` | string | Short description of the persona's intent |
|
||||
| `applies_to_kinds` | array | Workstream kinds the persona applies to (`interactive` / `coordinator`) |
|
||||
| `is_default` | bool | Whether this is the default persona for its kind |
|
||||
|
||||
> **Note:** For full persona management (create, edit, archive), use the
|
||||
> admin endpoints at `/v1/api/admin/personas` (requires the
|
||||
> `persona.{create,read,write}` permissions).
|
||||
|
||||
---
|
||||
|
||||
### `POST /v1/api/workstreams/{ws_id}/send`
|
||||
|
||||
Sends a user message to a workstream. Spawns a daemon worker thread that calls
|
||||
@@ -788,41 +712,32 @@ Sends a user message to a workstream. Spawns a daemon worker thread that calls
|
||||
**Request body:**
|
||||
|
||||
```json
|
||||
{"message": "Explain how the server works", "attachment_ids": ["a1"]}
|
||||
{"message": "Explain how the server works"}
|
||||
```
|
||||
|
||||
| Field | Type | Required | Description |
|
||||
|------------------|------------|----------|------------------------------------------------------|
|
||||
| `message` | string | yes | The user's message text |
|
||||
| `attachment_ids` | string[] | no | Staged uploads to attach (omit = auto-consume; `[]` = none) |
|
||||
| Field | Type | Required | Description |
|
||||
|-----------|--------|----------|-------------------------|
|
||||
| `message` | string | yes | The user's message text |
|
||||
|
||||
**Response.** Every 200 body carries `attached_ids` and
|
||||
`dropped_attachment_ids` (empty lists when no attachments are involved):
|
||||
**Response (success):**
|
||||
|
||||
- `{"status": "ok", ...}` — a fresh turn was dispatched.
|
||||
- `{"status": "queued", "priority", "msg_id", ...}` — folded into the live
|
||||
turn's interjection queue; delivered at the next tool-result seam.
|
||||
`DELETE .../send` with the `msg_id` retracts it before delivery.
|
||||
- `{"status": "queued", "deferred": true, ...}` — parked on the deferred-send
|
||||
list (a command window holds the slot, or earlier deferred sends are
|
||||
pending) and dispatched as its own full-fidelity send afterwards; see the
|
||||
defer contract under `POST /v1/api/command`.
|
||||
- `{"status": "queue_full", ...}` — the send was refused with retry-shortly
|
||||
semantics: the live worker's interjection queue is at capacity, the
|
||||
deferred-send list hit its saturation bound (10 pending — the same
|
||||
backpressure contract), or the deferred-send drain could not be started
|
||||
under resource exhaustion (the message was **not** accepted; nothing is
|
||||
parked).
|
||||
- `{"status": "attachments_busy", ...}` — attachments can't ride a queued
|
||||
turn; the staged uploads survive for a retry once the worker idles.
|
||||
```json
|
||||
{"status": "ok"}
|
||||
```
|
||||
|
||||
**Response (busy):** Returned if the workstream's worker thread is still alive
|
||||
from a previous request. Also pushes a `busy_error` event to the SSE stream.
|
||||
|
||||
```json
|
||||
{"status": "busy"}
|
||||
```
|
||||
|
||||
**Error responses:**
|
||||
|
||||
| Status | Body | Condition |
|
||||
|--------|-------------------------------------------------|----------------------------------------|
|
||||
| 400 | `{"error": "message is required"}` | Message is empty |
|
||||
| 404 | `{"error": "Unknown workstream"}` | `ws_id` not found (or closed mid-send) |
|
||||
| 409 | `{"status": "cross_user_interjection", ...}` | Another participant's turn is in flight |
|
||||
| Status | Body | Condition |
|
||||
|--------|------------------------------------|------------------------|
|
||||
| 400 | `{"error": "Empty message"}` | Message is empty |
|
||||
| 404 | `{"error": "Unknown workstream"}` | `ws_id` not found |
|
||||
|
||||
---
|
||||
|
||||
@@ -865,56 +780,7 @@ automatically approved without prompting.
|
||||
|
||||
### `POST /v1/api/command`
|
||||
|
||||
Executes a slash command in the given workstream. Commands run on the
|
||||
workstream's worker slot (mutual exclusion against sends, a running
|
||||
compaction, and each other) — the endpoint is **not** unconditionally
|
||||
synchronous:
|
||||
|
||||
- **Quick commands** (everything except `/compact`): the endpoint waits for
|
||||
completion, so `{"status": "ok"}` means the command ran. A command still
|
||||
running after 25 s answers `{"status": "running"}` — the worker keeps
|
||||
going, its output reaches the pane via SSE, and the post-command pane
|
||||
refreshes below still fire when it completes. (The bound sits under
|
||||
common 30 s client/proxy timeouts — the console proxy's included — so
|
||||
the degraded answer actually reaches bounded callers.)
|
||||
- **`/compact`**: dispatched fire-and-forget — `{"status": "ok"}` means the
|
||||
compaction *started*. A large context can legitimately compact for many
|
||||
minutes; progress streams as `compaction` SSE events (see the event
|
||||
reference) and the persisted marker row lands on completion. Do not read
|
||||
`/history` expecting the compacted transcript immediately after the
|
||||
response.
|
||||
- **Busy refusal**: if a turn or another command holds the worker slot, the
|
||||
command is refused with HTTP **409** `{"status": "busy", "error": ...}` and
|
||||
did **not** run. Retry after the current turn finishes. (The old inline
|
||||
endpoint executed commands unconditionally mid-turn; the 409 makes the
|
||||
refusal loud for callers that only check the HTTP status.)
|
||||
|
||||
While a command holds the slot — and afterwards, while earlier deferred
|
||||
sends are still waiting (the pending list is the order authority: a fresh
|
||||
send never overtakes a message already acknowledged) — `POST .../send`
|
||||
requests are **deferred**: the server answers `{"status": "queued",
|
||||
"deferred": true, "msg_id": ...}` immediately and dispatches the message
|
||||
as an ordinary full-fidelity send (attachments and sender identity
|
||||
included) in arrival order once the slot frees — it is never routed
|
||||
through the mid-turn interjection queue (no length cap, no cross-user
|
||||
rejection). The response arrives within normal round-trip time, so
|
||||
timeout-bounded clients (SDKs, proxies, the coordinator) need no special
|
||||
handling. To retract a deferred send before it dispatches, issue the same
|
||||
`DELETE .../send` with its `msg_id` used for queued interjections —
|
||||
`{"status": "removed"}` confirms it will not dispatch; `"not_found"` means
|
||||
it already dispatched (or is dispatching). Retracting a deferred send
|
||||
discards any attachments it carried; re-attach to send them again. When a
|
||||
deferred send dispatches, panes receive a `message_dispatched` event
|
||||
(`msg_id`, plus `folded: true` when it folded into a live turn's
|
||||
interjection queue rather than spawning its own turn) so queued-message
|
||||
UI can settle the right way.
|
||||
|
||||
Durability: deferred sends are **node-local and in-memory** (the same
|
||||
lifetime as the interjection queue). `"queued"` is at-most-once intake, not
|
||||
durable acceptance — if the workstream is closed or the node restarts before
|
||||
the window ends, the message is dropped. Anything that must survive a
|
||||
restart should be re-sent after confirming dispatch (the turn appears on the
|
||||
SSE stream / in `/history`).
|
||||
Executes a slash command in the given workstream.
|
||||
|
||||
**Request body:**
|
||||
|
||||
@@ -927,12 +793,10 @@ SSE stream / in `/history`).
|
||||
| `command` | string | yes | The slash command (e.g. `/clear`) |
|
||||
| `ws_id` | string | yes | Target workstream ID |
|
||||
|
||||
If the command is `/clear`, `/new`, or `/resume`, the server pushes a
|
||||
`clear_ui` SSE event to instruct the client to reset its message display and
|
||||
re-fetch the transcript via `GET .../history` (there is no SSE event that
|
||||
carries the messages themselves). These follow-ups are emitted by the
|
||||
command worker itself, so they fire even when the endpoint already answered
|
||||
`{"status": "running"}`.
|
||||
If the command is `/clear` or `/new`, the server pushes a `clear_ui` SSE event
|
||||
to instruct the client to reset its message display. If the command is
|
||||
`/resume`, the server pushes `clear_ui` followed by a `history` event
|
||||
containing the resumed session's messages.
|
||||
|
||||
**Response:**
|
||||
|
||||
@@ -940,16 +804,12 @@ command worker itself, so they fire even when the endpoint already answered
|
||||
{"status": "ok"}
|
||||
```
|
||||
|
||||
or `{"status": "running"}` as above.
|
||||
|
||||
**Error responses:**
|
||||
|
||||
| Status | Body | Condition |
|
||||
|--------|-------------------------------------|--------------------------------------------------|
|
||||
| 400 | `{"error": "Empty command"}` | Command is empty |
|
||||
| 404 | `{"error": "Unknown workstream"}` | `ws_id` not found |
|
||||
| 409 | `{"status": "busy", "error": ...}` | A turn/command holds the worker |
|
||||
| 503 | `{"status": "error", "error": ...}` | The command worker could not be started (resource exhaustion) — the command did **not** run; retry shortly |
|
||||
| Status | Body | Condition |
|
||||
|--------|------------------------------------|----------------------|
|
||||
| 400 | `{"error": "Empty command"}` | Command is empty |
|
||||
| 404 | `{"error": "Unknown workstream"}` | `ws_id` not found |
|
||||
|
||||
---
|
||||
|
||||
@@ -1035,7 +895,6 @@ All fields are optional. The body can be empty or an empty JSON object.
|
||||
| `auto_approve` | bool | false | Auto-approve all tool calls for this workstream |
|
||||
| `resume_ws` | string | "" | Workstream ID to resume atomically during creation (empty = fresh)|
|
||||
| `skill` | string | "" | Skill name. Applies content (system prompt), model, temperature, reasoning effort, max tokens, auto-approve policy, token budget, and other session config from the skill. Returns 400 if not found or disabled. Ignored when `resume_ws` is set (resumed sessions restore their own skill). |
|
||||
| `persona` | string | "" | Persona slug. Resolved and snapshotted into the workstream at creation; empty selects the kind's default. |
|
||||
| `judge_model` | string | "" | Optional model alias for the judge (overrides default judge model for this workstream) |
|
||||
|
||||
> **Skill behavior:** When `skill` is specified, the skill's content is injected as a system message and its session config fields (model, temperature, auto-approve, token budget, etc.) override system defaults for the new workstream.
|
||||
@@ -1052,7 +911,6 @@ All fields are optional. The body can be empty or an empty JSON object.
|
||||
| `name` | string | Auto-generated workstream name |
|
||||
| `resumed` | bool | Whether a previous session was successfully resumed |
|
||||
| `message_count` | int | Number of messages in the resumed session (0 if fresh) |
|
||||
| `initial_message_status` | string | Present ONLY when the workstream was created but its `initial_message` could not be delivered: `"queue_full"` (a raced live worker's interjection queue was at capacity — resend via `/send`; any uploads stay staged) or `"refused_closed"` (the workstream was closed mid-create). Absent whenever the message was dispatched. |
|
||||
|
||||
**Error (limit reached):**
|
||||
|
||||
|
||||
+37
-148
@@ -19,8 +19,7 @@ plugs in.
|
||||
| `turnstone` | `turnstone.cli` | `TerminalUI` | Interactive terminal REPL |
|
||||
| `turnstone-server` | `turnstone.server` | `WebUI` | Browser-based chat (HTTP + SSE) |
|
||||
| `turnstone-console` | `turnstone.console.server` | ClusterCollector | Cluster dashboard (aggregates all nodes) |
|
||||
| `turnstone-eval` | `turnstone.eval.cli` | `NullUI` | Headless measurement (scores tool-use against expected actions) |
|
||||
| `turnstone-optimizer` | `turnstone.optimizer` | `NullUI` | Prompt/tool optimization (UCB self-modify loop over the eval substrate) |
|
||||
| `turnstone-eval` | `turnstone.eval` | `NullUI` | Headless evaluation and prompt optimization |
|
||||
| `turnstone-channel` | `turnstone.channels.cli` | ChannelAdapter | Channel gateway (Discord, Slack, etc.) |
|
||||
| `turnstone-admin` | `turnstone.admin` | — | Offline user and API token management |
|
||||
| `turnstone-doctor` | `turnstone.doctor` | — | LLM-backed cluster diagnostics |
|
||||
@@ -91,7 +90,7 @@ turnstone/
|
||||
discord/ Discord adapter (bot, cog, views, streaming, config)
|
||||
slack/ Slack adapter (Socket Mode bot, DM routing, approval buttons)
|
||||
shared_static/ Shared design system (base.css, auth.js, theme.js, toast.js, utils.js, kb.js)
|
||||
katex-0.18.1/ Vendored KaTeX math rendering library (MIT, woff2 fonts)
|
||||
katex-0.17.0/ Vendored KaTeX math rendering library (MIT, woff2 fonts)
|
||||
ui/
|
||||
colors.py ANSI color constants with NO_COLOR support
|
||||
markdown.py Streaming terminal markdown renderer (line-buffered)
|
||||
@@ -268,7 +267,7 @@ the per-workstream events stream in
|
||||
|-------|--------|-------|
|
||||
| `TerminalUI` | `turnstone.cli` | ANSI colors, `MarkdownRenderer`, `Spinner`, readline-based `input()` for approval |
|
||||
| `WebUI` | `turnstone.server` | SSE event queue per workstream + global broadcast, `threading.Event` for blocking on approval. `on_state_change` sends to both per-workstream and global SSE (the browser UI uses per-workstream `state_change` events to manage busy/idle transitions; `stream_end` only finalizes markdown rendering). |
|
||||
| `NullUI` | `turnstone.eval.core` | Discards all output; `approve_tools` always returns `(True, None)` |
|
||||
| `NullUI` | `turnstone.eval` | Discards all output; `approve_tools` always returns `(True, None)` |
|
||||
|
||||
### WorkstreamTerminalUI
|
||||
|
||||
@@ -609,7 +608,8 @@ LLMProvider (protocol)
|
||||
|
||||
| Method | Purpose |
|
||||
|--------|---------|
|
||||
| `create_streaming()` | The one transport: streaming request, yields normalized `StreamChunk` objects (single-shot callers accumulate via `drain_stream()` into a `CompletionResult`) |
|
||||
| `create_streaming()` | Streaming request, yields normalized `StreamChunk` objects |
|
||||
| `create_completion()` | Non-streaming request, returns `CompletionResult` |
|
||||
| `get_capabilities()` | Per-model flags (`ModelCapabilities`) |
|
||||
| `convert_tools()` | Translate OpenAI tool schemas to provider format |
|
||||
| `retryable_error_names` | Exception class names that trigger retry |
|
||||
@@ -621,30 +621,20 @@ LLMProvider (protocol)
|
||||
|------|--------|
|
||||
| `StreamChunk` | `content_delta`, `reasoning_delta`, `tool_call_deltas`, `info_delta`, `usage`, `finish_reason`, `provider_blocks` |
|
||||
| `CompletionResult` | `content`, `tool_calls`, `finish_reason`, `usage`, `provider_blocks` |
|
||||
| `ModelCapabilities` | `context_window`, `max_output_tokens`, `supports_temperature`, `token_param`, `thinking_mode`, `supports_effort`, `supports_web_search`, `supports_tool_search`, `supports_vision`, `supports_reasoning_replay`, `supports_verbosity`, `verbosity`, `supports_pro_mode`, `reasoning_mode` |
|
||||
| `ModelCapabilities` | `context_window`, `max_output_tokens`, `supports_temperature`, `token_param`, `thinking_mode`, `supports_effort`, `supports_web_search`, `supports_tool_search`, `supports_vision`, `supports_reasoning_replay` |
|
||||
| `UsageInfo` | `prompt_tokens`, `completion_tokens`, `total_tokens`, `cache_creation_tokens`, `cache_read_tokens` |
|
||||
|
||||
**OpenAIProvider** (`_openai.py`): passes messages through unchanged (they are
|
||||
already in OpenAI format), including multi-part content blocks (text + images)
|
||||
in tool results. Model capability lookup covers GPT-5 through GPT-5.6,
|
||||
in tool results. Model capability lookup table covers GPT-5/5.1/5.2/5.3/5.4,
|
||||
O-series, and search models (`gpt-5-search-api`) — all with `supports_vision`.
|
||||
For search models, injects `web_search_options` and removes the `web_search`
|
||||
function tool (the model always searches). Citations from `url_citation`
|
||||
annotations are formatted as footnotes. Pre-5.6 GPT-5 models request extended
|
||||
prompt-cache retention (`prompt_cache_retention: "24h"`); GPT-5.6 uses
|
||||
`prompt_cache_options.ttl: "30m"`. Cache reads and writes are extracted from
|
||||
`cached_tokens` and `cache_write_tokens`. Unknown models get permissive
|
||||
defaults with `supports_vision=False` and use SearxNG for web search. The
|
||||
`openai-compatible` lane never consults this table at all — on either API
|
||||
surface (the responses pin is served by a compat-mode
|
||||
`OpenAIResponsesProvider`, mirroring `AnthropicProvider(compat=True)`): a
|
||||
local server serves whatever the operator named it (vLLM
|
||||
`--served-model-name` is a free string), so a prefix collision with a cloud
|
||||
model id must not inherit that model's sampling/effort contract — every
|
||||
local model gets the plain defaults, commercial prompt-cache controls are not
|
||||
injected by model-name prefix, and anything beyond those defaults is declared
|
||||
on the model definition (capabilities JSON + `server_compat`), matching the
|
||||
`anthropic-compatible` lane.
|
||||
annotations are formatted as footnotes. Extended prompt cache retention
|
||||
(`prompt_cache_retention: "24h"`) is enabled for GPT-5.x models at no
|
||||
additional cost. Cached token counts are extracted from
|
||||
`usage.prompt_tokens_details.cached_tokens`. Unknown models (local servers) get
|
||||
permissive defaults with `supports_vision=False` and use SearxNG for web search.
|
||||
|
||||
**AnthropicProvider** (`_anthropic.py`): converts OpenAI-format messages to
|
||||
Anthropic content blocks, maps `system`/`developer` roles to the `system`
|
||||
@@ -662,7 +652,7 @@ display). Automatic prompt caching is enabled via top-level `cache_control:
|
||||
cacheable block and advances it as conversations grow (90% input cost
|
||||
reduction on cache hits, 1.25x write on first turn). Cache metrics
|
||||
(`cache_creation_input_tokens`, `cache_read_input_tokens`) are extracted from
|
||||
the stream's usage events. The `anthropic` SDK is a core
|
||||
both streaming and non-streaming responses. The `anthropic` SDK is a core
|
||||
dependency — the Anthropic provider is first-class alongside OpenAI.
|
||||
|
||||
**GoogleProvider** (`_google.py`): extends `OpenAIChatCompletionsProvider` for
|
||||
@@ -807,105 +797,15 @@ model = "deepseek-ai/DeepSeek-V4-Flash"
|
||||
supports_vision = true # multimodal checkpoints only
|
||||
supports_mid_conversation_system = true # template-dependent
|
||||
context_window = 131072
|
||||
thinking_mode = "manual" # session effort knob drives the template toggle
|
||||
thinking_param = "enable_thinking" # Qwen/Gemma key; "thinking" for Granite/DeepSeek
|
||||
```
|
||||
|
||||
Reasoning control does NOT use Anthropic's `thinking` request param —
|
||||
the levers live in the chat template, reached through
|
||||
`chat_template_kwargs` in the request body. Two channels, dynamic first:
|
||||
|
||||
* **Session effort knob (dynamic).** Set the model's thinking mode to
|
||||
"Effort-knob controlled" in the admin Models form (or
|
||||
`thinking_mode = "manual"` + `thinking_param` under
|
||||
`[models.*.capabilities]`) and the provider maps the session's
|
||||
reasoning-effort knob onto the template toggle per-request: effort
|
||||
`none` sends `{<thinking_param>: false}`, any other level sends
|
||||
`true` — the same contract as the real lane's manual mode. ("Always
|
||||
on" / `thinking_mode = "adaptive"` instead always sends `true`: the
|
||||
model self-regulates, so the knob never force-disables — mirroring
|
||||
the native adaptive branch.) The graded effort value always rides
|
||||
alongside the toggle: under `effort_param` when the operator names
|
||||
the template's key, else under the conventional fallback key
|
||||
(`reasoning_effort`) on the anthropic-compatible lane — the user's
|
||||
effort setting always reaches the wire, and a template that doesn't
|
||||
reference the kwarg ignores it. On the openai-compatible lane the
|
||||
undeclared-key case rides the flat top-level `reasoning_effort`
|
||||
param instead (the documented compat field), forwarded verbatim.
|
||||
Optional `reasoning_effort_values` / `default_reasoning_effort`
|
||||
validate the knob before it reaches the server; without declared
|
||||
values the knob is forwarded as-is. The knob is ordinal, and validation
|
||||
respects that: an off-list knob value rounds UP onto the declared
|
||||
list and a value above the ceiling rides the ceiling
|
||||
(`snap_reasoning_effort`) — asking for more effort than the model
|
||||
declares never falls back to a lower default tier. The knob's
|
||||
`none` position is forwarded verbatim when the model declares an
|
||||
explicit `none` level (gpt-5.1+, grok-4.3) — omitting it there would
|
||||
leave a reasoning-on server default (e.g. gpt-5.5's `medium`) in
|
||||
charge of a knob that promises off — and omitted otherwise; `none`
|
||||
is never a snap target for other positions.
|
||||
`default_reasoning_effort` only catches values the ordinal snap
|
||||
cannot rank (custom strings). Declare values that match the
|
||||
template's documented vocabulary: for DeepSeek-V4, which officially
|
||||
accepts `high`/`max` (Think High is the default thinking tier;
|
||||
`low`/`medium` alias to `high`, `xhigh` to `max`), a
|
||||
`("high", "max")` values list reproduces the official aliasing
|
||||
exactly — `low`/`medium` round up to `high`, `xhigh` to `max` —
|
||||
and freeform passthrough matches it too. To map an undocumented
|
||||
template, probe with per-request `chat_template_kwargs` and compare
|
||||
`input_tokens`. Setting `effort_param` also suppresses the
|
||||
flat top-level `reasoning_effort` request param on the
|
||||
openai-compatible lane — the template channel replaces it, never
|
||||
doubles it. With the default `thinking_mode = "none"` nothing is
|
||||
injected and the server's template default decides.
|
||||
|
||||
Upgrade note: before 1.7.0a7 the openai-compatible lane sent the
|
||||
toggle unconditionally `true` whenever thinking mode was enabled. A
|
||||
stored per-model `reasoning_effort = "none"` now disables thinking
|
||||
on such models — pick any real level (or clear the override) to keep
|
||||
it on. Also since 1.7.0a7 the effort level itself always reaches the
|
||||
wire on the local lanes (previously dropped unless
|
||||
`reasoning_effort_values` was declared): flat `reasoning_effort` on
|
||||
openai-compatible, the `effort_param`-or-fallback template key on
|
||||
anthropic-compatible when reasoning control is engaged.
|
||||
* **Operator pin (static).** Entries under `{"chat_template_kwargs":
|
||||
...}` in the admin Models extra-body field ride the SDK's
|
||||
`extra_body` unconditionally and win over the knob mapping on key
|
||||
collision — e.g. pin `{"enable_thinking": true}` to keep thinking on
|
||||
regardless of the session knob. (Server type and API surface remain
|
||||
openai-compatible-only knobs and stay hidden for this provider.)
|
||||
|
||||
The same knob mapping drives the `openai-compatible` lane's Chat
|
||||
Completions requests — `merge_reasoning_template_kwargs` is shared by
|
||||
both local-server lanes, so `thinking_mode`/`thinking_param`/
|
||||
`effort_param` mean the same thing whichever endpoint serves the model.
|
||||
Only the Responses API surface (native reasoning) ignores it.
|
||||
|
||||
The console surfaces this projection as an *effective effort ladder*:
|
||||
the admin model form's per-model effort select and the skill
|
||||
launch-config effort select annotate each position with what the
|
||||
request will carry, in plain words — a position whose delivered level
|
||||
matches its name stays plain ("Max"), a snapped position says so
|
||||
("Low — sends high"), the adaptive lanes' none position warns
|
||||
"thinking stays on", and budget detail lives in the tooltip. A
|
||||
position is never labeled after a sibling that shares its wire (that
|
||||
rendered "Max (= minimal)", implying a downgrade the wire doesn't
|
||||
contain). Computed server-side by `providers/effort_ladder.py` from
|
||||
the same mapping functions the providers use at request time and
|
||||
shipped on `/v1/api/models` rows (every row carries `effort_ladder`,
|
||||
empty when the capabilities column fails to parse) and
|
||||
`POST /v1/api/admin/models/effort-ladder`. The ladder describes what
|
||||
Turnstone sends — a server-side template may alias further (DeepSeek-V4
|
||||
folds `low`/`medium` into its default `high` tier).
|
||||
|
||||
The `anthropic-compatible` lane never sends Anthropic's native
|
||||
`thinking`/`output_config` params — they are not in vLLM's request
|
||||
schema. The real `anthropic` provider is unaffected: official Claude
|
||||
models keep native thinking, budget mapping, and `output_config`
|
||||
effort. A gateway fronting *real* Claude on a Messages-shaped URL
|
||||
(e.g. a LiteLLM `anthropic/` route to the Claude API) should use
|
||||
`provider = "anthropic"` with a custom `base_url`, which keeps the
|
||||
native thinking params.
|
||||
The reasoning toggle does NOT use Anthropic's `thinking` request param.
|
||||
Toggle it through the chat template instead: set `{"chat_template_kwargs":
|
||||
{"thinking": false}}` as extra body params in the admin Models
|
||||
server-compat section (for this provider the section shows only the
|
||||
extra-body field — server type, API surface, and thinking mode are
|
||||
openai-compatible-only knobs); the provider forwards it via the SDK's
|
||||
`extra_body`.
|
||||
|
||||
Verified quirks of vLLM's Anthropic endpoint:
|
||||
|
||||
@@ -1117,10 +1017,9 @@ reconstructs the OpenAI message format from database rows:
|
||||
in the same workstream
|
||||
|
||||
**Config persistence:** LLM-affecting parameters (`temperature`,
|
||||
`reasoning_effort`, `max_tokens`, `instructions`, and the persona
|
||||
snapshot — see `docs/personas.md`) are persisted to the
|
||||
`workstream_config` table on creation and whenever changed via slash
|
||||
commands. `resume()` restores these values so resumed workstreams
|
||||
`reasoning_effort`, `max_tokens`, `instructions`, `creative_mode`) are
|
||||
persisted to the `workstream_config` table on creation and whenever changed
|
||||
via slash commands. `resume()` restores these values so resumed workstreams
|
||||
behave identically to the original.
|
||||
|
||||
**`/clear` vs `/new`:** `/clear` wipes in-memory context but preserves
|
||||
@@ -1149,30 +1048,20 @@ Named (aliased) workstreams are never age-pruned. Configure with
|
||||
|
||||
### API Retry
|
||||
|
||||
Every model call streams (#831); retry lives at two stacked layers:
|
||||
`ChatSession._create_stream_with_retry()` (streaming path) and the agent
|
||||
`_api_call()` (non-streaming) both use the same retry pattern:
|
||||
|
||||
- **Caller ladders** — `ChatSession._create_stream_with_retry()` (chat
|
||||
loop) and the agent `_api_call()` (drained via `model_turn`) use the
|
||||
same pattern: 4 total attempts (1 initial + 3 retries,
|
||||
`_MAX_RETRIES = 3`), exponential backoff base 1 second
|
||||
(`delay = 1s * 2^attempt`), `ui.on_info()` on retry, exception
|
||||
propagates on final failure. `_compact_messages()` wraps its drained
|
||||
call in the same loop.
|
||||
- **`model_turn`'s drain ladder** — inside every single-shot call,
|
||||
mid-stream deaths (errors raised while draining, e.g.
|
||||
`IncompleteStreamError`) are re-issued up to 2 more times with a
|
||||
0.5s-base exponential backoff (±50% jitter); request-time failures
|
||||
keep the SDK's own retry policy. The two ladders stack
|
||||
multiplicatively on transient-shaped failures.
|
||||
- **Retryable errors** are matched by class name against each
|
||||
provider's `retryable_error_names` (avoids importing
|
||||
backend-specific exception hierarchies): `RateLimitError`,
|
||||
`APITimeoutError`, `APIConnectionError`, `InternalServerError`,
|
||||
`ServiceUnavailableError`, `APIError`, plus the drained-transport
|
||||
errors `IncompleteStreamError` (stream ended with no terminal
|
||||
signal — for servers that never send one, declare
|
||||
`finish_reason_optional` in the model's capabilities JSON) and
|
||||
`ResponsesStreamFailedError` (transient in-band Responses failure).
|
||||
- **Retries**: 4 total attempts (1 initial + 3 retries, `_MAX_RETRIES = 3`)
|
||||
- **Backoff**: exponential, base 1 second (`delay = 1s * 2^attempt`)
|
||||
- **Retryable errors**: `RateLimitError`, `APITimeoutError`,
|
||||
`APIConnectionError`, `InternalServerError`, `ServiceUnavailableError`,
|
||||
`APIError` (matched by class name to avoid importing backend-specific
|
||||
exception hierarchies)
|
||||
- On retry: `ui.on_info()` notification
|
||||
- On final failure: exception propagates
|
||||
|
||||
`_compact_messages()` also wraps its non-streaming API call in the same
|
||||
retry loop.
|
||||
|
||||
### Finish Reason Handling
|
||||
|
||||
@@ -1185,7 +1074,7 @@ Every model call streams (#831); retry lives at two stacked layers:
|
||||
blocked.
|
||||
|
||||
Agent sub-sessions (`_run_agent()`) check `finish_reason` on each
|
||||
drained turn and stop the agent early on `"length"` or
|
||||
non-streaming response and stop the agent early on `"length"` or
|
||||
`"content_filter"`.
|
||||
|
||||
`_compact_messages()` checks `finish_reason` on the compaction response and
|
||||
|
||||
+3
-4
@@ -379,7 +379,6 @@ Breadcrumb: `Cluster > Running` or `Cluster > db-west-04`. Server-side paginated
|
||||
Triggered by the "+ new" header button. A modal dialog with:
|
||||
|
||||
- **Node selector** — dropdown with three targeting modes: "Auto (best available)" picks the node with the most headroom, "General pool (any node)" picks a node with available capacity using round-robin, or a specific node from the list (showing capacity).
|
||||
- **Persona** — optional dropdown listing the enabled personas for the workstream kind. Sets the system-message composition and capability envelope at creation, snapshotted server-side; empty uses the kind's default. Picking one requires no `persona.*` permission.
|
||||
- **Profile** — optional dropdown listing enabled skills. Applies the skill's model, auto-approve policy, token budget, and other behavioral settings at creation time.
|
||||
- **Name** — optional text input. Auto-generated if left empty.
|
||||
- **Model** — optional text input for a model alias from the target node's registry.
|
||||
@@ -397,9 +396,9 @@ The browser maintains a local `clusterState` object that mirrors the cluster sna
|
||||
|
||||
Accessed via the "admin" button in the header (visible when authenticated
|
||||
with `approve` scope). Provides user, API token, channel link, MCP server,
|
||||
and skill management with tabs that include Users, API Tokens, Channels,
|
||||
Schedules, Watches, Personas, Roles, Policies, Prompts, Judge, Skills,
|
||||
MCP Servers, Usage, Audit, Memories, Models, Nodes, Settings, and TLS. See also
|
||||
and skill management with 18 tabs (Users, API Tokens, Channels, Schedules,
|
||||
Watches, Roles, Policies, Prompts, Judge, Skills, MCP Servers, Usage,
|
||||
Audit, Memories, Models, Nodes, Settings, TLS). See also
|
||||
[Governance](governance.md) for the Roles, Policies, Skills, Usage, and
|
||||
Audit tabs, and [Settings](settings.md) for the database-backed
|
||||
configuration editor.
|
||||
|
||||
@@ -366,7 +366,7 @@ deleted.
|
||||
## Further reading
|
||||
|
||||
- [coordinator-skills.md](coordinator-skills.md) — writing a skill
|
||||
that runs on a coordinator session (orchestrator framing,
|
||||
that runs on a coordinator session (orchestrator persona,
|
||||
workflow patterns, `SkillKind` classifier).
|
||||
- [bulk-endpoints.md](bulk-endpoints.md) — the two bulk-shape
|
||||
idioms (`{results, denied, truncated}` vs
|
||||
|
||||
+11
-11
@@ -1,13 +1,13 @@
|
||||
# Writing a coordinator-specific skill
|
||||
|
||||
A skill is prompt-level framing that steers a Turnstone session
|
||||
Skills are prompt-level personas that steer a Turnstone session
|
||||
toward a narrow task. Most skills target **interactive** sessions —
|
||||
the single-workstream "do this thing" surface where the model wields
|
||||
`bash`, `edit_file`, `web_fetch`, and the rest of the maker toolset.
|
||||
|
||||
A **coordinator skill** is different. It runs on a session whose job
|
||||
is to orchestrate other sessions. The toolset is smaller and
|
||||
narrower, the role is an orchestrator instead of a maker, and the
|
||||
narrower, the persona is an orchestrator instead of a maker, and the
|
||||
success metric is "did the plan resolve" instead of "did the code
|
||||
compile". This doc covers the differences a skill author has to
|
||||
care about.
|
||||
@@ -22,8 +22,8 @@ migration 044 added the column). Three values:
|
||||
|
||||
| `SkillKind` enum | Stored as | Meaning |
|
||||
|-------------------------|-----------------|----------------------------------------------------------------------------|
|
||||
| `SkillKind.INTERACTIVE` | `"interactive"` | Authored for the interactive maker role (single-workstream "do this"). |
|
||||
| `SkillKind.COORDINATOR` | `"coordinator"` | Authored for the orchestrator role (delegate, monitor, synthesise). |
|
||||
| `SkillKind.INTERACTIVE` | `"interactive"` | Authored for the interactive maker persona (single-workstream "do this"). |
|
||||
| `SkillKind.COORDINATOR` | `"coordinator"` | Authored for the orchestrator persona (delegate, monitor, synthesise). |
|
||||
| `SkillKind.ANY` | `"any"` | Either surface (or audience-neutral). Default on create. |
|
||||
|
||||
The `kind` field is a `StrEnum` — drop-in `str` compatible — so DB
|
||||
@@ -96,20 +96,20 @@ for the output. The coordinator stays the orchestrator.
|
||||
|
||||
---
|
||||
|
||||
## Framing differences
|
||||
## Persona differences
|
||||
|
||||
Interactive skills compose on top of `base_interactive.md` — a
|
||||
"maker" framing: get the work done, use the tools, edit the code,
|
||||
"maker" persona: get the work done, use the tools, edit the code,
|
||||
close the loop.
|
||||
|
||||
Coordinator skills compose on top of
|
||||
[`personas/orchestrator.md`](../turnstone/prompts/personas/orchestrator.md) —
|
||||
an "orchestrator" framing: decompose, delegate, monitor, synthesise.
|
||||
[`base_coordinator.md`](../turnstone/prompts/base_coordinator.md) —
|
||||
an "orchestrator" persona: decompose, delegate, monitor, synthesise.
|
||||
The base text is short but sets the tone every coordinator skill
|
||||
inherits:
|
||||
|
||||
> You are a coordinator. Your role is to orchestrate work across
|
||||
> the cluster... You do
|
||||
> You are a coordinator on a small, focused infrastructure team.
|
||||
> Your role is to orchestrate work across the cluster... You do
|
||||
> not edit files, run shell commands, browse the web, or manipulate
|
||||
> the codebase directly. Children do that.
|
||||
|
||||
@@ -339,7 +339,7 @@ For a new coordinator skill:
|
||||
A full end-to-end test isn't required for every skill; a
|
||||
prepare-step unit test that asserts "given this initial message, the
|
||||
first tool call is X with Y args" is usually sufficient to catch
|
||||
framing drift without a real LLM in the loop.
|
||||
persona drift without a real LLM in the loop.
|
||||
|
||||
---
|
||||
|
||||
|
||||
@@ -66,7 +66,8 @@ class "NullUI" as NullUI {
|
||||
interface "LLMProvider" as LLMProvider <<Protocol>> {
|
||||
+ provider_name: str {property}
|
||||
+ get_capabilities(model) → ModelCapabilities
|
||||
+ create_streaming(client, model, messages, ..., cancel_ref, replay_reasoning_to_model) → Iterator[StreamChunk]
|
||||
+ create_streaming(client, model, messages, ..., replay_reasoning_to_model) → Iterator[StreamChunk]
|
||||
+ create_completion(client, model, messages, ..., replay_reasoning_to_model) → CompletionResult
|
||||
+ convert_tools(tools) → list[dict]
|
||||
+ extract_reasoning_text(provider_blocks) → str
|
||||
+ retryable_error_names: frozenset[str] {property}
|
||||
@@ -176,7 +177,7 @@ class "HeadlessSession" as HeadlessSession {
|
||||
+ send_headless(input, max_turns, ...)
|
||||
- _override_system_prompt(content)
|
||||
--
|
||||
eval.py: drained single-shot turns,
|
||||
eval.py: non-streaming,
|
||||
records all tool calls
|
||||
}
|
||||
|
||||
|
||||
@@ -84,8 +84,8 @@ end note
|
||||
|
||||
loop up to 3 turns (timeout budget)
|
||||
|
||||
Judge -> LLM : model_turn(lane, judge_turns,\ntools=[read_file, list_directory])\nvia drained create_streaming
|
||||
LLM --> Judge : ModelTurnResult
|
||||
Judge -> LLM : create_completion(\nmodel, judge_messages,\ntools=[read_file, list_directory])
|
||||
LLM --> Judge : CompletionResult
|
||||
|
||||
alt tool_calls present (turn < 3)
|
||||
Judge -> Judge : _exec_read_only_tool()
|
||||
|
||||
+1
-31
@@ -252,7 +252,6 @@ interface, or anyone who can reach it can search through your instance.
|
||||
| Variable | Default | Description |
|
||||
|----------|---------|-------------|
|
||||
| `WORKSPACE_MOUNT` | empty volume | Host directory bind-mounted at `/workspace` for the model to read/write |
|
||||
| `TURNSTONE_WORKSPACE` | `/workspace` (image env) | Directory named as the user's workspace in the model's tool descriptions; informational only — see [Working directory](#working-directory) |
|
||||
| `SKIP_PERMISSIONS` | — | Set to any value to auto-approve all tool calls (dev only) |
|
||||
| `MCP_CONFIG` | — | Path to an MCP server config file |
|
||||
| `TURNSTONE_IMAGE_TAG` | `latest` | ghcr.io image tag — production stack |
|
||||
@@ -261,7 +260,7 @@ interface, or anyone who can reach it can search through your instance.
|
||||
|
||||
Both stacks install all entry points into a single image (`turnstone`,
|
||||
`turnstone-server`, `turnstone-console`, `turnstone-channel`, `turnstone-admin`,
|
||||
`turnstone-eval`, `turnstone-optimizer`, `turnstone-doctor`):
|
||||
`turnstone-eval`, `turnstone-doctor`):
|
||||
|
||||
```bash
|
||||
docker compose build # build the dev image
|
||||
@@ -277,35 +276,6 @@ docker compose build --no-cache # rebuild from scratch
|
||||
| `workspace` | `/workspace` (unless `WORKSPACE_MOUNT` is set) |
|
||||
| `caddy-data` / `caddy-config` | Caddy's local CA and config (dev stack) |
|
||||
|
||||
## Working directory
|
||||
|
||||
Node processes run with `/data` as their working directory (the image's
|
||||
`WORKDIR`), and that is where the model's shell commands execute and
|
||||
relative file paths resolve — **not** `/workspace`. The shell and file
|
||||
tool descriptions state both paths (the working directory, and the
|
||||
workspace named by `TURNSTONE_WORKSPACE`), so the model knows to look in
|
||||
`/workspace` for your files without being told each session.
|
||||
|
||||
To make tools start inside the mount instead, override the working
|
||||
directory on the node services:
|
||||
|
||||
```yaml
|
||||
services:
|
||||
turnstone-node:
|
||||
working_dir: /workspace
|
||||
```
|
||||
|
||||
Two caveats before overriding:
|
||||
|
||||
- **SQLite fallback**: when a node runs without PostgreSQL, its fallback
|
||||
database `.turnstone.db` is created in the process working directory.
|
||||
Changing `working_dir` on an existing SQLite-fallback deployment makes
|
||||
the node create a fresh database inside the mount and your prior state
|
||||
appears lost (it is still in the `turnstone-data` volume under `/data`).
|
||||
The stock compose stacks use PostgreSQL and are unaffected.
|
||||
- Migrations (`entrypoint.sh`) run in the same working directory, so the
|
||||
same SQLite caveat applies to them.
|
||||
|
||||
## Cleanup
|
||||
|
||||
```bash
|
||||
|
||||
+24
-55
@@ -1,19 +1,11 @@
|
||||
# Evaluation and Prompt Optimization (turnstone-eval, turnstone-optimizer)
|
||||
# Evaluation and Prompt Optimization (turnstone-eval)
|
||||
|
||||
Evaluation for turnstone is split into two commands:
|
||||
`turnstone-eval` is the evaluation and prompt optimization system for turnstone. It
|
||||
runs test cases against the LLM, scores tool call sequences against expected
|
||||
actions, and optionally uses a multi-agent pipeline to optimize the developer
|
||||
prompt and tool descriptions.
|
||||
|
||||
- **`turnstone-eval`** — the measurement substrate. Runs test cases against the LLM
|
||||
and scores tool call sequences against expected actions. A single measurement pass,
|
||||
no self-modification.
|
||||
- **`turnstone-optimizer`** — the prompt/tool optimizer. Loops over the measurement
|
||||
substrate, using a multi-agent pipeline (analyst, optimizer, observer, diversifier,
|
||||
tool optimizer) to edit the developer prompt and tool descriptions so more tests pass.
|
||||
|
||||
The dependency is strictly one-way: the optimizer consumes the eval substrate; the
|
||||
substrate never depends on the optimizer.
|
||||
|
||||
Source: `turnstone/eval/core.py` (measurement substrate), `turnstone/eval/cli.py`
|
||||
(the `turnstone-eval` CLI), `turnstone/optimizer.py` (the `turnstone-optimizer` CLI).
|
||||
Source: `turnstone/eval.py`
|
||||
|
||||
---
|
||||
|
||||
@@ -35,8 +27,8 @@ This approach (inspired by [Learning to Self-Evolve](https://arxiv.org/abs/2603.
|
||||
prevents irrecoverable collapse from bad edits — UCB naturally backtracks to
|
||||
high-scoring ancestors instead of following a linear chain.
|
||||
|
||||
The `turnstone-eval` command (or `turnstone-optimizer --no-optimize`) executes only
|
||||
steps 2-4: a single measurement pass over the root prompt, no optimization.
|
||||
When optimization is disabled (`--no-optimize`), only steps 2-4 execute
|
||||
(a single iteration evaluating the root node).
|
||||
|
||||
---
|
||||
|
||||
@@ -460,46 +452,30 @@ structure is:
|
||||
|
||||
## CLI Usage
|
||||
|
||||
Two console scripts (installed as entry points), or the equivalent `python -m`
|
||||
invocations:
|
||||
|
||||
- `turnstone-eval` / `python -m turnstone.eval.cli` — measure only.
|
||||
- `turnstone-optimizer` / `python -m turnstone.optimizer` — optimize.
|
||||
|
||||
### Measure (`turnstone-eval`)
|
||||
The entry point is `turnstone-eval` (installed as a console script) or
|
||||
`python -m turnstone.eval`.
|
||||
|
||||
```
|
||||
turnstone-eval tests.json # one measurement pass, print scores
|
||||
turnstone-eval tests.json --prompt custom.txt # measure a custom prompt
|
||||
turnstone-eval tests.json --n-runs 5 # more runs per case
|
||||
turnstone-eval tests.json --parallel 4 # run cases across 4 workers
|
||||
turnstone-eval tests.json -v # verbose per-turn logging
|
||||
turnstone-eval tests.json # evaluate + optimize
|
||||
turnstone-eval tests.json --no-optimize # evaluate only (single iteration)
|
||||
turnstone-eval tests.json --n-runs 5 --max-iter 10 # more thorough evaluation
|
||||
turnstone-eval tests.json --prompt custom.txt # start from a custom prompt
|
||||
turnstone-eval tests.json --optimize-tools # optimize tool descriptions only
|
||||
turnstone-eval tests.json --diversify 10 # test with prompt variants
|
||||
turnstone-eval tests.json -v # verbose per-turn logging
|
||||
```
|
||||
|
||||
### Optimize (`turnstone-optimizer`)
|
||||
### Multi-model setup (local test model, cloud optimizer)
|
||||
|
||||
```
|
||||
turnstone-optimizer tests.json # evaluate + optimize
|
||||
turnstone-optimizer tests.json --no-optimize # single pass, no optimization
|
||||
turnstone-optimizer tests.json --n-runs 5 --max-iter 10 # more thorough optimization
|
||||
turnstone-optimizer tests.json --prompt custom.txt # start from a custom prompt
|
||||
turnstone-optimizer tests.json --optimize-tools # optimize tool descriptions only
|
||||
turnstone-optimizer tests.json --diversify 10 # test with prompt variants
|
||||
```
|
||||
|
||||
#### Multi-model setup (local test model, cloud optimizer)
|
||||
|
||||
```
|
||||
turnstone-optimizer tests.json \
|
||||
turnstone-eval tests.json \
|
||||
--base-url http://localhost:8000/v1 \
|
||||
--optimizer-base-url https://api.anthropic.com \
|
||||
--optimizer-model claude-sonnet-4-6 \
|
||||
--analyst-model claude-opus-4-6
|
||||
```
|
||||
|
||||
### Measurement Options
|
||||
|
||||
Accepted by **both** commands.
|
||||
### All Options
|
||||
|
||||
| Flag | Default | Description |
|
||||
|-------------------------|----------------------------|-------------|
|
||||
@@ -508,26 +484,19 @@ Accepted by **both** commands.
|
||||
| `--model` | auto-detect | Model name. Auto-detected from the API if not specified. |
|
||||
| `--prompt` | turnstone built-in prompt | Path to initial prompt text file. |
|
||||
| `--n-runs` | from tests.json or 3 | Number of runs per test case. |
|
||||
| `--max-iter` | 5 | Maximum optimization iterations. |
|
||||
| `--no-optimize` | false | Run evaluation only (sets max-iter to 1). |
|
||||
| `--temperature` | 0.7 | Sampling temperature. |
|
||||
| `--max-tokens` | 32768 | Max completion tokens. |
|
||||
| `--reasoning-effort` | `medium` | Reasoning effort: `low`, `medium`, or `high`. |
|
||||
| `--context-window` | 131072 | Context window size. |
|
||||
| `--output` | `eval_results.json` | Output results file path. |
|
||||
| `-v`, `--verbose` | false | Show detailed per-turn logging. |
|
||||
| `--explore-constant` | 1.414 (sqrt(2)) | UCB exploration constant C. |
|
||||
| `--test-timeout` | 300 | Per-test timeout in seconds. |
|
||||
| `--suite-timeout` | 0 (unlimited) | Total suite timeout in seconds. |
|
||||
| `--no-fast-fail` | false | Disable early termination on all-zero initial runs. |
|
||||
| `--parallel` | 1 (serial) | Parallel workers (0=auto, N=use N workers). |
|
||||
|
||||
### Optimizer Options
|
||||
|
||||
Accepted by **`turnstone-optimizer`** only.
|
||||
|
||||
| Flag | Default | Description |
|
||||
|-------------------------|----------------------------|-------------|
|
||||
| `--max-iter` | 5 | Maximum optimization iterations. |
|
||||
| `--no-optimize` | false | Run a single measurement pass (sets max-iter to 1). |
|
||||
| `--explore-constant` | 1.414 (sqrt(2)) | UCB exploration constant C. |
|
||||
| `--suite-timeout` | 0 (unlimited) | Total suite timeout in seconds. |
|
||||
| `--optimizer-model` | same as `--model` | Model for prompt optimization. |
|
||||
| `--optimizer-base-url` | same as `--base-url` | Base URL for optimizer model. |
|
||||
| `--observer-model` | same as optimizer | Model for meta-optimization (observer). |
|
||||
|
||||
+6
-14
@@ -13,7 +13,7 @@ The permission model has two layers:
|
||||
|
||||
1. **Scopes** (legacy) — `read`, `write`, `approve`. Checked by `AuthMiddleware`
|
||||
on every request based on URL path classification.
|
||||
2. **Permissions** (granular) — named permission strings checked per-endpoint by
|
||||
2. **Permissions** (granular) — 15 permission strings checked per-endpoint by
|
||||
`require_permission()`.
|
||||
|
||||
**Built-in roles** (seeded by migration 008):
|
||||
@@ -24,11 +24,7 @@ The permission model has two layers:
|
||||
| operator | read, write, workstreams.create, workstreams.close |
|
||||
| viewer | read |
|
||||
|
||||
Custom roles can be created with any subset of the valid permissions.
|
||||
The `persona.create` / `persona.read` / `persona.write` family gates
|
||||
persona administration; migration `063` seeds all three onto
|
||||
`builtin-admin`, and any role can be granted them through the standard
|
||||
role and permission-override editors.
|
||||
Custom roles can be created with any subset of the 15 valid permissions.
|
||||
|
||||
**Auth flow:**
|
||||
1. User logs in (password or API token) → `_load_user_permissions()` aggregates
|
||||
@@ -131,12 +127,9 @@ Per-LLM-request token and tool call metrics:
|
||||
LLM response with prompt/completion tokens, cache tokens, tool call count,
|
||||
model, ws_id
|
||||
- **Prompt caching**: Anthropic automatic caching (`cache_control: ephemeral`)
|
||||
and OpenAI caching are enabled by default. Pre-5.6 GPT-5 models request
|
||||
`prompt_cache_retention: 24h`; GPT-5.6 uses
|
||||
`prompt_cache_options: {"ttl": "30m"}`. GPT-5.6 cache writes use the
|
||||
provider's 1.25× input-token rate. `cache_creation_tokens` and
|
||||
`cache_read_tokens` are tracked per request in `usage_events` and surfaced
|
||||
in the Usage admin tab
|
||||
and OpenAI extended retention (`prompt_cache_retention: 24h` for GPT-5.x)
|
||||
are enabled by default. `cache_creation_tokens` and `cache_read_tokens` are
|
||||
tracked per request in `usage_events` and surfaced in the Usage admin tab
|
||||
- **Querying**: `GET /v1/api/admin/usage` with `group_by` (day/hour/model/user)
|
||||
and time range filtering — includes cache token aggregates
|
||||
- **Prometheus**: `turnstone_tokens_total{type="cache_creation|cache_read"}`
|
||||
@@ -184,7 +177,6 @@ All under `/v1/api/admin/` (requires `approve` scope + granular permission).
|
||||
| Orgs | 3 (list, get, update) | `admin.orgs` |
|
||||
| Tool Policies | 4 (CRUD) | `admin.policies` |
|
||||
| Skills | 4 (CRUD) | `admin.skills` |
|
||||
| Personas | 4 (list, create, get, edit/archive) | `persona.read` / `persona.create` / `persona.write` |
|
||||
| Schedules | 6 (CRUD + runs) | `admin.schedules` |
|
||||
| Watches | 3 (list, create, cancel) | `admin.watches` |
|
||||
| Usage | 1 (aggregated query) | `admin.usage` |
|
||||
@@ -230,7 +222,7 @@ Both Python and TypeScript console SDKs expose governance methods:
|
||||
- **Privilege escalation prevented**: `admin_assign_role` blocks self-assignment
|
||||
and requires caller to hold a superset of the target role's permissions
|
||||
- **Permission validation**: Role create/update validates permissions against
|
||||
the permission allowlist (`_VALID_PERMISSIONS`)
|
||||
a 15-item allowlist (`_VALID_PERMISSIONS`)
|
||||
- **Self-deletion blocked**: `admin_delete_user` rejects attempts to delete
|
||||
your own account (matching the self-assignment guard on role endpoints)
|
||||
- **Field allowlists**: Storage `update_*` methods filter fields against
|
||||
|
||||
+2
-8
@@ -249,14 +249,8 @@ are withheld from the live surfaces (a reused call_id must never ride a stale
|
||||
`approve` into Smart Approvals) but still persist with
|
||||
`user_decision = "superseded"` so the audit trail records the judge's answer.
|
||||
|
||||
Sub-agent (task agent) tool calls are judge-gated too. Each runs the same
|
||||
intent pipeline as its own `agent_gate` generation, grounded in that sub-agent's
|
||||
own trajectory -- its task prompt is the delegation contract the operator
|
||||
approved, so "does this call serve the task" is the right local question.
|
||||
Agent-gate generations never occupy the main loop's supersede slot (parallel
|
||||
siblings would otherwise make each other's verdicts look stale); per-cycle
|
||||
generation checks enforce staleness instead, and `judge.cancel_on_approval`
|
||||
fires per gate exactly like the main loop.
|
||||
Sub-agents (plan agent, task agent) are exempt from intent validation -- they
|
||||
always get full tool visibility without judge evaluation.
|
||||
|
||||
---
|
||||
|
||||
|
||||
+4
-64
@@ -17,9 +17,8 @@ The MCP server admin form exposes three authorization modes ("Multitenant Author
|
||||
| `none` | No headers attached. Open MCP server (or one gated by network policy only). | Internal MCP servers on a trusted network. |
|
||||
| `static` | One static bearer token, configured per server, sent on every request from every user. | Service-to-service MCP servers where per-user attribution doesn't matter, or single-tenant deployments. |
|
||||
| `oauth_user` *(recommended for user-data servers)* | Each user authorizes separately via OAuth 2.1 + PKCE; Turnstone stores per-user tokens encrypted at rest. | MCP servers that expose user-specific data or that want per-user audit attribution. |
|
||||
| `oauth_obo` *(sign-in passthrough)* | Each user's Turnstone **org sign-in** (OIDC) mints a per-server access token on demand — no separate per-server consent. One captured credential per user covers every `oauth_obo` server. | Enterprise deployments where the identity provider governs access (Entra, Keycloak) and you want zero per-user connect clicks. See the dedicated section below. |
|
||||
|
||||
Switching `auth_type` away from `oauth_user` / `oauth_obo` **deletes** that server's per-user rows (consents / minted cache) — see the transition table below. Switching back later starts clean: users re-consent (or re-mint) on next use. The admin **bulk-revoke** / **flush cache** affordance clears rows without an auth-type change.
|
||||
Switching `auth_type` away from `oauth_user` orphans existing per-user tokens. Use the admin **bulk-revoke** affordance on the server row (Phase 9) to clear them, or let them expire naturally — they're inert without the matching `auth_type` value.
|
||||
|
||||
---
|
||||
|
||||
@@ -66,58 +65,6 @@ Keep this in `config.toml` rather than environment variables. An in-process LLM
|
||||
|
||||
---
|
||||
|
||||
## `auth_type=oauth_obo` — single-credential sign-in passthrough
|
||||
|
||||
Where `oauth_user` makes each user complete a **separate** browser consent per MCP server, `oauth_obo` reuses the user's Turnstone **org sign-in** (OIDC). Turnstone captures one refresh credential per user at login and, on each tool call, mints a short-lived access token scoped to that server's audience. There is no per-server connect step, and one credential covers every `oauth_obo` server. This is the right shape when your identity provider already governs who may reach each backend (an Entra tenant with Entra-protected MCP servers; a Keycloak realm with token exchange).
|
||||
|
||||
Access is governed **downstream** by the IdP: a user can only mint a token for a server their delegated permissions allow. Removing that grant at the IdP cuts the user off regardless of their Turnstone state.
|
||||
|
||||
### Deployment configuration (`[oidc]` in `config.toml`)
|
||||
|
||||
`oauth_obo` requires OIDC SSO to be configured (it is the credential source), plus:
|
||||
|
||||
```toml
|
||||
[oidc]
|
||||
# ... your existing issuer / client_id / client_secret ...
|
||||
capture_user_credential = true # persist the IdP refresh token at login
|
||||
obo_grant_profile = "entra" # "entra" | "rfc8693" — how tokens are minted
|
||||
```
|
||||
|
||||
- **`capture_user_credential`** (default `false`): when enabled, Turnstone appends `offline_access` to the login scopes and stores the returned refresh token, encrypted with the same `[security] mcp_token_encryption_key` as `oauth_user` tokens. **The encryption key is required** — Turnstone refuses to start with an `oauth_obo` row (or capture enabled) and no key.
|
||||
- **`obo_grant_profile`** picks the mint mechanism (the IdP determines which one is valid; this is deployment-wide, not per-server):
|
||||
- **`entra`** — redeems the user's refresh token directly for a token scoped to `<audience>/.default`. `oauth_scopes` on the server row is **not used** (the admin form rejects it under this profile).
|
||||
- **`rfc8693`** — a refresh grant for a subject token, then an RFC 8693 token exchange for the server audience. Per-server `oauth_scopes` **are** sent on the exchange (some IdPs require the audience scope explicitly).
|
||||
|
||||
### Adding an `oauth_obo` server
|
||||
|
||||
In the admin MCP form, choose **Sign-in passthrough** and set **Audience** (required — the downstream resource the token is minted for, e.g. `api://<app-id>` on Entra or the client id on Keycloak). The client-id / secret / registration fields do not apply and are hidden.
|
||||
|
||||
`oauth_obo` servers are accepted only when **OIDC sign-in is configured and enabled** and `[oidc] obo_grant_profile` is a valid profile — the write is rejected otherwise, since a row that can never mint would surface to users as a permanent "please retry" that never heals.
|
||||
|
||||
### Identity-provider setup
|
||||
|
||||
**Entra (`obo_grant_profile = "entra"`):**
|
||||
1. Turnstone's app registration must hold **delegated permissions** to each MCP server's exposed API, with **admin consent granted** (or the MCP app listed in Turnstone's `preAuthorizedApplications`).
|
||||
2. Set the server row's Audience to the MCP app's Application ID URI (`api://<guid>`).
|
||||
3. **Gotcha (verified):** admin-consent issued *immediately* after creating the app/service principal can silently skip a not-yet-propagated resource — the only symptom is `AADSTS65001` at mint time. Verify the delegated grant landed (`az ad app permission list-grants` / the portal's *API permissions* blade shows *Granted*), or grant it explicitly per resource. A missing grant surfaces in Turnstone as a re-login prompt on the affected server (same rail as a revoked credential), and the `mcp_server.oauth.obo_mint_rejected` log line carries the raw `AADSTS…` text.
|
||||
|
||||
**Keycloak / RFC 8693 (`obo_grant_profile = "rfc8693"`):**
|
||||
1. Enable **standard token exchange** on Turnstone's client.
|
||||
2. Grant the audience: add an audience client scope for each MCP client and attach it to Turnstone's client (optional scopes must be requested — set the server row's Scopes to that scope, or the exchange returns *"Requested audience not available"*).
|
||||
3. Set the server row's Audience to the downstream client id.
|
||||
|
||||
### Revocation & custody
|
||||
|
||||
The captured credential is a single per-user secret that can mint for every `oauth_obo` server, so treat it like any long-lived credential:
|
||||
|
||||
- **Cut off one user:** unlink their OIDC identity in the admin console (**Users → OIDC identities → delete**). This revokes the captured credential **and** purges their minted cache rows, so future mints fail and cached tokens are dropped. (Warmed in-memory sessions on server nodes self-expire at the access-token TTL; there is no cross-node per-user session-kill.) Removing the user's access at the IdP is the authoritative cut-off.
|
||||
- **Flush a server's minted tokens** (e.g. after narrowing its audience): the server row's **flush cache** action drops all users' cached tokens for that server. This is **not** a revocation — users re-mint on next use from their still-valid sign-in. It is surfaced honestly (audit `mcp_server.oauth.obo_cache_flushed`, response `effect: cache_flush_remints`) so it is never mistaken for cutting access.
|
||||
- Per-server revocation in the `oauth_user` sense does not exist for `oauth_obo` — the credential is issuer-scoped and IdP-governed. Revoke at the IdP.
|
||||
|
||||
> **Interim for Entra without OBO:** if you don't want host-side minting, admin consent + `preAuthorizedApplications` on each MCP app registration removes the second consent prompt for the plain `oauth_user` flow too (a tenant-config change, no Turnstone code). Tracked in issue #682. It does not remove the per-server connect clicks or per-(user, server) token custody — that is what `oauth_obo` is for.
|
||||
|
||||
---
|
||||
|
||||
## Lifecycle
|
||||
|
||||
1. **First tool call** for a user against an `oauth_user` MCP server: pool dispatch finds no stored token, returns `mcp_consent_required` to the agent. Dashboard renders an inline "Connect" action card.
|
||||
@@ -128,7 +75,7 @@ The captured credential is a single per-user secret that can mint for every `oau
|
||||
|
||||
4. **Step-up scope**: when a tool call hits `403` with `WWW-Authenticate: error="insufficient_scope"`, Turnstone emits `mcp_insufficient_scope` with the parsed scope set; the dashboard offers a "Connect with additional scopes" affordance that opens `/v1/api/mcp/oauth/start?server=<name>&scopes=<extra>` so the union of original + new scopes flows into the AS authorize request.
|
||||
|
||||
5. **User revoke** (settings modal): `DELETE /v1/api/mcp/oauth/connections/{server_name}` runs the authoritative local delete + best-effort RFC 7009 upstream revoke (fire-and-forget, capped at 256 concurrent in-flight tasks). `oauth_obo` servers are excluded: their rows are mint cache, not consents — deleting one only forces a re-mint — so the connections list hides them and the endpoint refuses them with `409` (revocation for sign-in passthrough happens at the identity layer: unlink the identity or revoke at the IdP).
|
||||
5. **User revoke** (settings modal): `DELETE /v1/api/mcp/oauth/connections/{server_name}` runs the authoritative local delete + best-effort RFC 7009 upstream revoke (fire-and-forget, capped at 256 concurrent in-flight tasks).
|
||||
|
||||
6. **Admin bulk-revoke** (Phase 9): `POST /v1/api/admin/mcp-servers/{name}/bulk-revoke` drops every user's token for the server. Upstream RFC 7009 revoke is intentionally **not** attempted in bulk (avoids N upstream HTTP calls per admin click); tokens at the AS expire naturally. Use the per-user revoke endpoint if you need guaranteed upstream invalidation.
|
||||
|
||||
@@ -150,13 +97,10 @@ Additional indicators (circuit-breaker state, encryption-key mismatch) are expos
|
||||
| From | To | What happens |
|
||||
|---|---|---|
|
||||
| `none` / `static` → `oauth_user` | — | New code path activates for this server. Existing static headers (if any) are no longer sent. Users must authorize on first use. |
|
||||
| `oauth_user` → `none` / `static` | — | Existing `mcp_user_tokens` rows are **deleted**: the tokens are bound to the auth model + URL active at consent time, and rows left behind could silently rebind if a row with the old name/URL reappears. Switching back to `oauth_user` later starts clean — users re-consent on next use. This is **not reversible**; the AS-side grants are untouched (revoke upstream via the AS if needed). |
|
||||
| `oauth_user` → `none` / `static` | — | Existing `mcp_user_tokens` rows are **orphaned** — inert without a matching `auth_type`. Use admin bulk-revoke to drop them, or let them expire. Switching back to `oauth_user` later re-activates the orphaned rows if they haven't been deleted. |
|
||||
| OAuth `client_id` or `client_secret` rotated | — | Existing tokens may stop refreshing if the AS treats them as bound to the previous client. Bulk-revoke after rotation. |
|
||||
| `oauth_user` ↔ `oauth_obo` | — | The per-user rows are **deleted** on the flip (they mean different things: per-server AS refresh tokens vs. minted cache). `oauth_audience` and `oauth_scopes` mean different things in each model (a resource indicator vs. an IdP app identifier; AS-consent scopes vs. an rfc8693 exchange scope), so on a flip they **never carry** — each is taken from the request for the target model or set NULL. The admin console clears these fields when you change the auth type, so re-enter the correct values for the new mode; via the API, supply them explicitly (a flip into `oauth_obo` with no `oauth_audience` is rejected, and a non-empty `oauth_scopes` under the `entra` profile is rejected since that leg pins `<audience>/.default`). |
|
||||
| `oauth_obo` → `none` / `static` | — | Minted cache rows are deleted. |
|
||||
| `oauth_obo` **audience**, **URL**, or **`oauth_scopes`** changed | — | Minted cache rows are **deleted** (tokens are bound to the audience/URL/scopes at mint time), forcing a fresh mint — so an audience or scope narrowing takes effect immediately, not at token expiry. |
|
||||
|
||||
Every transition that changes what a stored row *means* deletes the rows outright — a stale consent or minted token must never be served under new semantics. There is no orphan-and-reactivate path.
|
||||
The orphan-by-default behavior is chosen so switching back to `oauth_user` is non-destructive. Bulk-revoke is the explicit cleanup path.
|
||||
|
||||
---
|
||||
|
||||
@@ -169,9 +113,5 @@ Every transition that changes what a stored row *means* deletes the rows outrigh
|
||||
| `mcp_oauth_url_insecure` | MCP server URL is `http://` (not `https://`) on a non-loopback host | Use `https://`. Per-user bearers must not transit cleartext. |
|
||||
| Tools fail in scheduled / Discord / Slack runs | OAuth-MCP requires browser-based consent | Users must pre-consent via the web UI. Phase 9 dashboard badge surfaces deferred consents from these runs on next login. |
|
||||
| Circuit breaker open repeatedly | Transport-level errors on the MCP server (DNS, TLS, 5xx) | Check the per-server error pill; auth errors do not trip the breaker. |
|
||||
| **`oauth_obo`**: every tool call fails, log shows `obo_misconfigured` | Server row has no Audience, or `obo_grant_profile` is unset/unknown | Set the Audience on the server row; set `[oidc] obo_grant_profile` to `entra` or `rfc8693`. |
|
||||
| **`oauth_obo`**: `obo_mint_rejected` with `AADSTS65001` | Turnstone's app lacks the (admin-consented) delegated grant to this MCP app — often admin consent that didn't propagate | Grant + admin-consent the delegated permission for this resource; verify it shows *Granted*. See the Entra gotcha above. |
|
||||
| **`oauth_obo`**: "Sign in to Turnstone again" on one server | Captured credential missing/rejected, or a Conditional Access challenge | User re-logs into Turnstone (re-captures the credential). If it persists, check the IdP grant / CA policy. |
|
||||
| **`oauth_obo`**: tools don't appear at all for a user | User has not signed in since `capture_user_credential` was enabled (no credential captured) | User logs out and back in via OIDC so the refresh credential is captured. |
|
||||
|
||||
See also: `docs/operations/mcp-oauth-headless.md` for the cron / channel-driven run caveat.
|
||||
|
||||
@@ -75,11 +75,6 @@ This means the model always has its most relevant memories available without
|
||||
explicit recall -- but can still use `memory(action='search')` for deeper
|
||||
lookup.
|
||||
|
||||
The persona memory lever gates this pathway: a workstream whose persona
|
||||
turns memory off receives no relevance injection at all -- the steps
|
||||
above run only when memory is enabled for the session. See
|
||||
[Personas](personas.md).
|
||||
|
||||
### Nudges
|
||||
|
||||
The metacognition layer can nudge the model to save memories at appropriate
|
||||
|
||||
+9
-46
@@ -41,7 +41,6 @@ are set.
|
||||
| `TURNSTONE_OIDC_PASSWORD_ENABLED` | No | `true` | Set to `false` to hide the password form and block all username/password logins (including admin). API tokens continue to work. |
|
||||
| `TURNSTONE_OIDC_REDIRECT_BASE` | Yes | — | Externally-reachable origin for the OIDC redirect URI (e.g. `https://app.example.com`). Without this, OIDC will refuse to start. The previous Host-header fallback was unsafe under permissive reverse proxies. |
|
||||
| `TURNSTONE_OIDC_TRUSTED_ENDPOINT_HOSTS` | No | — | Comma-separated list of additional hostnames whose endpoints the IdP discovery document is allowed to reference. See [Cross-host endpoints](#cross-host-endpoints). |
|
||||
| `TURNSTONE_OIDC_ALLOW_PRIVATE_NETWORK` | No | `false` | Allow the issuer (and its discovered endpoints) to resolve to private/internal addresses — needed for a self-hosted IdP on an internal network. See [Self-hosted and internal IdPs](#self-hosted-and-internal-idps). |
|
||||
|
||||
All four required fields — issuer, client ID, client secret, and
|
||||
`TURNSTONE_OIDC_REDIRECT_BASE` — must be set. If any are missing OIDC
|
||||
@@ -77,17 +76,17 @@ IdP from redirecting the token-exchange POST (which carries
|
||||
being aimed at internal services.
|
||||
|
||||
A few public IdPs legitimately split endpoints across hostnames. Google
|
||||
and Microsoft Entra ID are the canonical examples:
|
||||
is the canonical example:
|
||||
|
||||
| IdP | Issuer host | Cross-host endpoint(s) |
|
||||
|-----|-------------|------------------------|
|
||||
| Google | `accounts.google.com` | `oauth2.googleapis.com`, `www.googleapis.com`, `openidconnect.googleapis.com` |
|
||||
| Microsoft Entra | `login.microsoftonline.com` | `graph.microsoft.com` (userinfo) |
|
||||
| Field | Hostname |
|
||||
|-------|----------|
|
||||
| issuer | `accounts.google.com` |
|
||||
| token_endpoint | `oauth2.googleapis.com` |
|
||||
| jwks_uri | `www.googleapis.com` |
|
||||
| userinfo_endpoint | `openidconnect.googleapis.com` |
|
||||
|
||||
Both sets are built in — operators using `https://accounts.google.com` or
|
||||
`https://login.microsoftonline.com/<tenant>/v2.0` need no extra
|
||||
configuration. (Entra's discovery document advertises `userinfo_endpoint`
|
||||
on `graph.microsoft.com`, distinct from the issuer host.)
|
||||
Google's set is built in — operators using `https://accounts.google.com`
|
||||
need no extra configuration.
|
||||
|
||||
For other IdPs whose discovery document references a non-issuer host,
|
||||
extend the allow-list explicitly:
|
||||
@@ -100,40 +99,6 @@ The same scheme / no-userinfo / SSRF rules apply to allow-listed hosts —
|
||||
this knob only relaxes the same-origin check, not the security gates.
|
||||
Each entry is a hostname (no scheme, no path).
|
||||
|
||||
### Self-hosted and internal IdPs
|
||||
|
||||
By default Turnstone refuses an issuer whose hostname resolves to a
|
||||
private or internal address:
|
||||
|
||||
```
|
||||
OIDCError: endpoint URL resolves to non-public address (10.0.0.5): https://auth.example.site
|
||||
```
|
||||
|
||||
This is SSRF hardening, not a licensing or product restriction: the OIDC
|
||||
flow makes server-side HTTP requests (discovery, JWKS, token exchange),
|
||||
and refusing non-public destinations keeps a mistyped or maliciously
|
||||
steered issuer from aiming those fetches at internal services. For a
|
||||
self-hosted IdP (Keycloak, Authentik, Dex, …) on a private network,
|
||||
opt in explicitly in `config.toml`:
|
||||
|
||||
```toml
|
||||
[oidc]
|
||||
allow_private_network = true
|
||||
```
|
||||
|
||||
or via `TURNSTONE_OIDC_ALLOW_PRIVATE_NETWORK=true` (the env var wins
|
||||
when both are set).
|
||||
|
||||
The opt-in admits private-range (RFC 1918), unique-local, CGNAT
|
||||
(100.64/10 — tailnets), and loopback addresses. Link-local, multicast,
|
||||
and reserved ranges stay refused even with the opt-in — cloud metadata
|
||||
services (169.254.169.254) live there, and no legitimate IdP does. The
|
||||
HTTPS requirement and the same-origin endpoint checks are unaffected.
|
||||
|
||||
This knob only affects the login-flow IdP configured here. OAuth
|
||||
endpoints advertised by remote MCP servers are untrusted input and are
|
||||
always held to the strict public-address rule.
|
||||
|
||||
### config.toml alternative
|
||||
|
||||
```toml
|
||||
@@ -146,8 +111,6 @@ provider_name = "Google"
|
||||
role_claim = "groups"
|
||||
password_enabled = true
|
||||
redirect_base = "https://app.example.com"
|
||||
# Self-hosted IdP on an internal network (see "Self-hosted and internal IdPs")
|
||||
allow_private_network = false
|
||||
|
||||
[oidc.role_map]
|
||||
admin = "builtin-admin"
|
||||
|
||||
@@ -1,173 +0,0 @@
|
||||
# Personas
|
||||
|
||||
A **persona** is a named, reusable bundle attached to a workstream **at
|
||||
creation** that controls how its system message is composed and what
|
||||
capability envelope it runs with. Personas answer a recurring operational
|
||||
complaint: the default composition primes every session for heavy tool use,
|
||||
and there was no per-workstream dial to launch a "just write prose" or
|
||||
"evidence-first research" session.
|
||||
|
||||
A persona is exactly four levers — no more:
|
||||
|
||||
| Lever | What it does |
|
||||
|---|---|
|
||||
| **Base prompt** | Replaces the BASE module of the composed system message. *Only* BASE: ENV, CONTEXT, TOOLS, and POLICIES keep composing, so mandatory [prompt policies](governance.md) ride on top of every persona. Built-in personas source their prose from a repo file; operator personas store it inline — see [Where persona prompts live](#where-persona-prompts-live). |
|
||||
| **Tool visibility** | Which tools the session advertises. Tri-state: *unrestricted* (tracks tool growth and MCP catalogs), *no tools* (the TOOLS prompt block self-suppresses and zero definitions go on the wire), or an *exact set* of names. Including `tool_search` in a set makes it **soft** — tools the model discovers through search join the visible set; omitting it makes the set **hard** (the search pathway is disabled entirely). On commercial providers a soft set costs one prompt-cache re-prime per `tool_search` expansion, since each expansion rewrites the wire tool set and recomposes the prompt. |
|
||||
| **MCP** | Whether the workstream talks to MCP at all. **Session-wide**: off means no MCP tools for the persona's own hands *or* for in-process task agents, no resource/prompt catalogs, and no listener registrations. This lever expresses infrastructure intent, not behavior shaping. |
|
||||
| **Memory** | Whether the persona's **own hands** get memory: recalled-memory injection into the prompt, memory-directed metacognitive nudges, and the `memory` tool. Task agents keep their own envelope, and compaction spill/markers are session mechanics that are never persona-gated. An exact tool set that hides `memory` also mutes those nudges, and the compaction-resume pointer follows `recall`'s visibility. |
|
||||
|
||||
Visibility is behavior shaping, **not** a security boundary: any tool call
|
||||
that does reach the wire still clears the same approval, judge, and policy
|
||||
machinery as always. RBAC and tool policies remain the enforcement layers.
|
||||
|
||||
## Snapshot semantics — resolve once, stamp forever
|
||||
|
||||
The persona is resolved **once**, at workstream creation, and stamped into
|
||||
`workstream_config` as five keys (`persona`, `persona_prompt`,
|
||||
`persona_tools`, `persona_mcp`, `persona_memory`). From then on the session
|
||||
reads only the stamp:
|
||||
|
||||
- **Editing or archiving a persona never changes an existing workstream.**
|
||||
Rehydrate, resume, and post-compaction resume all run from the stamp.
|
||||
A mid-session REPL `/resume` adopts the target workstream's stamp for
|
||||
prompt, tools, and memory; for the MCP lever it can only narrow in
|
||||
place — adopting an MCP-off stamp drops the live MCP surface, while
|
||||
adopting an MCP-on stamp into a session whose persona dropped MCP at
|
||||
construction is refused with an error telling you to reopen the
|
||||
workstream fresh.
|
||||
- A workstream outlives its persona — an archived persona keeps labelling
|
||||
the workstreams stamped with it.
|
||||
- A partial or unparseable stamp is treated as corruption: session
|
||||
construction fails loudly rather than silently falling back to a default
|
||||
envelope the operator never chose.
|
||||
- Workstreams created before personas existed carry no stamp and keep
|
||||
legacy behavior, byte-identical to the `engineer` / `orchestrator`
|
||||
defaults below — with one exception: pre-1.7 workstreams that had
|
||||
`creative_mode` set are converted by migration `063` into full
|
||||
`writer` stamps, so they resume as writing sessions rather than as
|
||||
legacy defaults.
|
||||
- Forking (`resume_ws` on create) resumes the source's stamped persona; the
|
||||
fork does not re-resolve.
|
||||
|
||||
## Seed personas
|
||||
|
||||
Migration `063` seeds six personas. The two per-kind **defaults** carry no
|
||||
overrides at all, so a zero-touch launch behaves exactly as it did before
|
||||
personas existed:
|
||||
|
||||
| Persona | Kind | Base prompt | Tools | MCP | Memory |
|
||||
|---|---|---|---|---|---|
|
||||
| `engineer` *(default)* | interactive | stock | unrestricted | on | on |
|
||||
| `orchestrator` *(default)* | coordinator | stock | unrestricted | on | on |
|
||||
| `scribe` | interactive | custom (faithful structuring of given material) | none | off | off |
|
||||
| `researcher` | interactive | custom (evidence-first) | `read_file`, `search`, `web_fetch`, `web_search`, `recall`, `memory`, `tool_search` (soft) | off | on |
|
||||
| `writer` | interactive | custom (creative writing partner — replaces the removed `/creative`) | none | off | on |
|
||||
| `executive` | coordinator | custom (delegate, interrogate plans, judge outcomes) | spawn/inspect/lifecycle tools plus `memory`: `spawn_workstream`, `spawn_batch`, `send_to_workstream`, `wait_for_workstream`, `inspect_workstream`, `list_workstreams`, `list_nodes`, `close_workstream`, `cancel_workstream`, `memory` (hard) | off | on |
|
||||
|
||||
Notes:
|
||||
|
||||
- `scribe` turns memory off deliberately: recalled memories would
|
||||
contaminate faithful summarization with unrelated context.
|
||||
- `researcher`'s set is soft (includes `tool_search`): it starts with
|
||||
read and evidence tools but can pull in others on demand — e.g. load
|
||||
`bash` to run a snippet and verify a calculation. It is evidence-first,
|
||||
not sandboxed; any escalated tool still hits the normal approval path.
|
||||
- Coordinator sessions do not merge MCP today, so the MCP lever on
|
||||
coordinator personas is forward-compatible bookkeeping; it bites on
|
||||
interactive workstreams.
|
||||
|
||||
## Where persona prompts live
|
||||
|
||||
Prompt source is explicit in the persona row — two nullable columns, never both empty:
|
||||
|
||||
| `base_prompt_file` | `base_prompt` | Meaning |
|
||||
|---|---|---|
|
||||
| set (e.g. `scribe.md`) | — | **built-in**: prose lives in `prompts/personas/<file>`, code-owned and PR-reviewed |
|
||||
| set | set | built-in with an **operator override** layered on top (the inline text wins) |
|
||||
| — | set | **operator** persona, inline prose |
|
||||
|
||||
A `CHECK` forbids the both-empty row, so resolution is a plain coalesce —
|
||||
`base_prompt ?? load(base_prompt_file)` — with no implicit "inherit the default"
|
||||
branch in application logic. `base_prompt_file` is set only by the migration/code
|
||||
(the admin API never exposes it): it marks a persona as built-in and blocks
|
||||
archive, so `engineer` and `orchestrator` can't be removed. To customise a
|
||||
built-in, set `base_prompt` on it (clear it to revert), or create your own persona.
|
||||
|
||||
The resolved prompt is **frozen into the workstream at creation** — later edits to
|
||||
a built-in's file or an operator's row never change a running workstream; only new
|
||||
ones pick up the change. "No persona" is not a state: every workstream is stamped,
|
||||
and an empty `persona=` resolves to the kind's `is_default` (`engineer` /
|
||||
`orchestrator`).
|
||||
|
||||
## Choosing a persona
|
||||
|
||||
Every creation surface takes an optional persona; empty always means the
|
||||
kind's default (or plain legacy behavior on a database with no personas
|
||||
seeded):
|
||||
|
||||
- **Web/console**: the persona select on the console launcher, the server
|
||||
webui's new-workstream dialog, and the dashboard composer. Selecting a
|
||||
persona requires **no** `persona.*` permission — the picker feed
|
||||
(`GET /v1/api/personas`) is authenticated-only and returns display fields.
|
||||
- **API/SDK**: `CreateWorkstreamRequest.persona` (Python:
|
||||
`create_workstream(persona=...)`; TypeScript: `{ persona: ... }`).
|
||||
- **CLI**: `turnstone --persona <name>`. Unknown or disabled names error at
|
||||
startup. `--resume` ignores `--persona` and adopts the resumed
|
||||
workstream's stamp.
|
||||
- **Coordinator spawn**: `spawn_workstream` / `spawn_batch` take a
|
||||
`persona` argument, validated when the coordinator prepares the spawn
|
||||
and re-checked by the node that creates the child (children are always
|
||||
interactive-kind). Omitted means the interactive **default** — a child
|
||||
never inherits its parent coordinator's persona.
|
||||
- **Sub-agents**: `task_agent` takes a `persona` argument setting the
|
||||
sub-agent's identity and capability envelope (resolved against
|
||||
interactive-kind personas, frozen into the task at prep). Omitted keeps
|
||||
the default autonomous task-agent identity — never the parent's persona.
|
||||
|
||||
## How agents discover personas
|
||||
|
||||
Agents are told, not expected to guess: the live persona list (enabled,
|
||||
interactive-kind — children and sub-agents are always interactive) is
|
||||
injected into the `persona` parameter description of `task_agent`,
|
||||
`spawn_workstream`, and `spawn_batch` whenever the session's tool surface
|
||||
is rendered — session start, MCP catalog change, model-registry reload.
|
||||
Each entry carries the name, the default marker, and the persona's
|
||||
one-line description so the model can pick by purpose (descriptions drop
|
||||
out past 25 personas; the name list always enumerates completely).
|
||||
|
||||
A persona created after that render is still reachable — pass its name.
|
||||
Every resolve failure enumerates the names currently valid for the kind,
|
||||
so a stale list (or a typo) self-corrects on the next attempt.
|
||||
|
||||
Resolution is forgiving on all surfaces (they share one rule):
|
||||
|
||||
- names match case-insensitively (`Writer` resolves `writer`);
|
||||
- an input that uniquely matches a persona's **display name**
|
||||
(case-insensitive, among the kind's enabled personas — display names are
|
||||
not unique, and a same-label persona of another kind neither blocks nor
|
||||
wins) resolves to that persona; an ambiguous match errors, listing the
|
||||
candidate slugs;
|
||||
- whatever variant matched, the stamped identity, approval chrome, and
|
||||
wire always carry the canonical `name` slug.
|
||||
|
||||
## Authoring (console)
|
||||
|
||||
Personas are managed in the console's **Manage → Governance → Personas**
|
||||
tab. The admin shelf exposes exactly the four levers plus the kind
|
||||
list, the default marker, and archive. Rules:
|
||||
|
||||
- `name` is an immutable lowercase slug — and the identifier agents and
|
||||
the CLI launch the persona by (`persona=` on the spawn tools,
|
||||
`--persona` on the CLI); the create shelf says so under **Name**.
|
||||
`display_name` is a list label, editable any time, and deliberately
|
||||
not an identifier (a unique display name happens to resolve, as a
|
||||
forgiveness fallback — don't design workflows around it).
|
||||
- Exactly one default per kind, storage-enforced: flipping the flag on a
|
||||
successor demotes the incumbent atomically, defaults are single-kind,
|
||||
and a default cannot be archived.
|
||||
- **Archive only** — there is no delete verb, so every stamped
|
||||
workstream's provenance stays explicable.
|
||||
|
||||
RBAC: `persona.create` / `persona.read` / `persona.write` gate the admin
|
||||
CRUD (`/v1/api/admin/personas`); all three are granted to `builtin-admin`
|
||||
by migration `063`, and other roles opt in via role permission overrides.
|
||||
+2
-2
@@ -69,7 +69,7 @@ Both `TurnstoneServer` (sync) and `AsyncTurnstoneServer` (async) expose:
|
||||
|----------|--------|---------|
|
||||
| **Workstreams** | `list_workstreams()` | `ListWorkstreamsResponse` |
|
||||
| | `dashboard()` | `DashboardResponse` |
|
||||
| | `create_workstream(*, name, model, auto_approve, skill, persona, initial_message, attachments)` | `CreateWorkstreamResponse` |
|
||||
| | `create_workstream(*, name, model, auto_approve, skill, initial_message, attachments)` | `CreateWorkstreamResponse` |
|
||||
| | `close_workstream(ws_id)` | `StatusResponse` |
|
||||
| **Attachments** | `upload_attachment(ws_id, filename, data, *, mime_type=...)` | `UploadAttachmentResponse` |
|
||||
| | `list_attachments(ws_id)` | `ListAttachmentsResponse` |
|
||||
@@ -100,7 +100,7 @@ Both `TurnstoneConsole` (sync) and `AsyncTurnstoneConsole` (async) expose:
|
||||
| | `workstreams(*, state, node, search, sort, page, per_page)` | `ClusterWorkstreamsResponse` |
|
||||
| | `node_detail(node_id)` | `NodeDetailResponse` |
|
||||
| | `snapshot()` | `ClusterSnapshotResponse` |
|
||||
| | `create_workstream(*, node_id, name, model, initial_message, skill, persona)` | `ConsoleCreateWsResponse` |
|
||||
| | `create_workstream(*, node_id, name, model, initial_message, skill)` | `ConsoleCreateWsResponse` |
|
||||
| **Schedules** | `list_schedules()` | `ListSchedulesResponse` |
|
||||
| | `create_schedule(*, name, schedule_type, initial_message, ...)` | `ScheduleInfo` |
|
||||
| | `get_schedule(task_id)` | `ScheduleInfo` |
|
||||
|
||||
@@ -54,23 +54,6 @@ When a per-model override is `NULL` (empty in the UI), the global default is
|
||||
used. Switching models via `/model <alias>` re-resolves sampling parameters
|
||||
from the new model's overrides or global defaults.
|
||||
|
||||
### Responses output controls (per-model)
|
||||
|
||||
Models whose capability table declares Responses output controls expose two
|
||||
additional fields in the Models create/edit shelf:
|
||||
|
||||
| Field | Stored capability | Values | Effect |
|
||||
|-------|-------------------|--------|--------|
|
||||
| Output verbosity | `verbosity` | `low`, `medium`, `high` | Controls answer length independently of reasoning effort. |
|
||||
| Reasoning mode | `reasoning_mode` | `standard`, `pro` | Selects standard or higher-compute Pro execution without changing the model ID. |
|
||||
|
||||
An empty selection means provider default and omits the capability key. Known
|
||||
GPT-5.6 models inherit support from the built-in table without persisting
|
||||
redundant support flags. An OpenAI-compatible model pinned to the Responses API
|
||||
can opt in with the `supports_verbosity` and `supports_pro_mode` capability
|
||||
tiles. Chat Completions and non-Responses providers do not surface or submit
|
||||
these controls.
|
||||
|
||||
**Removed settings:** `model.name` and `model.context_window` have been removed
|
||||
from ConfigStore. Model names and context windows are now configured per-model
|
||||
in the Models tab. A startup warning is logged if these keys appear in
|
||||
|
||||
+18
-77
@@ -1,6 +1,6 @@
|
||||
# Tools Reference
|
||||
|
||||
turnstone exposes 17 built-in tools plus any number of external MCP tools to the
|
||||
turnstone exposes 16 built-in tools plus any number of external MCP tools to the
|
||||
LLM via the OpenAI function-calling interface. Built-in tools are defined as JSON
|
||||
files under `turnstone/tools/` and loaded at startup by `turnstone/core/tools.py`.
|
||||
MCP tools are discovered from configured MCP servers at startup by
|
||||
@@ -28,19 +28,13 @@ schema plus turnstone-specific metadata keys:
|
||||
}
|
||||
```
|
||||
|
||||
**Metadata keys** (stripped before sending the schema to the model; the full
|
||||
set lives in `_META_KEYS` in `turnstone/core/tools.py`):
|
||||
**Metadata keys** (stripped before sending the schema to the model):
|
||||
|
||||
| Key | Type | Meaning |
|
||||
|------------------|------|---------|
|
||||
| `task_agent` | bool | Tool is available to task sub-agents. |
|
||||
| `coordinator` | bool | Tool is available to coordinator sessions. Without `interactive: true` alongside it, this reads as coord-only and the tool is stripped from interactive sessions. |
|
||||
| `interactive` | bool | Opt a `coordinator: true` tool back into interactive sessions (dual-kind tools like `memory`). |
|
||||
| `auto_approve` | bool | Tool runs without user confirmation (read-only, safe operations). |
|
||||
| `primary_key` | str | When the model sends a bare string instead of JSON args, map it to this parameter name. |
|
||||
| `kind_variants` | dict | Per-kind description / parameter-schema overlays so each session kind sees only the surface it can use (see `memory.json`). |
|
||||
| `cwd_note` | str | Sentence appended to the description at session build time with `{working_dir}` substituted — declare on tools whose semantics depend on the process working directory (see `bash.json`, `apply_cwd_context`). |
|
||||
| `workspace_note` | str | Companion sentence naming the operator-configured workspace directory, `{workspace_dir}` substituted; dropped when no workspace is configured. |
|
||||
| Key | Type | Meaning |
|
||||
|----------------|------|---------|
|
||||
| `task_agent` | bool | Tool is available to task sub-agents. |
|
||||
| `auto_approve` | bool | Tool runs without user confirmation (read-only, safe operations). |
|
||||
| `primary_key` | str | When the model sends a bare string instead of JSON args, map it to this parameter name. |
|
||||
|
||||
---
|
||||
|
||||
@@ -50,10 +44,10 @@ set lives in `_META_KEYS` in `turnstone/core/tools.py`):
|
||||
|
||||
| Name | Description |
|
||||
|---------------------|-------------|
|
||||
| `TOOLS` | All 29 loaded built-in tool definitions (interactive + coordinator union). Sessions send a kind-specific subset (`INTERACTIVE_TOOLS` or `COORDINATOR_TOOLS`). |
|
||||
| `TOOLS` | All 28 loaded built-in tool definitions (interactive + coordinator union). Sessions send a kind-specific subset (`INTERACTIVE_TOOLS` or `COORDINATOR_TOOLS`). |
|
||||
| `TASK_AGENT_TOOLS` | Tools with `task_agent: true` -- available to task sub-agents. Includes write operations. |
|
||||
| `TASK_AUTO_TOOLS` | Set of all tool names with `auto_approve: true` -- used by task-agent sub-sessions to skip confirmation for matching available tools. |
|
||||
| `BUILTIN_TOOL_NAMES`| Frozenset of all 29 built-in tool names (interactive + coordinator union). Used by tool search to distinguish always-on tools from deferrable MCP tools. |
|
||||
| `BUILTIN_TOOL_NAMES`| Frozenset of all 28 built-in tool names (interactive + coordinator union). Used by tool search to distinguish always-on tools from deferrable MCP tools. |
|
||||
| `PRIMARY_KEY_MAP` | Dict mapping tool name to its `primary_key` parameter name. |
|
||||
|
||||
---
|
||||
@@ -71,7 +65,7 @@ Tool execution follows a three-phase pipeline inside `ChatSession._execute_tools
|
||||
- Parses the JSON arguments (with fallback for malformed JSON).
|
||||
- If JSON parsing fails entirely, uses `PRIMARY_KEY_MAP` to map a bare string
|
||||
to the correct parameter.
|
||||
- Dispatches to the matching `_prepare_{func_name}()` handler. There are 17
|
||||
- Dispatches to the matching `_prepare_{func_name}()` handler. There are 16
|
||||
built-in tools plus `tool_search` (synthetic, client-side BM25 fallback) and
|
||||
the generic `_prepare_mcp_tool()` handler for MCP tools.
|
||||
- Validates arguments and builds a preview dict containing:
|
||||
@@ -131,9 +125,6 @@ Each item's `execute` callable is invoked:
|
||||
- `web_fetch` -- fetches a URL (SSRF-protected, but makes network requests)
|
||||
- `web_search` -- web search via self-hosted SearxNG (makes network requests)
|
||||
- `task_agent` -- spawns an autonomous sub-agent
|
||||
- `open_preview` -- **URL targets only** (network access, gated like `web_fetch`);
|
||||
file-path and `attachment:` targets are local reads and run unprompted like
|
||||
`read_file`
|
||||
|
||||
Note: The JSON schema metadata key `auto_approve` controls membership in
|
||||
`TASK_AUTO_TOOLS` (used for task agent sub-sessions). The actual runtime
|
||||
@@ -166,7 +157,6 @@ Every tool defines a `primary_key`. The mapping is:
|
||||
| `search` | `query` |
|
||||
| `web_fetch` | `url` |
|
||||
| `web_search` | `query` |
|
||||
| `open_preview` | `target` |
|
||||
| `task_agent` | `prompt` |
|
||||
| `memory` | `name` |
|
||||
| `recall` | `query` |
|
||||
@@ -295,7 +285,7 @@ Fetch a URL and extract specific information from it.
|
||||
| `url` | string | yes | The URL to fetch (must start with `http://` or `https://`). |
|
||||
| `question` | string | yes | What to extract or answer from the page content. |
|
||||
|
||||
- **What it does**: Fetches the URL, strips HTML to plain text, and uses the LLM to extract the answer to the question from the page content. Every redirect hop is SSRF-screened before it is requested. Private/internal addresses are refused by default; enable `tools.allow_private_network` (console Settings → Tools) to make them approvable for self-hosted setups whose services live on the local network — the approval prompt marks such requests, and a public site redirecting into private space is refused regardless.
|
||||
- **What it does**: Fetches the URL, strips HTML to plain text, and uses the LLM to extract the answer to the question from the page content. Protected against SSRF (blocks private/internal IPs).
|
||||
- **Auto-approve**: No -- requires user confirmation (makes network requests).
|
||||
- **Agent availability**: `task_agent`.
|
||||
|
||||
@@ -355,39 +345,6 @@ It reports the score scale, whether the endpoint cleanly separates relevant from
|
||||
|
||||
---
|
||||
|
||||
### open_preview
|
||||
|
||||
Show the user rich content in a preview pane beside the conversation.
|
||||
|
||||
| Parameter | Type | Required | Description |
|
||||
|-----------|--------|----------|-------------|
|
||||
| `target` | string | yes | An http(s) URL, a file path, or `attachment:<id>` for a file attached to the conversation. |
|
||||
| `kind` | string | no | Rendering override: `web`, `pdf`, `image`, `table`, `text`, or `markdown`. Detected from the content when omitted. |
|
||||
| `title` | string | no | Pane header title. Defaults to the page title, filename, or URL. |
|
||||
|
||||
- **What it does**: Resolves the target to bytes (URLs fetch through the same
|
||||
SSRF-guarded path as `web_fetch`, screened per redirect hop, honoring the
|
||||
same `tools.allow_private_network` opt-in), classifies the
|
||||
content, stores it content-addressed against the workstream, and opens the
|
||||
frontend preview pane beside the conversation: web pages render in a fully
|
||||
sandboxed iframe (no scripts, opaque origin), PDFs in the browser viewer,
|
||||
images inline, CSV/TSV/JSON as a sortable table, text/markdown rendered. A
|
||||
previewed web page loads none of its remote images or styles by default, so
|
||||
opening it never reveals the viewer to the page's site; a toggle in the pane
|
||||
header turns remote content back on for that preview. The
|
||||
model receives only a one-line confirmation — to reason about content, use
|
||||
`web_fetch` / `read_file` instead. Preview content is size-capped per kind
|
||||
(pages 4 MB, PDFs 32 MB, images 4 MB, tables 2 MB, text 512 KB) and GC'd
|
||||
with the workstream.
|
||||
- **Auto-approve**: URL targets require confirmation (network access); file
|
||||
paths and `attachment:` targets run unprompted (local reads).
|
||||
- **Agent availability**: interactive sessions only (not `task_agent`, not
|
||||
coordinators).
|
||||
- **Surfaces**: the pane renders in the web UI (standalone and console). The
|
||||
CLI prints the confirmation line only — there is no terminal pane.
|
||||
|
||||
---
|
||||
|
||||
## Agent
|
||||
|
||||
The tool name uses the `_agent` suffix — bare `task` collides with
|
||||
@@ -588,7 +545,6 @@ pre-configure skills at workstream creation.
|
||||
| `search` | File Ops | Yes | Yes | `query` |
|
||||
| `web_fetch` | Info | No | Yes | `url` |
|
||||
| `web_search` | Info | No | Yes | `query` |
|
||||
| `open_preview`| Info | URL: no; path/attachment: yes | No | `target` |
|
||||
| `task_agent` | Agent | No | No | `prompt` |
|
||||
| `memory` | Memory | Yes | No | `name` |
|
||||
| `recall` | Memory | Yes | No | `query` |
|
||||
@@ -624,11 +580,6 @@ Tool search uses the best available mechanism for each provider:
|
||||
`_exec_tool_search()` runs a pure-Python BM25 index over tool names and
|
||||
descriptions, then expands the matched tools into the visible set.
|
||||
|
||||
A persona with a tool-visibility set overrides this selection: any exact
|
||||
set forces tool search into the client-side BM25 mechanism (tier 3)
|
||||
regardless of provider, and a **hard** set — one whose visible tools omit
|
||||
`tool_search` — disables tool search entirely.
|
||||
|
||||
### Configuration
|
||||
|
||||
Tool search is configured in `config.toml` under the `[tools]` section:
|
||||
@@ -698,7 +649,7 @@ MCP-compatible service.
|
||||
3. **Schema conversion**: Each MCP tool's `inputSchema` is converted to OpenAI
|
||||
function-calling format. The tool name is prefixed: `mcp__{server}__{tool}`.
|
||||
|
||||
4. **Merging**: MCP tools are appended after the 17 built-in tools via
|
||||
4. **Merging**: MCP tools are appended after the 16 built-in tools via
|
||||
`merge_mcp_tools()`. Built-in tools appear first, giving them natural LLM priority.
|
||||
When dynamic tool search is active, MCP tools are deferred rather than directly
|
||||
visible -- the model discovers them via search as needed (see
|
||||
@@ -785,10 +736,7 @@ MCP tool lists stay up-to-date without restart through two mechanisms:
|
||||
1. **Push notifications** -- MCP servers that declare `tools.listChanged: true` in
|
||||
their capabilities send `notifications/tools/list_changed` when their tool list
|
||||
changes. `MCPClientManager` registers a `message_handler` on each `ClientSession`
|
||||
that triggers an immediate refresh for that server (debounced per server and
|
||||
notification kind, and run off the receive loop). A refresh that fails while
|
||||
the connection stays up is retried automatically on the next health-loop tick
|
||||
until one completes.
|
||||
that triggers an immediate refresh for that server.
|
||||
|
||||
2. **Manual** -- `/mcp refresh` re-fetches tools from all servers immediately.
|
||||
`/mcp refresh <server>` targets a single server. If a server has disconnected,
|
||||
@@ -796,10 +744,6 @@ MCP tool lists stay up-to-date without restart through two mechanisms:
|
||||
same controls (refresh / reconnect buttons per server) for cluster-wide
|
||||
fan-out.
|
||||
|
||||
Reconnects (health-loop, dispatch-driven, or operator-forced) always end in a
|
||||
full catalog rediscovery, so a server that changed its tools while disconnected
|
||||
comes back current.
|
||||
|
||||
When tools change, `MCPClientManager` rebuilds its merged tool list using copy-on-write
|
||||
(new list/dict objects assigned atomically) and notifies all active `ChatSession`
|
||||
instances via registered listener callbacks. Each session rebuilds its `_tools`,
|
||||
@@ -870,16 +814,13 @@ catalog.
|
||||
|
||||
### Refresh
|
||||
|
||||
Resource lists stay current through the same mechanisms as tool lists:
|
||||
Resource lists stay current through the same three-tier mechanism as tool lists:
|
||||
|
||||
1. **Push** -- Servers declaring `resources.listChanged: true` send
|
||||
`notifications/resources/list_changed`, triggering an immediate refresh
|
||||
(with the same failed-refresh retry on the health-loop tick).
|
||||
2. **Manual** -- `/mcp refresh` re-fetches resources alongside tools.
|
||||
|
||||
Servers without push support are refreshed whenever they reconnect (every
|
||||
reconnect ends in full rediscovery) or when an operator refreshes manually;
|
||||
there is no periodic polling.
|
||||
`notifications/resources/list_changed`, triggering an immediate refresh.
|
||||
2. **Periodic** -- Servers without push are polled on the configured refresh
|
||||
interval (default 4 hours, same timer as tools).
|
||||
3. **Manual** -- `/mcp refresh` re-fetches resources alongside tools.
|
||||
|
||||
---
|
||||
|
||||
|
||||
@@ -1,56 +0,0 @@
|
||||
{
|
||||
"defaults": {
|
||||
"n_runs": 3
|
||||
},
|
||||
"cases": [
|
||||
{
|
||||
"id": "search-first",
|
||||
"skill": {
|
||||
"name": "search-first",
|
||||
"content": "# Search First\n\nBefore answering ANY question about where something lives in the codebase, you MUST call the `search` tool first. Never answer from memory."
|
||||
},
|
||||
"user_prompt": "Where is JWT token validation implemented in this project?",
|
||||
"expected_actions": [{ "tool": "search" }],
|
||||
"match_mode": "ordered_subset",
|
||||
"max_turns": 4
|
||||
},
|
||||
{
|
||||
"id": "test-after-edit",
|
||||
"skill": {
|
||||
"name": "test-after-edit",
|
||||
"content": "# Test After Edit\n\nAfter editing or writing ANY file, you MUST run the test suite with `python -m pytest` via bash before you finish. Do not report done until tests have run."
|
||||
},
|
||||
"user_prompt": "Add a function `clamp(x, lo, hi)` that clamps x to [lo, hi] in utils.py.",
|
||||
"setup": {
|
||||
"files": {
|
||||
"utils.py": ""
|
||||
}
|
||||
},
|
||||
"expected_actions": [
|
||||
{ "tool": "write_file" },
|
||||
{ "tool": "bash", "args_pattern": { "command": "pytest" } }
|
||||
],
|
||||
"match_mode": "ordered_subset",
|
||||
"max_turns": 8
|
||||
},
|
||||
{
|
||||
"id": "changelog-update",
|
||||
"skill": {
|
||||
"name": "changelog-update",
|
||||
"content": "# Changelog Discipline\n\nWhenever you modify a file, you MUST also append a one-line entry to CHANGELOG.md describing the change in the same task."
|
||||
},
|
||||
"user_prompt": "Fix the off-by-one so pager.py shows the last page. Edit pager.py.",
|
||||
"setup": {
|
||||
"files": {
|
||||
"pager.py": "def last_page(total_items, per_page):\n # off-by-one: drops the final partial page\n return total_items // per_page\n",
|
||||
"CHANGELOG.md": "# Changelog\n"
|
||||
}
|
||||
},
|
||||
"expected_actions": [
|
||||
{ "tool": "edit_file", "args_pattern": { "path": "CHANGELOG.md" } }
|
||||
],
|
||||
"match_mode": "subset",
|
||||
"max_turns": 8
|
||||
}
|
||||
]
|
||||
}
|
||||
+4
-6
@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
|
||||
|
||||
[project]
|
||||
name = "turnstone"
|
||||
version = "1.8.0a4"
|
||||
version = "1.7.0a6"
|
||||
description = "Multi-node AI orchestration platform with tool use, agent routing, and cluster simulation."
|
||||
readme = "README.md"
|
||||
license = "Apache-2.0"
|
||||
@@ -23,7 +23,7 @@ classifiers = [
|
||||
"Topic :: Scientific/Engineering :: Artificial Intelligence",
|
||||
]
|
||||
dependencies = [
|
||||
"openai>=2.45", # GPT-5.6: typed reasoning.mode, prompt_cache_options, and cache_write_tokens
|
||||
"openai>=2.37",
|
||||
"anthropic>=0.108", # claude-fable-5 support; hard runtime floor is 0.105 (mid-conversation system blocks)
|
||||
"httpx>=0.28",
|
||||
"mcp>=1.27,<2", # v2 is a breaking rewrite (2.0.0a1 live 2026-06-11; stable ~2026-07-27) — streamablehttp_client removed, 2-tuple transport, snake_case types; migrate deliberately
|
||||
@@ -64,8 +64,7 @@ all = ["turnstone[discord,slack]"]
|
||||
|
||||
[project.scripts]
|
||||
turnstone = "turnstone.cli:main"
|
||||
turnstone-eval = "turnstone.eval.cli:main"
|
||||
turnstone-optimizer = "turnstone.optimizer:main"
|
||||
turnstone-eval = "turnstone.eval:main"
|
||||
turnstone-server = "turnstone.server:main"
|
||||
turnstone-console = "turnstone.console.server:main"
|
||||
turnstone-admin = "turnstone.admin:main"
|
||||
@@ -88,7 +87,7 @@ include = [
|
||||
"turnstone/console/static/coordinator/*.js",
|
||||
"turnstone/shared_static/*.css",
|
||||
"turnstone/shared_static/*.js",
|
||||
"turnstone/shared_static/katex-0.18.1/**/*",
|
||||
"turnstone/shared_static/katex-0.17.0/**/*",
|
||||
"turnstone/shared_static/hljs-11.11.1/**/*",
|
||||
"turnstone/shared_static/mermaid-11.16.0/**/*",
|
||||
"turnstone/shared_static/hls-1.6.16/**/*",
|
||||
@@ -103,7 +102,6 @@ testpaths = ["tests"]
|
||||
markers = [
|
||||
"live: requires a running LLM backend",
|
||||
"allow_thread_leak: test intentionally leaves a background thread running (opts out of the leaked-thread guard)",
|
||||
"e2e_recovery: opt-in end-to-end SSE recovery harness (real server + real SSE consumers, scripted provider — NOT live, no LLM backend needed); tens of seconds each. CI lanes run ``-m 'not live and not e2e_recovery'``; select with ``-m e2e_recovery``.",
|
||||
]
|
||||
filterwarnings = [
|
||||
# mcp v1 deprecates streamablehttp_client for an entry point whose call
|
||||
|
||||
@@ -4,9 +4,8 @@
|
||||
#
|
||||
# curl -fsSL https://raw.githubusercontent.com/turnstonelabs/turnstone/main/run.sh | bash
|
||||
#
|
||||
# Autodetects your distro — Ubuntu/Debian, Fedora/RHEL, Arch, their common
|
||||
# derivatives (Mint, Pop!_OS, Nobara, AlmaLinux, …), and WSL on any of them —
|
||||
# and:
|
||||
# Autodetects your distro (Ubuntu/Debian, Fedora/RHEL, Arch, and WSL on any of
|
||||
# them) and:
|
||||
# 1. ensures git is installed, then clones the repo
|
||||
# 2. ensures Docker + the compose plugin are installed and the daemon is usable
|
||||
# 3. asks how many server nodes to run (1-10)
|
||||
@@ -66,18 +65,12 @@ ask() {
|
||||
|
||||
# -- distro / package manager detection --------------------------------------
|
||||
OS_ID=""; OS_LIKE=""; PKG=""; IS_WSL=0; SUDO=""
|
||||
# Extra os-release fields, captured only to pick Docker's upstream repo when
|
||||
# get.docker.com refuses a derivative it doesn't recognize (see install_docker).
|
||||
OS_PLATFORM_ID=""; OS_CODENAME=""; OS_UBUNTU_CODENAME=""
|
||||
|
||||
detect_os() {
|
||||
if [ -r /etc/os-release ]; then
|
||||
# shellcheck disable=SC1091
|
||||
. /etc/os-release
|
||||
OS_ID="${ID:-}"; OS_LIKE="${ID_LIKE:-}"
|
||||
OS_PLATFORM_ID="${PLATFORM_ID:-}"
|
||||
OS_CODENAME="${VERSION_CODENAME:-}"
|
||||
OS_UBUNTU_CODENAME="${UBUNTU_CODENAME:-}"
|
||||
fi
|
||||
if grep -qiE 'microsoft|wsl' /proc/version 2>/dev/null || [ -n "${WSL_DISTRO_NAME:-}" ]; then
|
||||
IS_WSL=1
|
||||
@@ -137,83 +130,11 @@ clone_repo() {
|
||||
# -- docker -------------------------------------------------------------------
|
||||
DOCKER="docker"
|
||||
|
||||
# Fallback when get.docker.com won't install here. That script keys off $ID alone
|
||||
# (never ID_LIKE), so it aborts with "Unsupported distribution '<id>'" on every
|
||||
# derivative — Nobara, Linux Mint, Pop!_OS, AlmaLinux, Oracle Linux, … — even
|
||||
# though the family is clear. We already know the family from detect_os, so we add
|
||||
# Docker's official CE repo for the matching upstream and install the same
|
||||
# packages get.docker.com would (including the compose plugin the rest of run.sh
|
||||
# relies on).
|
||||
install_docker_ce_repo() {
|
||||
local up
|
||||
case "$PKG" in
|
||||
apt)
|
||||
local codename arch
|
||||
# UBUNTU_CODENAME is set by Ubuntu and every Ubuntu-derived distro
|
||||
# (Mint/Pop!_OS/Zorin/…) and never by pure Debian, so it both routes
|
||||
# the family and gives the exact codename Docker's repo expects.
|
||||
if [ -n "$OS_UBUNTU_CODENAME" ]; then
|
||||
up=ubuntu; codename="$OS_UBUNTU_CODENAME"
|
||||
else
|
||||
up=debian; codename="$OS_CODENAME"
|
||||
fi
|
||||
[ -n "$codename" ] || die "couldn't determine the $up release codename for Docker's repo — install Docker manually and re-run."
|
||||
arch="$(dpkg --print-architecture 2>/dev/null || echo amd64)"
|
||||
info "Adding Docker's $up repository ($codename)."
|
||||
$SUDO install -m 0755 -d /etc/apt/keyrings
|
||||
curl -fsSL "https://download.docker.com/linux/$up/gpg" | $SUDO tee /etc/apt/keyrings/docker.asc >/dev/null
|
||||
$SUDO chmod a+r /etc/apt/keyrings/docker.asc
|
||||
printf 'deb [arch=%s signed-by=/etc/apt/keyrings/docker.asc] https://download.docker.com/linux/%s %s stable\n' \
|
||||
"$arch" "$up" "$codename" | $SUDO tee /etc/apt/sources.list.d/docker.list >/dev/null
|
||||
$SUDO apt-get update -y
|
||||
$SUDO apt-get install -y docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin
|
||||
;;
|
||||
dnf|yum)
|
||||
# A Fedora spin and a RHEL clone can both carry "fedora" in ID_LIKE
|
||||
# (Nobara's is "rhel centos fedora"), so ID_LIKE can't separate them.
|
||||
# PLATFORM_ID can: Fedora is platform:fNN, Enterprise Linux platform:elN.
|
||||
case "$OS_PLATFORM_ID" in
|
||||
platform:f*) up=fedora ;;
|
||||
platform:el*) up=centos ;;
|
||||
*) if [ -e /etc/fedora-release ]; then up=fedora; else up=centos; fi ;;
|
||||
esac
|
||||
info "Adding Docker's $up repository."
|
||||
$SUDO curl -fsSL "https://download.docker.com/linux/$up/docker-ce.repo" \
|
||||
-o /etc/yum.repos.d/docker-ce.repo \
|
||||
|| die "couldn't add Docker's $up repository — install Docker manually and re-run."
|
||||
pkg_install docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin
|
||||
;;
|
||||
esac
|
||||
}
|
||||
|
||||
# The distro IDs get.docker.com installs directly: it matches $ID against this
|
||||
# exact set (ignoring ID_LIKE) and aborts on anything else. Mirrors the dispatch
|
||||
# in get.docker.com, including its fedora-asahi-remix -> fedora alias.
|
||||
get_docker_com_supports() {
|
||||
case "$1" in
|
||||
ubuntu|debian|raspbian|centos|fedora|rhel|rocky|sles|fedora-asahi-remix) return 0 ;;
|
||||
*) return 1 ;;
|
||||
esac
|
||||
}
|
||||
|
||||
install_docker() {
|
||||
case "$PKG" in
|
||||
apt|dnf|yum)
|
||||
# Decide up front which installer applies, rather than treating every
|
||||
# get.docker.com failure as "unsupported distro": for an ID it knows,
|
||||
# let it run and surface any real failure (network, apt lock, EOL) via
|
||||
# die instead of masking it with the repo path. Only unrecognized
|
||||
# derivatives (Nobara, Mint, …) — which it would just abort on — skip
|
||||
# straight to adding Docker's repo ourselves.
|
||||
if [ -n "$OS_ID" ] && ! get_docker_com_supports "$OS_ID"; then
|
||||
info "get.docker.com doesn't support '$OS_ID' — using Docker's official repository directly."
|
||||
install_docker_ce_repo
|
||||
else
|
||||
info "Installing Docker via the official get.docker.com script"
|
||||
curl -fsSL https://get.docker.com | $SUDO sh \
|
||||
|| die "get.docker.com failed to install Docker (see the output above). Fix the issue and re-run — the script resumes."
|
||||
fi
|
||||
;;
|
||||
info "Installing Docker via the official get.docker.com script"
|
||||
curl -fsSL https://get.docker.com | $SUDO sh ;;
|
||||
pacman)
|
||||
pkg_install docker docker-compose ;;
|
||||
esac
|
||||
@@ -445,15 +366,12 @@ ${GREEN}${BOLD}Turnstone is running${RESET} (${NODE_COUNT} node$([ "$NODE_COUNT"
|
||||
${DIM}cd $INSTALL_DIR && $DOCKER compose exec caddy cat /data/caddy/pki/authorities/local/root.crt${RESET}
|
||||
|
||||
Finish setup
|
||||
1. Open ${BOLD}${url}${RESET} and create the admin account when prompted —
|
||||
the first user created there gets full admin access.
|
||||
2. Log in, then add a model backend in the ${BOLD}Models${RESET} tab —
|
||||
1. Create the first admin user:
|
||||
${DIM}cd $INSTALL_DIR && $DOCKER compose exec node-1 turnstone-admin create-user --username admin --name "Admin"${RESET}
|
||||
2. Open ${url}, log in, and add a model backend in the ${BOLD}Models${RESET} tab —
|
||||
a local server (vLLM / llama.cpp) or an OpenAI / Anthropic / Gemini key.
|
||||
Nodes boot without a model and pick it up live; no restart needed.
|
||||
|
||||
${DIM}No browser? Create the admin from the CLI instead:
|
||||
cd $INSTALL_DIR && $DOCKER compose exec node-1 turnstone-admin create-admin --username admin --name "Admin"${RESET}
|
||||
|
||||
Scale Running ${scale}
|
||||
|
||||
Manage ${DIM}cd $INSTALL_DIR${RESET}
|
||||
|
||||
+31
-5
@@ -399,7 +399,7 @@ CONSOLE_TEMPLATE = """<!doctype html>
|
||||
known: true,
|
||||
capabilities: {
|
||||
context_window: 200000, supports_tools: true,
|
||||
supports_vision: true,
|
||||
supports_streaming: true, supports_vision: true,
|
||||
supports_web_search: true, supports_temperature: true,
|
||||
supports_effort: true,
|
||||
},
|
||||
@@ -1267,6 +1267,12 @@ PERF_TEMPLATE = """<!doctype html>
|
||||
let phase = "mount";
|
||||
try {
|
||||
const pane = new InteractivePane("perf-ws");
|
||||
// ?window= overrides the pane's transcript window (message count),
|
||||
// e.g. ?window=100000 disables windowing to isolate the
|
||||
// content-visibility/block-flow effect from the windowing effect.
|
||||
// Default (0) measures shipped behavior.
|
||||
const WINDOW = parseInt(q.get("window") || "0", 10);
|
||||
if (WINDOW > 0) pane._historyWindow = WINDOW;
|
||||
document.getElementById("mount").appendChild(pane.el);
|
||||
const msgs = buildHistory(N);
|
||||
report.heap_start = heapBytes();
|
||||
@@ -1576,7 +1582,14 @@ def _await_report(
|
||||
|
||||
|
||||
def _perf_run_one(
|
||||
chrome: str, out: Path, port: int, store: _PerfStore, n: int, turns: int, timeout: float
|
||||
chrome: str,
|
||||
out: Path,
|
||||
port: int,
|
||||
store: _PerfStore,
|
||||
n: int,
|
||||
turns: int,
|
||||
timeout: float,
|
||||
extra_query: str = "",
|
||||
) -> dict[str, object] | None:
|
||||
"""One headless-Chrome perf pass; returns the page's report or None."""
|
||||
base_flags = [
|
||||
@@ -1602,6 +1615,8 @@ def _perf_run_one(
|
||||
url = (
|
||||
f"http://127.0.0.1:{port}/perf/livepass.html?n={n}&turns={turns}&post=1&run={run_token}"
|
||||
)
|
||||
if extra_query:
|
||||
url += "&" + extra_query.lstrip("&")
|
||||
store.event.clear()
|
||||
store.data = None
|
||||
profile = out / f".chrome-perf-{n}"
|
||||
@@ -1624,7 +1639,9 @@ def _perf_run_one(
|
||||
return None
|
||||
|
||||
|
||||
def run_perf(out: Path, sizes: list[int], turns: int, timeout: float) -> bool:
|
||||
def run_perf(
|
||||
out: Path, sizes: list[int], turns: int, timeout: float, extra_query: str = ""
|
||||
) -> bool:
|
||||
"""Build, serve, and run the perf page once per history size; print a table."""
|
||||
import functools
|
||||
import threading
|
||||
@@ -1644,7 +1661,7 @@ def run_perf(out: Path, sizes: list[int], turns: int, timeout: float) -> bool:
|
||||
try:
|
||||
for n in sizes:
|
||||
print(f"perf: n={n} turns={turns} … ", end="", flush=True)
|
||||
report = _perf_run_one(chrome, out, port, store, n, turns, timeout)
|
||||
report = _perf_run_one(chrome, out, port, store, n, turns, timeout, extra_query)
|
||||
if report is None:
|
||||
print("FAILED (no report — timeout or chrome startup failure)")
|
||||
continue
|
||||
@@ -1721,11 +1738,20 @@ def main() -> None:
|
||||
)
|
||||
ap.add_argument("--perf-turns", type=int, default=20)
|
||||
ap.add_argument("--perf-timeout", type=float, default=420.0)
|
||||
ap.add_argument(
|
||||
"--perf-extra",
|
||||
default="",
|
||||
help="extra query params for the perf page (e.g. 'window=100000' to disable windowing)",
|
||||
)
|
||||
args = ap.parse_args()
|
||||
build(args.out)
|
||||
if args.perf:
|
||||
sizes = [int(s) for s in str(args.perf_n).split(",") if s.strip()]
|
||||
raise SystemExit(0 if run_perf(args.out, sizes, args.perf_turns, args.perf_timeout) else 1)
|
||||
raise SystemExit(
|
||||
0
|
||||
if run_perf(args.out, sizes, args.perf_turns, args.perf_timeout, args.perf_extra)
|
||||
else 1
|
||||
)
|
||||
if args.serve:
|
||||
import functools
|
||||
|
||||
|
||||
@@ -1,19 +0,0 @@
|
||||
# Entra config for the Entra e2e / spike harnesses. Copy to `.env` (gitignored)
|
||||
# and fill in from your tenant. `entra_setup.sh setup` creates the app
|
||||
# registrations and writes a populated `.env` for you.
|
||||
#
|
||||
# cp scripts/obo-e2e/.env.example scripts/obo-e2e/.env
|
||||
# # then edit, or run: ./scripts/obo-e2e/entra_setup.sh setup
|
||||
|
||||
export ENTRA_TENANT_ID=<tenant-guid-or-domain>
|
||||
export ENTRA_CLIENT_ID=<turnstone-spike-app-client-id>
|
||||
export ENTRA_CLIENT_SECRET=<client-secret>
|
||||
export SPIKE_AUDIENCE_A=api://<resource-app-a-guid> # a consented resource
|
||||
export SPIKE_AUDIENCE_B=api://<resource-app-b-guid> # a second consented resource
|
||||
export SPIKE_AUDIENCE_UNCONSENTED=api://<resource-app-c-guid> # NOT granted (negative case)
|
||||
export SPIKE_RUN_OBO=1
|
||||
# export SPIKE_PORT=8765 # redirect-listener port (default 8765)
|
||||
# export SPIKE_CALLBACK_FILE=/tmp/obo_cb.txt # remote-browser mode: paste the redirect URL here
|
||||
|
||||
# The Keycloak / OSS-path harness needs no config — keycloak_e2e.sh sets
|
||||
# everything and stands up an ephemeral container.
|
||||
@@ -1,214 +0,0 @@
|
||||
# OBO e2e harnesses — single-credential MCP token minting (`auth_type=oauth_obo`)
|
||||
|
||||
Manual test harnesses for the `oauth_obo` feature (issue #551). They exercise
|
||||
the **real** Turnstone mint path (`get_obo_access_token_classified` →
|
||||
`_obo_mint_entra` / `_obo_mint_rfc8693`) against a real identity provider — not
|
||||
mocks, not the unit suite. Two grant legs:
|
||||
|
||||
- **Entra** (`entra_e2e.py`) — real tenant, one interactive sign-in.
|
||||
- **Keycloak / RFC 8693** (`keycloak_e2e.py` + `.sh`) — ephemeral docker, fully
|
||||
headless.
|
||||
|
||||
There is also `entra_spike.py` (raw-OAuth **wire** probe, pre-implementation
|
||||
reference) and `entra_setup.sh` (creates the Entra app registrations + writes a
|
||||
populated `.env`).
|
||||
|
||||
**Secrets:** these read config from env. Real credentials live in a **gitignored
|
||||
`.env`** (copy `.env.example`); nothing tenant-specific is committed. The only
|
||||
literal secret in the tree is the ephemeral Keycloak container's throwaway
|
||||
`spike-secret`, which lives and dies with the container.
|
||||
|
||||
Not part of CI — run by hand when validating the feature against a live IdP.
|
||||
|
||||
## `entra_e2e.py` — end-to-end product exercise (post-implementation)
|
||||
|
||||
`entra_spike.py` verified the raw OAuth WIRE (before code existed). `entra_e2e.py`
|
||||
verifies the SHIPPED Turnstone code: it does a real Entra login, feeds the
|
||||
credential through the real `MCPTokenStore.upsert_oidc_credential` (the call the
|
||||
OIDC callback makes on capture), then drives the real
|
||||
`get_obo_access_token_classified` → `_obo_mint_entra` against the live Entra token
|
||||
endpoint. Checks E1–E7: real mint + aud claim, cache-hit (0 Entra calls),
|
||||
single-credential→audiences A&B, rotation write-back, force_refresh re-mint,
|
||||
unconsented-audience classification with the credential surviving, and
|
||||
flush→re-mint. Reuses the same `.env` and interactive login (SPIKE_CALLBACK_FILE
|
||||
for remote browser).
|
||||
|
||||
```bash
|
||||
source scripts/obo-e2e/.env
|
||||
uv run python scripts/obo-e2e/entra_e2e.py
|
||||
# one interactive sign-in; E1–E7 then run against the real product code. Results below.
|
||||
```
|
||||
|
||||
Results — RUN 2026-07-12 on the real tenant, ALL VERIFIED (exit 0): capture
|
||||
persisted; E1 mint A (aud=A app-id, cache row refresh_token_ct NULL); E2 cache
|
||||
hit (0 extra Entra calls); E3 mint B from the SAME credential (aud=B app-id); E4
|
||||
rotation write-back (RT rotated 2040→2091 chars, newest persisted); E5
|
||||
force_refresh re-mint (1 Entra call); E6 unconsented C → refresh_failed and the
|
||||
credential SURVIVES; E7 flush→re-mint. The real `get_obo_access_token_classified`
|
||||
→ `_obo_mint_entra` path against the live Entra token endpoint.
|
||||
|
||||
## `keycloak_e2e.py` + `keycloak_e2e.sh` — OSS path (RFC 8693), headless
|
||||
|
||||
The rfc8693 equivalent of `entra_e2e.py`: `keycloak_e2e.sh` spins up ephemeral
|
||||
Keycloak, configures the realm (turnstone client with standard token exchange,
|
||||
mcp-a/b/c clients, aud-mcp-a/b audience scopes, a test user), runs the harness
|
||||
against the real `get_obo_access_token_classified` → `_obo_mint_rfc8693`
|
||||
(refresh grant → token exchange), then tears down. No browser (password grant).
|
||||
|
||||
```bash
|
||||
./scripts/obo-e2e/keycloak_e2e.sh
|
||||
```
|
||||
|
||||
Results — RUN 2026-07-12, ALL VERIFIED: capture persisted; E1 mint A
|
||||
(refresh→exchange, aud=mcp-a, cache row refresh_token_ct NULL); E2 cache hit (0
|
||||
extra KC calls); E3 mint B from the SAME credential (aud=mcp-b); E4 rotation
|
||||
write-back (KC rotated the RT on the refresh leg, newest persisted); E5
|
||||
force_refresh re-mint (**2 KC calls** = the two-leg chain); E6 unconsented C →
|
||||
refresh_failed_transient (KC returns invalid_request for a missing audience
|
||||
scope → classified transient; credential SURVIVES either way); E7 flush→re-mint.
|
||||
Gotcha: dev-mode Keycloak boot is slow on a loaded host — the script now waits on
|
||||
kcadm auth (up to ~6 min) rather than a fixed sleep. Port 8091 (8090 = the dev
|
||||
console).
|
||||
|
||||
## Leg 1 — Entra (`entra_spike.py`) — NEEDS TENANT ACCESS
|
||||
|
||||
### Tenant / app-registration setup (one-time, ~15 min)
|
||||
|
||||
1. **Spike client app** (stands in for Turnstone's OIDC app registration):
|
||||
- New app registration, single tenant. Platform **Web**, redirect URI
|
||||
`http://localhost:8765/callback`. Create a **client secret**.
|
||||
2. **Two resource apps** (stand in for MCP servers A and B):
|
||||
- New app registrations `spike-mcp-a`, `spike-mcp-b`. In each:
|
||||
**Expose an API** → set Application ID URI (`api://<guid>`) → add a scope
|
||||
(e.g. `mcp.access`).
|
||||
3. **Delegated grants** (this is metaclassing's "proper tenant and app reg setup"):
|
||||
- On the spike client app → **API permissions** → add delegated permission to
|
||||
`spike-mcp-a` and `spike-mcp-b` scopes → **Grant admin consent**.
|
||||
- Optionally also add the spike client's app id to each resource app's
|
||||
`preAuthorizedApplications` (Expose an API → Add a client application) to
|
||||
compare against pure admin consent.
|
||||
4. **Unconsented control** (for V5): a third resource app `spike-mcp-c` with an
|
||||
exposed API but NO permission granted to the spike client.
|
||||
|
||||
### Run
|
||||
|
||||
```bash
|
||||
export ENTRA_TENANT_ID=... ENTRA_CLIENT_ID=... ENTRA_CLIENT_SECRET=...
|
||||
export SPIKE_AUDIENCE_A=api://<a-guid> SPIKE_AUDIENCE_B=api://<b-guid>
|
||||
export SPIKE_AUDIENCE_UNCONSENTED=api://<c-guid> # optional (V5)
|
||||
export SPIKE_RUN_OBO=1 # optional (V6)
|
||||
uv run python scripts/obo-e2e/entra_spike.py
|
||||
```
|
||||
|
||||
A browser opens for one interactive login (any tenant user). Everything after is
|
||||
non-interactive — that IS the feature.
|
||||
|
||||
### What each check pins down
|
||||
|
||||
| Check | Design assumption it verifies |
|
||||
| --- | --- |
|
||||
| V1 | `offline_access` on the login yields a client-bound RT (capture layer) |
|
||||
| V2/V3 | ONE RT redeems for access tokens of DIFFERENT audiences (`scope=<aud>/.default`) — the load-bearing Entra behavior |
|
||||
| V4 | rotation semantics → whether RT write-back on every mint is convenience or correctness-critical |
|
||||
| V5 | unconsented audience fails `AADSTS65001 consent_required` → maps to the reconnect-rail fallback, never a silent failure |
|
||||
| V6 | OBO jwt-bearer middle-tier variant works with the same app registration (comparison data only) |
|
||||
|
||||
Also record (manual): whether Conditional Access / MFA policies in the tenant
|
||||
produce `interaction_required` on redemption — that's the fallback path's other
|
||||
trigger.
|
||||
|
||||
### Results — RUN 2026-07-11 on a real tenant, ALL SIX VERIFIED
|
||||
|
||||
Tenant: personal default directory (Global Admin), user is an MSA member.
|
||||
Setup via `entra_setup.sh setup`; V3 initially failed (see gotcha below),
|
||||
passed after fixing the grant. Second run: V1-V6 all VERIFIED, exit 0.
|
||||
|
||||
| Check | Result |
|
||||
| --- | --- |
|
||||
| V1 offline_access login -> RT | VERIFIED (confidential client + PKCE, RT ~2KB) |
|
||||
| V2 RT -> audience A token | VERIFIED (`aud=<A app guid>`, ~70 min TTL, new RT returned) |
|
||||
| V3 SAME RT -> audience B token | **VERIFIED — the load-bearing claim: one RT, many audiences** |
|
||||
| V4 rotation | VERIFIED: RT rotates on every redemption, but the OLD RT stays valid (reuse HTTP 200) -> write-back-newest is required; races are benign on Entra |
|
||||
| V5 unconsented audience | VERIFIED: `invalid_grant` + `AADSTS65001` (error_codes=[65001]) -> clean mapping to the reconnect-rail fallback |
|
||||
| V6 OBO jwt-bearer variant | VERIFIED: middle-tier shape also works with the same app registration |
|
||||
|
||||
**Operator gotcha (feeds #682 + product docs):** `az ad app permission
|
||||
admin-consent` run immediately after SP creation SILENTLY skips
|
||||
not-yet-propagated resource SPs — grant A landed, grant B didn't, and the only
|
||||
symptom was AADSTS65001 at redemption. Verify grants after consent
|
||||
(`oauth2PermissionGrants` filter on the client SP) or write them directly with
|
||||
`az ad app permission grant --id <client> --api <resource> --scope <scope>`.
|
||||
Product-side implication: a missing tenant grant for a NEW oauth_obo server
|
||||
surfaces as AADSTS65001 -> the same reconnect-rail path as revocation; the
|
||||
admin docs must say "grant first, then add the server".
|
||||
|
||||
## Leg 2 — Keycloak RFC 8693 (portability check) — runnable locally
|
||||
|
||||
Ephemeral `quay.io/keycloak/keycloak:26.3` (`start-dev`, port 8089), realm
|
||||
`spike`, confidential client `turnstone` with **standard token exchange**
|
||||
enabled, resource clients `mcp-a`/`mcp-b`, user `alice`. Pipeline mirrors the
|
||||
product design for a generic-8693 IdP:
|
||||
|
||||
```
|
||||
stored user RT --(refresh grant)--> user AT --(RFC 8693 exchange, audience=mcp-X)--> audience-scoped AT
|
||||
```
|
||||
|
||||
i.e. the per-user credential stays ONE refresh token; per-server tokens are
|
||||
minted via standard token exchange instead of Entra's multi-resource RT
|
||||
redemption. Same substrate, different grant leg.
|
||||
|
||||
### Results — RUN 2026-07-11, VERIFIED (Keycloak 26.3, ephemeral)
|
||||
|
||||
```
|
||||
alice ONE stored RT
|
||||
-> refresh grant -> user AT (azp=turnstone); RT ROTATED on refresh
|
||||
-> 8693 exchange audience=mcp-a scope=aud-mcp-a -> AT aud=mcp-a user=alice 300s, NO RT
|
||||
-> 8693 exchange audience=mcp-b scope=aud-mcp-b -> AT aud=mcp-b (same subject AT)
|
||||
negative control audience=mcp-c -> invalid_client "Audience not found"
|
||||
```
|
||||
|
||||
Findings that feed the design:
|
||||
1. **One per-user credential -> N audience tokens: VERIFIED on a second IdP.**
|
||||
The substrate is portable; only the grant leg differs per IdP.
|
||||
2. **Exchanged tokens are cache-shaped** (short TTL, no RT) — per-server
|
||||
`mcp_user_tokens` rows as short-lived mint cache is the right model.
|
||||
3. **RT rotation happens here too** — newest-RT write-back on every redemption
|
||||
is a correctness requirement of the capture layer, not an Entra quirk.
|
||||
4. **The IdP-side "delegated grant" has a per-IdP shape**: Entra = API
|
||||
permissions + admin consent; Keycloak = audience client scopes attached to
|
||||
the requester client (optional scopes activate via `scope=` at exchange).
|
||||
Operator runbooks are per-IdP (#682 pattern), code is not.
|
||||
5. Gotchas hit: KC user needs a complete profile for direct grant ("Account is
|
||||
not fully set up"); optional audience scope must be requested explicitly or
|
||||
the exchange 400s with "Requested audience not available".
|
||||
|
||||
Repro (ephemeral, ~2 min):
|
||||
|
||||
```bash
|
||||
docker run -d --name kc-obo-spike -p 127.0.0.1:8089:8080 \
|
||||
-e KC_BOOTSTRAP_ADMIN_USERNAME=admin -e KC_BOOTSTRAP_ADMIN_PASSWORD=admin \
|
||||
quay.io/keycloak/keycloak:26.3 start-dev
|
||||
KC="docker exec kc-obo-spike /opt/keycloak/bin/kcadm.sh"
|
||||
$KC config credentials --server http://localhost:8080 --realm master --user admin --password admin
|
||||
$KC create realms -s realm=spike -s enabled=true
|
||||
$KC create clients -r spike -s clientId=turnstone -s enabled=true -s publicClient=false \
|
||||
-s secret=spike-secret -s directAccessGrantsEnabled=true \
|
||||
-s 'attributes={"standard.token.exchange.enabled":"true"}'
|
||||
$KC create clients -r spike -s clientId=mcp-a -s enabled=true -s publicClient=false -s secret=x
|
||||
$KC create clients -r spike -s clientId=mcp-b -s enabled=true -s publicClient=false -s secret=x
|
||||
$KC create users -r spike -s username=alice -s enabled=true -s email=a@s.test \
|
||||
-s emailVerified=true -s firstName=A -s lastName=S
|
||||
$KC set-password -r spike --username alice --new-password alice-pw
|
||||
TURNSTONE_UUID=$($KC get clients -r spike -q clientId=turnstone --fields id --format csv --noquotes)
|
||||
for t in mcp-a mcp-b; do
|
||||
SID=$($KC create client-scopes -r spike -s name=aud-$t -s protocol=openid-connect -i)
|
||||
$KC create client-scopes/$SID/protocol-mappers/models -r spike -s name=aud-$t \
|
||||
-s protocol=openid-connect -s protocolMapper=oidc-audience-mapper \
|
||||
-s "config={\"included.client.audience\":\"$t\",\"access.token.claim\":\"true\"}"
|
||||
$KC update clients/$TURNSTONE_UUID/optional-client-scopes/$SID -r spike
|
||||
done
|
||||
# then: password grant -> refresh grant -> token-exchange with
|
||||
# grant_type=urn:ietf:params:oauth:grant-type:token-exchange,
|
||||
# subject_token=<user AT>, subject_token_type=...:access_token,
|
||||
# audience=mcp-a, scope=aud-mcp-a
|
||||
```
|
||||
@@ -1,286 +0,0 @@
|
||||
"""End-to-end exercise of the oauth_obo feature against a REAL Entra tenant.
|
||||
|
||||
Unlike ``entra_spike.py`` (which verified the raw OAuth wire shapes), this
|
||||
drives the ACTUAL Turnstone product code — real ``MCPTokenStore``, real
|
||||
``get_obo_access_token_classified`` → ``_obo_mint_entra`` → the real Entra
|
||||
token endpoint — so a green run proves the shipped mint engine works against
|
||||
live Entra, not just that the protocol does.
|
||||
|
||||
Flow:
|
||||
1. Interactive Entra login (auth-code + PKCE + offline_access) → a real
|
||||
refresh credential. This is what ``handle_oidc_callback`` receives.
|
||||
2. Persist it via ``MCPTokenStore.upsert_oidc_credential`` — the exact call
|
||||
the OIDC callback makes on capture (auth.py). The rest of the callback
|
||||
(JWKS validation, user provisioning) is OIDC-generic and unit-tested; the
|
||||
novel path is capture + mint, which this exercises for real.
|
||||
3. Seed real ``oauth_obo`` ``mcp_servers`` rows (audiences A/B consented, C
|
||||
not) and drive ``get_obo_access_token_classified`` — the real dispatch-time
|
||||
entry point — asserting on the minted tokens, cache, rotation, and
|
||||
classification.
|
||||
|
||||
Checks (VERIFIED / FAILED per line):
|
||||
E1 mint for audience A → kind=token; decoded aud == A; cache row written with
|
||||
refresh_token_ct NULL (cache, not custody); expires_at set
|
||||
E2 second call for A → cache hit, ZERO additional Entra calls
|
||||
E3 mint for audience B from the SAME captured credential → aud == B
|
||||
(the single-credential-many-audiences thesis, through the real engine)
|
||||
E4 rotation write-back: the stored credential holds the newest refresh token
|
||||
E5 force_refresh → a fresh mint (Entra call count increments)
|
||||
E6 unconsented audience C → NOT kind=token, and the shared credential SURVIVES
|
||||
(never auto-deleted — the load-bearing custody invariant)
|
||||
E7 cache flush → re-mint: deleting the cache row makes the next call re-mint
|
||||
|
||||
Run:
|
||||
source scripts/obo-e2e/.env
|
||||
uv run python scripts/obo-e2e/entra_e2e.py
|
||||
Env (from .env): ENTRA_TENANT_ID, ENTRA_CLIENT_ID, ENTRA_CLIENT_SECRET,
|
||||
SPIKE_AUDIENCE_A, SPIKE_AUDIENCE_B, SPIKE_AUDIENCE_UNCONSENTED, SPIKE_PORT.
|
||||
Remote browser: set SPIKE_CALLBACK_FILE to paste the redirect URL (as before).
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import base64
|
||||
import os
|
||||
import sys
|
||||
import tempfile
|
||||
from types import SimpleNamespace
|
||||
from typing import Any
|
||||
|
||||
import httpx
|
||||
|
||||
# Reuse the verified interactive-login machinery from the wire spike.
|
||||
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
|
||||
from entra_spike import interactive_login, jwt_claims_unverified, redact # noqa: E402
|
||||
|
||||
from turnstone.core.mcp_crypto import ( # noqa: E402
|
||||
MCPTokenCipher,
|
||||
MCPTokenCipherConfig,
|
||||
MCPTokenStore,
|
||||
)
|
||||
from turnstone.core.mcp_oauth import get_obo_access_token_classified # noqa: E402
|
||||
from turnstone.core.oidc import OIDCConfig # noqa: E402
|
||||
from turnstone.core.storage._sqlite import SQLiteBackend # noqa: E402
|
||||
|
||||
USER = "e2e-user"
|
||||
RESULTS: list[tuple[str, str]] = []
|
||||
|
||||
|
||||
def record(status: str, msg: str) -> None:
|
||||
RESULTS.append((status, msg))
|
||||
print(f"[{status:>8}] {msg}")
|
||||
|
||||
|
||||
def aud_matches(token: str, want_audience: str) -> tuple[bool, str]:
|
||||
"""Compare a minted access token's aud claim to the configured audience.
|
||||
|
||||
Entra returns aud as the bare app-id GUID or the full ``api://<guid>`` URI;
|
||||
accept either.
|
||||
"""
|
||||
claims = jwt_claims_unverified(token)
|
||||
aud = str(claims.get("aud", "<none>"))
|
||||
want = want_audience.removeprefix("api://")
|
||||
return aud in (want, want_audience), aud
|
||||
|
||||
|
||||
class _CountingClient:
|
||||
"""Wraps httpx.AsyncClient, counting token-endpoint POSTs so cache hits
|
||||
(which must issue zero) are observable."""
|
||||
|
||||
def __init__(self, inner: httpx.AsyncClient) -> None:
|
||||
self._inner = inner
|
||||
self.posts = 0
|
||||
|
||||
async def post(self, *args: Any, **kwargs: Any) -> httpx.Response:
|
||||
self.posts += 1
|
||||
return await self._inner.post(*args, **kwargs)
|
||||
|
||||
|
||||
def _make_app_state(
|
||||
storage: SQLiteBackend,
|
||||
store: MCPTokenStore,
|
||||
oidc_config: OIDCConfig,
|
||||
http_client: _CountingClient,
|
||||
) -> SimpleNamespace:
|
||||
return SimpleNamespace(
|
||||
auth_storage=storage,
|
||||
mcp_token_store=store,
|
||||
oidc_config=oidc_config,
|
||||
obo_http_client=http_client,
|
||||
mcp_oauth_refresh_locks={},
|
||||
mcp_oauth_refresh_backoff={},
|
||||
)
|
||||
|
||||
|
||||
def _seed_obo_server(storage: SQLiteBackend, name: str, audience: str) -> None:
|
||||
storage.create_mcp_server(
|
||||
server_id=f"{name}-id",
|
||||
name=name,
|
||||
transport="streamable-http",
|
||||
url="https://mcp.example.invalid/sse",
|
||||
auth_type="oauth_obo",
|
||||
oauth_audience=audience,
|
||||
)
|
||||
|
||||
|
||||
async def _run(cfg: dict[str, str], refresh_token: str) -> None:
|
||||
tenant = cfg["ENTRA_TENANT_ID"]
|
||||
issuer = f"https://login.microsoftonline.com/{tenant}/v2.0"
|
||||
token_endpoint = f"https://login.microsoftonline.com/{tenant}/oauth2/v2.0/token"
|
||||
aud_a = cfg["SPIKE_AUDIENCE_A"]
|
||||
aud_b = cfg["SPIKE_AUDIENCE_B"]
|
||||
aud_c = cfg.get("SPIKE_AUDIENCE_UNCONSENTED", "")
|
||||
|
||||
# Real Turnstone objects.
|
||||
db_path = os.path.join(tempfile.mkdtemp(prefix="obo-e2e-"), "e2e.db")
|
||||
storage = SQLiteBackend(db_path)
|
||||
from cryptography.fernet import Fernet
|
||||
|
||||
raw = base64.urlsafe_b64decode(Fernet.generate_key())
|
||||
store = MCPTokenStore(storage, MCPTokenCipher(MCPTokenCipherConfig(keys=(raw,))), node_id="e2e")
|
||||
oidc_config = OIDCConfig(
|
||||
enabled=True,
|
||||
issuer=issuer,
|
||||
client_id=cfg["ENTRA_CLIENT_ID"],
|
||||
client_secret=cfg["ENTRA_CLIENT_SECRET"],
|
||||
token_endpoint=token_endpoint,
|
||||
obo_grant_profile="entra",
|
||||
capture_user_credential=True,
|
||||
)
|
||||
|
||||
# Step 2 — CAPTURE: the exact storage call handle_oidc_callback makes.
|
||||
store.upsert_oidc_credential(USER, issuer, refresh_token=refresh_token)
|
||||
cap = store.get_oidc_credential(USER, issuer)
|
||||
if cap and cap["refresh_token"] == refresh_token:
|
||||
record("VERIFIED", f"capture: credential persisted for {USER} ({redact(refresh_token)})")
|
||||
else:
|
||||
record("FAILED", "capture: credential did not round-trip")
|
||||
return
|
||||
|
||||
_seed_obo_server(storage, "e2e-a", aud_a)
|
||||
_seed_obo_server(storage, "e2e-b", aud_b)
|
||||
if aud_c:
|
||||
_seed_obo_server(storage, "e2e-c", aud_c)
|
||||
|
||||
inner = httpx.AsyncClient(timeout=20.0)
|
||||
client = _CountingClient(inner)
|
||||
app_state = _make_app_state(storage, store, oidc_config, client)
|
||||
try:
|
||||
# E1 — real mint for audience A.
|
||||
r = await get_obo_access_token_classified(
|
||||
app_state=app_state, user_id=USER, server_name="e2e-a"
|
||||
)
|
||||
if r.kind == "token" and r.token:
|
||||
ok, aud = aud_matches(r.token, aud_a)
|
||||
row = storage.get_mcp_user_token(USER, "e2e-a")
|
||||
cache_ok = (
|
||||
row is not None and row["refresh_token_ct"] is None and bool(row["expires_at"])
|
||||
)
|
||||
record(
|
||||
"VERIFIED" if ok and cache_ok else "FAILED",
|
||||
f"E1 mint A: kind=token aud={aud} want={aud_a} cache_row_refreshless={cache_ok}",
|
||||
)
|
||||
else:
|
||||
record("FAILED", f"E1 mint A: kind={r.kind} (expected token)")
|
||||
return
|
||||
|
||||
# E2 — cache hit issues zero Entra calls.
|
||||
posts_before = client.posts
|
||||
r2 = await get_obo_access_token_classified(
|
||||
app_state=app_state, user_id=USER, server_name="e2e-a"
|
||||
)
|
||||
record(
|
||||
"VERIFIED" if r2.kind == "token" and client.posts == posts_before else "FAILED",
|
||||
f"E2 cache hit: kind={r2.kind} extra_entra_calls={client.posts - posts_before} (want 0)",
|
||||
)
|
||||
|
||||
# E3 — same credential, audience B.
|
||||
rb = await get_obo_access_token_classified(
|
||||
app_state=app_state, user_id=USER, server_name="e2e-b"
|
||||
)
|
||||
if rb.kind == "token" and rb.token:
|
||||
ok_b, aud_bclaim = aud_matches(rb.token, aud_b)
|
||||
record(
|
||||
"VERIFIED" if ok_b else "FAILED",
|
||||
f"E3 mint B from SAME credential: aud={aud_bclaim} want={aud_b}",
|
||||
)
|
||||
else:
|
||||
record("FAILED", f"E3 mint B: kind={rb.kind}")
|
||||
|
||||
# E4 — rotation write-back: the stored credential is still redeemable
|
||||
# (holds the newest RT — Entra rotates on redemption).
|
||||
cred_now = store.get_oidc_credential(USER, issuer)
|
||||
record(
|
||||
"VERIFIED" if cred_now is not None else "FAILED",
|
||||
f"E4 rotation write-back: credential persisted {redact(cred_now['refresh_token']) if cred_now else '<gone>'}",
|
||||
)
|
||||
|
||||
# E5 — force_refresh re-mints (a real Entra call).
|
||||
posts_before = client.posts
|
||||
rf = await get_obo_access_token_classified(
|
||||
app_state=app_state, user_id=USER, server_name="e2e-a", force_refresh=True
|
||||
)
|
||||
record(
|
||||
"VERIFIED" if rf.kind == "token" and client.posts > posts_before else "FAILED",
|
||||
f"E5 force_refresh re-mint: kind={rf.kind} entra_calls={client.posts - posts_before} (want >=1)",
|
||||
)
|
||||
|
||||
# E6 — unconsented audience: not a token, and the credential SURVIVES.
|
||||
if aud_c:
|
||||
rc = await get_obo_access_token_classified(
|
||||
app_state=app_state, user_id=USER, server_name="e2e-c"
|
||||
)
|
||||
cred_after = store.get_oidc_credential(USER, issuer)
|
||||
record(
|
||||
"VERIFIED" if rc.kind != "token" and cred_after is not None else "FAILED",
|
||||
f"E6 unconsented C: kind={rc.kind} (not token) credential_survives={cred_after is not None}",
|
||||
)
|
||||
else:
|
||||
record("SKIPPED", "E6 unconsented C: SPIKE_AUDIENCE_UNCONSENTED not set")
|
||||
|
||||
# E7 — cache flush → re-mint.
|
||||
store.delete_user_token(USER, "e2e-a")
|
||||
posts_before = client.posts
|
||||
r7 = await get_obo_access_token_classified(
|
||||
app_state=app_state, user_id=USER, server_name="e2e-a"
|
||||
)
|
||||
record(
|
||||
"VERIFIED" if r7.kind == "token" and client.posts > posts_before else "FAILED",
|
||||
f"E7 flush→re-mint: kind={r7.kind} entra_calls={client.posts - posts_before} (want >=1)",
|
||||
)
|
||||
finally:
|
||||
await inner.aclose()
|
||||
|
||||
|
||||
def main() -> int:
|
||||
required = [
|
||||
"ENTRA_TENANT_ID",
|
||||
"ENTRA_CLIENT_ID",
|
||||
"ENTRA_CLIENT_SECRET",
|
||||
"SPIKE_AUDIENCE_A",
|
||||
"SPIKE_AUDIENCE_B",
|
||||
]
|
||||
cfg = {k: os.environ[k] for k in os.environ if k.startswith(("ENTRA_", "SPIKE_"))}
|
||||
missing = [k for k in required if not cfg.get(k)]
|
||||
if missing:
|
||||
print(f"Missing env: {', '.join(missing)} — did you `source scripts/obo-e2e/.env`?")
|
||||
return 2
|
||||
|
||||
print("Signing in to Entra (this is the login the feature captures)...")
|
||||
tokens = interactive_login(cfg)
|
||||
refresh_token = tokens.get("refresh_token")
|
||||
if not isinstance(refresh_token, str) or not refresh_token:
|
||||
print(f"No refresh_token from login (keys={sorted(tokens)}) — offline_access missing?")
|
||||
return 1
|
||||
|
||||
asyncio.run(_run(cfg, refresh_token))
|
||||
|
||||
print("\n=== summary ===")
|
||||
for status, msg in RESULTS:
|
||||
print(f" {status:>8} {msg}")
|
||||
return 0 if all(s in ("VERIFIED", "SKIPPED") for s, _ in RESULTS) else 1
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
sys.exit(main())
|
||||
@@ -1,134 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
# Entra spike setup for entra_spike.py (#551 re-scope boundary spike).
|
||||
# Manual test tooling — not run in CI. Creates throwaway Entra app registrations.
|
||||
#
|
||||
# ./entra_setup.sh setup create app registrations + consent + .env
|
||||
# ./entra_setup.sh cleanup delete everything it created (incl. .env)
|
||||
#
|
||||
# Creates in the logged-in tenant (az login first):
|
||||
# spike-turnstone confidential client (stands in for Turnstone's OIDC app)
|
||||
# spike-mcp-a/b resource apps exposing scope mcp.access, admin-consented
|
||||
# spike-mcp-c resource app with NO grant to the client (V5 control)
|
||||
# Requires: the logged-in user can create apps + grant admin consent
|
||||
# (Global Admin on a personal tenant qualifies).
|
||||
|
||||
set -euo pipefail
|
||||
cd "$(dirname "$0")"
|
||||
ENV_FILE=".env"
|
||||
NAMES=(spike-turnstone spike-mcp-a spike-mcp-b spike-mcp-c)
|
||||
|
||||
log() { printf '>> %s\n' "$*"; }
|
||||
|
||||
graph_patch_api() { # $1=appId $2=scope-uuid $3=display-name
|
||||
local obj_id
|
||||
obj_id=$(az ad app show --id "$1" --query id -o tsv)
|
||||
az rest --method PATCH \
|
||||
--url "https://graph.microsoft.com/v1.0/applications/${obj_id}" \
|
||||
--headers 'Content-Type=application/json' \
|
||||
--body "{
|
||||
\"identifierUris\": [\"api://$1\"],
|
||||
\"api\": {
|
||||
\"requestedAccessTokenVersion\": 2,
|
||||
\"oauth2PermissionScopes\": [{
|
||||
\"id\": \"$2\",
|
||||
\"value\": \"mcp.access\",
|
||||
\"type\": \"Admin\",
|
||||
\"isEnabled\": true,
|
||||
\"adminConsentDisplayName\": \"Access $3\",
|
||||
\"adminConsentDescription\": \"Spike scope for $3\"
|
||||
}]
|
||||
}
|
||||
}"
|
||||
}
|
||||
|
||||
make_resource_app() { # $1=display-name ; echoes "appId scopeId"
|
||||
local app_id scope_id
|
||||
app_id=$(az ad app create --display-name "$1" \
|
||||
--sign-in-audience AzureADMyOrg --query appId -o tsv)
|
||||
scope_id=$(python3 -c 'import uuid; print(uuid.uuid4())')
|
||||
graph_patch_api "$app_id" "$scope_id" "$1" >/dev/null
|
||||
az ad sp create --id "$app_id" >/dev/null 2>&1 || true
|
||||
echo "$app_id $scope_id"
|
||||
}
|
||||
|
||||
cmd_setup() {
|
||||
local tenant_id
|
||||
tenant_id=$(az account show --query tenantId -o tsv)
|
||||
log "tenant: ${tenant_id}"
|
||||
|
||||
log "creating resource apps (a, b, c)..."
|
||||
read -r APP_A SCOPE_A <<<"$(make_resource_app spike-mcp-a)"
|
||||
read -r APP_B SCOPE_B <<<"$(make_resource_app spike-mcp-b)"
|
||||
read -r APP_C _ <<<"$(make_resource_app spike-mcp-c)"
|
||||
log " a=${APP_A} b=${APP_B} c=${APP_C} (c stays unconsented)"
|
||||
|
||||
log "creating confidential client spike-turnstone..."
|
||||
CLIENT_ID=$(az ad app create --display-name spike-turnstone \
|
||||
--sign-in-audience AzureADMyOrg \
|
||||
--web-redirect-uris "http://localhost:8765/callback" \
|
||||
--query appId -o tsv)
|
||||
az ad sp create --id "$CLIENT_ID" >/dev/null 2>&1 || true
|
||||
# No stderr suppression here: the secret is load-bearing (it lands in .env),
|
||||
# so under `set -e` a reset failure must abort LOUDLY, not silently.
|
||||
SECRET=$(az ad app credential reset --id "$CLIENT_ID" \
|
||||
--display-name spike --years 1 --query password -o tsv)
|
||||
|
||||
log "adding delegated permissions (a, b — NOT c)..."
|
||||
# Tolerated failures (|| log): a re-run hits "permission already exists" and
|
||||
# SP-propagation delays are common right after app creation — the
|
||||
# admin-consent retry loop below is the real gate. `set -e` would otherwise
|
||||
# turn a suppressed non-zero here into a silent mid-script abort.
|
||||
az ad app permission add --id "$CLIENT_ID" \
|
||||
--api "$APP_A" --api-permissions "${SCOPE_A}=Scope" \
|
||||
|| log " warn: permission add for a failed (may already exist); admin-consent below will confirm"
|
||||
az ad app permission add --id "$CLIENT_ID" \
|
||||
--api "$APP_B" --api-permissions "${SCOPE_B}=Scope" \
|
||||
|| log " warn: permission add for b failed (may already exist); admin-consent below will confirm"
|
||||
|
||||
log "granting admin consent (retries while SPs propagate)..."
|
||||
local ok=""
|
||||
for i in 1 2 3 4 5; do
|
||||
if az ad app permission admin-consent --id "$CLIENT_ID" 2>/dev/null; then
|
||||
ok=1; break
|
||||
fi
|
||||
log " not yet (attempt $i) — waiting 15s"
|
||||
sleep 15
|
||||
done
|
||||
[ -n "$ok" ] || { log "admin-consent failed after retries — grant manually in the portal (API permissions blade) and re-run the spike"; }
|
||||
|
||||
# Single-quote the values in the generated .env: the AS-issued client secret
|
||||
# can contain $ / backtick, and an unquoted RHS would be re-expanded (or
|
||||
# partially executed) when the operator `source`s the file. The heredoc still
|
||||
# interpolates ${...} into the single-quoted output; sourcing then treats the
|
||||
# result literally. (Azure secrets are base64-ish — no single quotes to escape.)
|
||||
umask 177
|
||||
cat > "$ENV_FILE" <<EOF
|
||||
export ENTRA_TENANT_ID='${tenant_id}'
|
||||
export ENTRA_CLIENT_ID='${CLIENT_ID}'
|
||||
export ENTRA_CLIENT_SECRET='${SECRET}'
|
||||
export SPIKE_AUDIENCE_A='api://${APP_A}'
|
||||
export SPIKE_AUDIENCE_B='api://${APP_B}'
|
||||
export SPIKE_AUDIENCE_UNCONSENTED='api://${APP_C}'
|
||||
export SPIKE_RUN_OBO=1
|
||||
EOF
|
||||
log "wrote ${ENV_FILE} (chmod 600). Next:"
|
||||
log " source scripts/obo-e2e/.env && uv run python scripts/obo-e2e/entra_spike.py"
|
||||
log "cleanup later with: ./entra_setup.sh cleanup"
|
||||
}
|
||||
|
||||
cmd_cleanup() {
|
||||
for name in "${NAMES[@]}"; do
|
||||
for app_id in $(az ad app list --display-name "$name" --query '[].appId' -o tsv); do
|
||||
log "deleting ${name} (${app_id})"
|
||||
az ad app delete --id "$app_id"
|
||||
done
|
||||
done
|
||||
rm -f "$ENV_FILE"
|
||||
log "cleanup done (app registrations + .env removed)"
|
||||
}
|
||||
|
||||
case "${1:-}" in
|
||||
setup) cmd_setup ;;
|
||||
cleanup) cmd_cleanup ;;
|
||||
*) echo "usage: $0 setup|cleanup"; exit 2 ;;
|
||||
esac
|
||||
@@ -1,333 +0,0 @@
|
||||
"""Entra boundary spike for single-credential MCP token minting (#551 re-scope).
|
||||
|
||||
Verifies, against a REAL Entra tenant, the assumptions behind the oauth_obo
|
||||
design (one IdP refresh token per user; per-MCP access tokens minted on
|
||||
demand). Each check prints VERIFIED / FAILED / SKIPPED plus redacted evidence.
|
||||
|
||||
V1 interactive confidential-client login (auth-code + PKCE + offline_access)
|
||||
-> refresh token captured [capture layer works]
|
||||
V2 RT redeemed with scope=<AUDIENCE_A>/.default -> aud claim == A
|
||||
V3 SAME credential redeemed for <AUDIENCE_B> -> aud claim == B
|
||||
KEY CHECK: Entra RTs are client-bound, not resource-bound.
|
||||
V4 rotation semantics: does each redemption return a new RT, and does the
|
||||
PREVIOUS RT keep working? [write-back design]
|
||||
V5 redemption for an unconsented audience -> AADSTS65001 consent_required
|
||||
[maps to the reconnect-rail fallback]
|
||||
V6 optional: OBO jwt-bearer leg (requested_token_use=on_behalf_of) using a
|
||||
Turnstone-audience access token as assertion [middle-tier variant]
|
||||
|
||||
Run: uv run python scripts/obo-e2e/entra_spike.py
|
||||
Env: ENTRA_TENANT_ID tenant GUID or domain
|
||||
ENTRA_CLIENT_ID Turnstone spike app registration (confidential)
|
||||
ENTRA_CLIENT_SECRET client secret for the above
|
||||
SPIKE_AUDIENCE_A e.g. api://<guid-a> (exposes a scope, consented)
|
||||
SPIKE_AUDIENCE_B e.g. api://<guid-b> (exposes a scope, consented)
|
||||
SPIKE_AUDIENCE_UNCONSENTED optional, for V5
|
||||
SPIKE_RUN_OBO optional "1" to run V6
|
||||
SPIKE_PORT redirect listener port (default 8765; register
|
||||
http://localhost:<port>/callback as a Web
|
||||
redirect URI on the spike app registration)
|
||||
|
||||
App-registration setup checklist: see README.md next to this file.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import base64
|
||||
import hashlib
|
||||
import json
|
||||
import os
|
||||
import secrets
|
||||
import sys
|
||||
import threading
|
||||
import urllib.parse
|
||||
import webbrowser
|
||||
from http.server import BaseHTTPRequestHandler, HTTPServer
|
||||
from typing import Any
|
||||
|
||||
import httpx
|
||||
|
||||
RESULTS: list[tuple[str, str, str]] = [] # (check, status, evidence)
|
||||
|
||||
|
||||
def record(check: str, status: str, evidence: str) -> None:
|
||||
RESULTS.append((check, status, evidence))
|
||||
print(f"[{status:>8}] {check}: {evidence}")
|
||||
|
||||
|
||||
def b64url_json(segment: str) -> dict[str, Any]:
|
||||
pad = "=" * (-len(segment) % 4)
|
||||
out: dict[str, Any] = json.loads(base64.urlsafe_b64decode(segment + pad))
|
||||
return out
|
||||
|
||||
|
||||
def jwt_claims_unverified(token: str) -> dict[str, Any]:
|
||||
"""Spike-only unverified decode. NEVER do this in product code."""
|
||||
try:
|
||||
return b64url_json(token.split(".")[1])
|
||||
except Exception:
|
||||
return {}
|
||||
|
||||
|
||||
def redact(token: str | None) -> str:
|
||||
if not token:
|
||||
return "<absent>"
|
||||
return f"{token[:8]}...({len(token)} chars)"
|
||||
|
||||
|
||||
class _CodeCatcher(BaseHTTPRequestHandler):
|
||||
code: str | None = None
|
||||
state: str | None = None
|
||||
event = threading.Event()
|
||||
|
||||
def do_GET(self) -> None: # noqa: N802 - stdlib API name
|
||||
q = urllib.parse.parse_qs(urllib.parse.urlparse(self.path).query)
|
||||
_CodeCatcher.code = (q.get("code") or [None])[0]
|
||||
_CodeCatcher.state = (q.get("state") or [None])[0]
|
||||
body = b"Spike login captured - return to the terminal."
|
||||
if q.get("error"):
|
||||
body = f"IdP error: {q}".encode()
|
||||
self.send_response(200)
|
||||
self.send_header("Content-Type", "text/plain")
|
||||
self.end_headers()
|
||||
self.wfile.write(body)
|
||||
_CodeCatcher.event.set()
|
||||
|
||||
def log_message(self, *args: Any) -> None:
|
||||
pass
|
||||
|
||||
|
||||
def interactive_login(cfg: dict[str, str]) -> dict[str, Any]:
|
||||
"""V1: authorization-code + PKCE + offline_access as a confidential client.
|
||||
|
||||
Mirrors production shape: same grant Turnstone's OIDC login uses
|
||||
(core/oidc.py exchange_code), plus offline_access.
|
||||
"""
|
||||
port = int(cfg.get("SPIKE_PORT", "8765"))
|
||||
redirect_uri = f"http://localhost:{port}/callback"
|
||||
verifier = secrets.token_urlsafe(48)
|
||||
challenge = (
|
||||
base64.urlsafe_b64encode(hashlib.sha256(verifier.encode()).digest()).rstrip(b"=").decode()
|
||||
)
|
||||
state = secrets.token_urlsafe(16)
|
||||
authorize = (
|
||||
f"https://login.microsoftonline.com/{cfg['ENTRA_TENANT_ID']}/oauth2/v2.0/authorize?"
|
||||
+ urllib.parse.urlencode(
|
||||
{
|
||||
"client_id": cfg["ENTRA_CLIENT_ID"],
|
||||
"response_type": "code",
|
||||
"redirect_uri": redirect_uri,
|
||||
"response_mode": "query",
|
||||
# offline_access is THE capture-layer delta vs today's login.
|
||||
# No resource scope here: the RT is minted client-bound.
|
||||
"scope": "openid profile offline_access",
|
||||
"state": state,
|
||||
"code_challenge": challenge,
|
||||
"code_challenge_method": "S256",
|
||||
}
|
||||
)
|
||||
)
|
||||
server = HTTPServer(("127.0.0.1", port), _CodeCatcher)
|
||||
threading.Thread(target=server.serve_forever, daemon=True).start()
|
||||
print(f"\nOpen (or auto-opened) in a browser with a tenant user:\n {authorize}\n")
|
||||
cb_file = cfg.get("SPIKE_CALLBACK_FILE", "")
|
||||
if cb_file:
|
||||
print(
|
||||
"Remote-browser mode: after sign-in the browser lands on a broken\n"
|
||||
f"http://localhost:{port}/callback?... page. Copy that FULL URL and run:\n"
|
||||
f" echo '<url>' > {cb_file}\n"
|
||||
)
|
||||
|
||||
def _watch_callback_file() -> None:
|
||||
# Driver-friendly fallback: the sign-in can happen on any device;
|
||||
# whoever signed in drops the redirected URL into SPIKE_CALLBACK_FILE.
|
||||
import time as _time
|
||||
|
||||
while not _CodeCatcher.event.is_set():
|
||||
try:
|
||||
with open(cb_file) as _f:
|
||||
pasted = _f.read().strip()
|
||||
except OSError:
|
||||
pasted = ""
|
||||
if "?" in pasted:
|
||||
q = urllib.parse.parse_qs(urllib.parse.urlparse(pasted).query)
|
||||
_CodeCatcher.code = (q.get("code") or [None])[0]
|
||||
_CodeCatcher.state = (q.get("state") or [None])[0]
|
||||
_CodeCatcher.event.set()
|
||||
return
|
||||
_time.sleep(1.0)
|
||||
|
||||
if cb_file:
|
||||
threading.Thread(target=_watch_callback_file, daemon=True).start()
|
||||
webbrowser.open(authorize)
|
||||
if not _CodeCatcher.event.wait(timeout=600):
|
||||
server.shutdown()
|
||||
raise SystemExit("Timed out waiting for the redirect (10 min).")
|
||||
server.shutdown()
|
||||
if _CodeCatcher.state != state:
|
||||
raise SystemExit("state mismatch on redirect - aborting.")
|
||||
if not _CodeCatcher.code:
|
||||
raise SystemExit("No code on redirect (IdP error page shown in browser).")
|
||||
resp = httpx.post(
|
||||
f"https://login.microsoftonline.com/{cfg['ENTRA_TENANT_ID']}/oauth2/v2.0/token",
|
||||
data={
|
||||
"grant_type": "authorization_code",
|
||||
"code": _CodeCatcher.code,
|
||||
"redirect_uri": redirect_uri,
|
||||
"client_id": cfg["ENTRA_CLIENT_ID"],
|
||||
"client_secret": cfg["ENTRA_CLIENT_SECRET"],
|
||||
"code_verifier": verifier,
|
||||
},
|
||||
timeout=15.0,
|
||||
)
|
||||
tokens: dict[str, Any] = resp.json()
|
||||
if resp.status_code != 200:
|
||||
raise SystemExit(f"code exchange failed: {json.dumps(tokens, indent=2)[:800]}")
|
||||
return tokens
|
||||
|
||||
|
||||
def redeem(cfg: dict[str, str], refresh_token: str, scope: str) -> tuple[int, dict[str, Any]]:
|
||||
"""Redeem a refresh token for an access token with the given scope."""
|
||||
resp = httpx.post(
|
||||
f"https://login.microsoftonline.com/{cfg['ENTRA_TENANT_ID']}/oauth2/v2.0/token",
|
||||
data={
|
||||
"grant_type": "refresh_token",
|
||||
"refresh_token": refresh_token,
|
||||
"client_id": cfg["ENTRA_CLIENT_ID"],
|
||||
"client_secret": cfg["ENTRA_CLIENT_SECRET"],
|
||||
"scope": scope,
|
||||
},
|
||||
timeout=15.0,
|
||||
)
|
||||
body: dict[str, Any] = resp.json()
|
||||
return resp.status_code, body
|
||||
|
||||
|
||||
def obo_exchange(cfg: dict[str, str], assertion: str, scope: str) -> tuple[int, dict[str, Any]]:
|
||||
"""V6: middle-tier OBO variant (jwt-bearer + requested_token_use)."""
|
||||
resp = httpx.post(
|
||||
f"https://login.microsoftonline.com/{cfg['ENTRA_TENANT_ID']}/oauth2/v2.0/token",
|
||||
data={
|
||||
"grant_type": "urn:ietf:params:oauth:grant-type:jwt-bearer",
|
||||
"assertion": assertion,
|
||||
"client_id": cfg["ENTRA_CLIENT_ID"],
|
||||
"client_secret": cfg["ENTRA_CLIENT_SECRET"],
|
||||
"scope": scope,
|
||||
"requested_token_use": "on_behalf_of",
|
||||
},
|
||||
timeout=15.0,
|
||||
)
|
||||
body: dict[str, Any] = resp.json()
|
||||
return resp.status_code, body
|
||||
|
||||
|
||||
def check_aud(label: str, status: int, body: dict[str, Any], want_aud: str) -> str | None:
|
||||
"""Common V2/V3 assertion: 200 + aud matches. Returns the new RT if any."""
|
||||
if status != 200:
|
||||
record(label, "FAILED", f"HTTP {status}: {json.dumps(body)[:300]}")
|
||||
return None
|
||||
claims = jwt_claims_unverified(body.get("access_token", ""))
|
||||
aud = str(claims.get("aud", "<none>"))
|
||||
ok = aud == want_aud or aud == want_aud.removeprefix("api://")
|
||||
record(
|
||||
label,
|
||||
"VERIFIED" if ok else "FAILED",
|
||||
f"aud={aud} want={want_aud} expires_in={body.get('expires_in')} "
|
||||
f"new_rt={redact(body.get('refresh_token'))}",
|
||||
)
|
||||
new_rt = body.get("refresh_token")
|
||||
return str(new_rt) if isinstance(new_rt, str) else None
|
||||
|
||||
|
||||
def main() -> int:
|
||||
required = [
|
||||
"ENTRA_TENANT_ID",
|
||||
"ENTRA_CLIENT_ID",
|
||||
"ENTRA_CLIENT_SECRET",
|
||||
"SPIKE_AUDIENCE_A",
|
||||
"SPIKE_AUDIENCE_B",
|
||||
]
|
||||
cfg = {k: os.environ[k] for k in required if k in os.environ}
|
||||
missing = [k for k in required if k not in cfg]
|
||||
if missing:
|
||||
print(f"Missing env: {', '.join(missing)}\nSee module docstring.")
|
||||
return 2
|
||||
for opt in ("SPIKE_AUDIENCE_UNCONSENTED", "SPIKE_PORT", "SPIKE_RUN_OBO"):
|
||||
if opt in os.environ:
|
||||
cfg[opt] = os.environ[opt]
|
||||
|
||||
# V1 - capture
|
||||
tokens = interactive_login(cfg)
|
||||
rt0 = tokens.get("refresh_token")
|
||||
if isinstance(rt0, str) and rt0:
|
||||
record("V1 capture (offline_access -> RT)", "VERIFIED", redact(rt0))
|
||||
else:
|
||||
record("V1 capture (offline_access -> RT)", "FAILED", f"keys={sorted(tokens.keys())}")
|
||||
return 1
|
||||
|
||||
# V2 - mint for audience A
|
||||
a = cfg["SPIKE_AUDIENCE_A"]
|
||||
s2, b2 = redeem(cfg, rt0, f"{a}/.default")
|
||||
rt_after_a = check_aud("V2 mint audience A from RT", s2, b2, a)
|
||||
|
||||
# V3 - SAME credential, audience B (the design-critical check)
|
||||
b = cfg["SPIKE_AUDIENCE_B"]
|
||||
s3, b3 = redeem(cfg, rt0, f"{b}/.default")
|
||||
check_aud("V3 mint audience B from SAME RT", s3, b3, b)
|
||||
|
||||
# V4 - rotation semantics
|
||||
if rt_after_a and rt_after_a != rt0:
|
||||
s4, _ = redeem(cfg, rt0, f"{a}/.default")
|
||||
record(
|
||||
"V4 rotation (new RT returned; old still valid?)",
|
||||
"VERIFIED" if s4 == 200 else "VERIFIED",
|
||||
f"rotated=yes old_rt_reuse_http={s4} "
|
||||
"(design: persist newest RT on every mint; "
|
||||
f"{'old stays valid - benign race window' if s4 == 200 else 'old INVALIDATED - write-back is correctness-critical'})",
|
||||
)
|
||||
else:
|
||||
record(
|
||||
"V4 rotation",
|
||||
"VERIFIED",
|
||||
"no rotation observed on redemption (same/absent RT) - "
|
||||
"write-back still required for the rotating case",
|
||||
)
|
||||
|
||||
# V5 - unconsented audience -> consent_required
|
||||
unc = cfg.get("SPIKE_AUDIENCE_UNCONSENTED")
|
||||
if unc:
|
||||
s5, b5 = redeem(cfg, rt0, f"{unc}/.default")
|
||||
codes = b5.get("error_codes", [])
|
||||
hit = s5 == 400 and (65001 in codes or b5.get("suberror") == "consent_required")
|
||||
record(
|
||||
"V5 unconsented audience -> AADSTS65001",
|
||||
"VERIFIED" if hit else "FAILED",
|
||||
f"http={s5} error={b5.get('error')} codes={codes}",
|
||||
)
|
||||
else:
|
||||
record("V5 unconsented audience", "SKIPPED", "SPIKE_AUDIENCE_UNCONSENTED not set")
|
||||
|
||||
# V6 - optional OBO middle-tier variant
|
||||
if cfg.get("SPIKE_RUN_OBO") == "1":
|
||||
s6a, b6a = redeem(cfg, rt0, f"{cfg['ENTRA_CLIENT_ID']}/.default")
|
||||
at_self = b6a.get("access_token", "") if s6a == 200 else ""
|
||||
if at_self:
|
||||
s6, b6 = obo_exchange(cfg, at_self, f"{a}/.default")
|
||||
check_aud("V6 OBO jwt-bearer variant", s6, b6, a)
|
||||
else:
|
||||
record(
|
||||
"V6 OBO jwt-bearer variant",
|
||||
"FAILED",
|
||||
f"could not mint self-audience assertion: HTTP {s6a}",
|
||||
)
|
||||
else:
|
||||
record("V6 OBO jwt-bearer variant", "SKIPPED", "SPIKE_RUN_OBO != 1")
|
||||
|
||||
print("\n=== summary ===")
|
||||
for check, status, _ in RESULTS:
|
||||
print(f" {status:>8} {check}")
|
||||
return 0 if all(s != "FAILED" for _, s, _ in RESULTS) else 1
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
sys.exit(main())
|
||||
@@ -1,271 +0,0 @@
|
||||
"""End-to-end exercise of the oauth_obo feature on the OSS path (RFC 8693).
|
||||
|
||||
Parallel to ``entra_e2e.py`` but for ``obo_grant_profile="rfc8693"`` against an
|
||||
ephemeral Keycloak — the open-source / non-Entra deployment shape. Fully
|
||||
headless (password grant, no browser), so it runs unattended.
|
||||
|
||||
Drives the REAL Turnstone code: ``MCPTokenStore.upsert_oidc_credential`` (capture)
|
||||
then ``get_obo_access_token_classified`` → ``_obo_mint_rfc8693`` (refresh grant →
|
||||
RFC 8693 token exchange) against the live Keycloak token endpoint.
|
||||
|
||||
Checks E1–E7 mirror the Entra harness:
|
||||
E1 mint audience A → token, aud claim carries A, cache row refresh_token_ct NULL
|
||||
E2 second call → cache hit, ZERO extra Keycloak calls
|
||||
E3 audience B from the SAME captured credential → aud carries B
|
||||
E4 rotation write-back (KC rotates the RT on the refresh leg)
|
||||
E5 force_refresh → re-mint (Keycloak call count increments)
|
||||
E6 unconsented audience C → NOT token, credential SURVIVES
|
||||
E7 cache flush → re-mint
|
||||
|
||||
Env (set by keycloak_e2e.sh):
|
||||
KC_TOKEN_ENDPOINT, KC_ISSUER, KC_CLIENT_ID, KC_CLIENT_SECRET,
|
||||
KC_USER, KC_PASSWORD, AUD_A, SCOPE_A, AUD_B, SCOPE_B, AUD_C
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import base64
|
||||
import json
|
||||
import os
|
||||
import sys
|
||||
import tempfile
|
||||
from types import SimpleNamespace
|
||||
from typing import Any
|
||||
|
||||
import httpx
|
||||
|
||||
from turnstone.core.mcp_crypto import (
|
||||
MCPTokenCipher,
|
||||
MCPTokenCipherConfig,
|
||||
MCPTokenStore,
|
||||
)
|
||||
from turnstone.core.mcp_oauth import get_obo_access_token_classified
|
||||
from turnstone.core.oidc import OIDCConfig
|
||||
from turnstone.core.storage._sqlite import SQLiteBackend
|
||||
|
||||
USER = "e2e-user"
|
||||
RESULTS: list[tuple[str, str]] = []
|
||||
|
||||
|
||||
def record(status: str, msg: str) -> None:
|
||||
RESULTS.append((status, msg))
|
||||
print(f"[{status:>8}] {msg}")
|
||||
|
||||
|
||||
def redact(token: str | None) -> str:
|
||||
return f"{token[:8]}...({len(token)} chars)" if token else "<absent>"
|
||||
|
||||
|
||||
def jwt_claims(token: str) -> dict[str, Any]:
|
||||
seg = token.split(".")[1]
|
||||
pad = "=" * (-len(seg) % 4)
|
||||
out: dict[str, Any] = json.loads(base64.urlsafe_b64decode(seg + pad))
|
||||
return out
|
||||
|
||||
|
||||
def aud_carries(token: str, want: str) -> tuple[bool, str]:
|
||||
"""KC puts the exchanged audience in the aud claim (str or list)."""
|
||||
aud = jwt_claims(token).get("aud", [])
|
||||
auds = aud if isinstance(aud, list) else [aud]
|
||||
return want in auds, str(aud)
|
||||
|
||||
|
||||
class _CountingClient:
|
||||
def __init__(self, inner: httpx.AsyncClient) -> None:
|
||||
self._inner = inner
|
||||
self.posts = 0
|
||||
|
||||
async def post(self, *args: Any, **kwargs: Any) -> httpx.Response:
|
||||
self.posts += 1
|
||||
return await self._inner.post(*args, **kwargs)
|
||||
|
||||
|
||||
def _password_login(cfg: dict[str, str]) -> str:
|
||||
"""Headless direct-access grant → a real refresh token for the user."""
|
||||
resp = httpx.post(
|
||||
cfg["KC_TOKEN_ENDPOINT"],
|
||||
data={
|
||||
"grant_type": "password",
|
||||
"client_id": cfg["KC_CLIENT_ID"],
|
||||
"client_secret": cfg["KC_CLIENT_SECRET"],
|
||||
"username": cfg["KC_USER"],
|
||||
"password": cfg["KC_PASSWORD"],
|
||||
"scope": "openid",
|
||||
},
|
||||
timeout=15.0,
|
||||
)
|
||||
resp.raise_for_status()
|
||||
return str(resp.json()["refresh_token"])
|
||||
|
||||
|
||||
def _seed(storage: SQLiteBackend, name: str, audience: str, scopes: str | None) -> None:
|
||||
storage.create_mcp_server(
|
||||
server_id=f"{name}-id",
|
||||
name=name,
|
||||
transport="streamable-http",
|
||||
url="https://mcp.example.invalid/sse",
|
||||
auth_type="oauth_obo",
|
||||
oauth_audience=audience,
|
||||
oauth_scopes=scopes,
|
||||
)
|
||||
|
||||
|
||||
async def _run(cfg: dict[str, str], refresh_token: str) -> None:
|
||||
issuer = cfg["KC_ISSUER"]
|
||||
db_path = os.path.join(tempfile.mkdtemp(prefix="obo-kc-e2e-"), "e2e.db")
|
||||
storage = SQLiteBackend(db_path)
|
||||
from cryptography.fernet import Fernet
|
||||
|
||||
raw = base64.urlsafe_b64decode(Fernet.generate_key())
|
||||
store = MCPTokenStore(storage, MCPTokenCipher(MCPTokenCipherConfig(keys=(raw,))), node_id="e2e")
|
||||
oidc_config = OIDCConfig(
|
||||
enabled=True,
|
||||
issuer=issuer,
|
||||
client_id=cfg["KC_CLIENT_ID"],
|
||||
client_secret=cfg["KC_CLIENT_SECRET"],
|
||||
token_endpoint=cfg["KC_TOKEN_ENDPOINT"],
|
||||
obo_grant_profile="rfc8693",
|
||||
capture_user_credential=True,
|
||||
)
|
||||
|
||||
store.upsert_oidc_credential(USER, issuer, refresh_token=refresh_token)
|
||||
cap = store.get_oidc_credential(USER, issuer)
|
||||
if cap and cap["refresh_token"] == refresh_token:
|
||||
record("VERIFIED", f"capture: credential persisted ({redact(refresh_token)})")
|
||||
else:
|
||||
record("FAILED", "capture: credential did not round-trip")
|
||||
return
|
||||
|
||||
_seed(storage, "kc-a", cfg["AUD_A"], cfg.get("SCOPE_A"))
|
||||
_seed(storage, "kc-b", cfg["AUD_B"], cfg.get("SCOPE_B"))
|
||||
if cfg.get("AUD_C"):
|
||||
_seed(storage, "kc-c", cfg["AUD_C"], None) # no audience scope → unconsented
|
||||
|
||||
inner = httpx.AsyncClient(timeout=20.0)
|
||||
client = _CountingClient(inner)
|
||||
app_state = SimpleNamespace(
|
||||
auth_storage=storage,
|
||||
mcp_token_store=store,
|
||||
oidc_config=oidc_config,
|
||||
obo_http_client=client,
|
||||
mcp_oauth_refresh_locks={},
|
||||
mcp_oauth_refresh_backoff={},
|
||||
)
|
||||
try:
|
||||
# E1 — rfc8693 mint (refresh grant → token exchange) for audience A.
|
||||
r = await get_obo_access_token_classified(
|
||||
app_state=app_state, user_id=USER, server_name="kc-a"
|
||||
)
|
||||
if r.kind == "token" and r.token:
|
||||
ok, aud = aud_carries(r.token, cfg["AUD_A"])
|
||||
row = storage.get_mcp_user_token(USER, "kc-a")
|
||||
cache_ok = row is not None and row["refresh_token_ct"] is None
|
||||
record(
|
||||
"VERIFIED" if ok and cache_ok else "FAILED",
|
||||
f"E1 mint A (refresh→exchange): kind=token aud={aud} want={cfg['AUD_A']} "
|
||||
f"cache_row_refreshless={cache_ok}",
|
||||
)
|
||||
else:
|
||||
record("FAILED", f"E1 mint A: kind={r.kind} (expected token)")
|
||||
return
|
||||
|
||||
# E2 — cache hit.
|
||||
posts_before = client.posts
|
||||
r2 = await get_obo_access_token_classified(
|
||||
app_state=app_state, user_id=USER, server_name="kc-a"
|
||||
)
|
||||
record(
|
||||
"VERIFIED" if r2.kind == "token" and client.posts == posts_before else "FAILED",
|
||||
f"E2 cache hit: kind={r2.kind} extra_kc_calls={client.posts - posts_before} (want 0)",
|
||||
)
|
||||
|
||||
# E3 — audience B from the SAME credential.
|
||||
rb = await get_obo_access_token_classified(
|
||||
app_state=app_state, user_id=USER, server_name="kc-b"
|
||||
)
|
||||
if rb.kind == "token" and rb.token:
|
||||
ok_b, aud_b = aud_carries(rb.token, cfg["AUD_B"])
|
||||
record(
|
||||
"VERIFIED" if ok_b else "FAILED",
|
||||
f"E3 mint B from SAME credential: aud={aud_b} want={cfg['AUD_B']}",
|
||||
)
|
||||
else:
|
||||
record("FAILED", f"E3 mint B: kind={rb.kind}")
|
||||
|
||||
# E4 — rotation write-back (KC rotates the RT on the refresh leg).
|
||||
cred_now = store.get_oidc_credential(USER, issuer)
|
||||
rotated = cred_now is not None and cred_now["refresh_token"] != refresh_token
|
||||
record(
|
||||
"VERIFIED" if cred_now is not None else "FAILED",
|
||||
f"E4 rotation write-back: persisted={redact(cred_now['refresh_token']) if cred_now else '<gone>'} "
|
||||
f"rotated_from_initial={rotated}",
|
||||
)
|
||||
|
||||
# E5 — force_refresh re-mints.
|
||||
posts_before = client.posts
|
||||
rf = await get_obo_access_token_classified(
|
||||
app_state=app_state, user_id=USER, server_name="kc-a", force_refresh=True
|
||||
)
|
||||
record(
|
||||
"VERIFIED" if rf.kind == "token" and client.posts > posts_before else "FAILED",
|
||||
f"E5 force_refresh re-mint: kind={rf.kind} kc_calls={client.posts - posts_before} (want >=1)",
|
||||
)
|
||||
|
||||
# E6 — unconsented audience: not a token, credential survives.
|
||||
if cfg.get("AUD_C"):
|
||||
rc = await get_obo_access_token_classified(
|
||||
app_state=app_state, user_id=USER, server_name="kc-c"
|
||||
)
|
||||
cred_after = store.get_oidc_credential(USER, issuer)
|
||||
record(
|
||||
"VERIFIED" if rc.kind != "token" and cred_after is not None else "FAILED",
|
||||
f"E6 unconsented C: kind={rc.kind} (not token) credential_survives={cred_after is not None}",
|
||||
)
|
||||
else:
|
||||
record("SKIPPED", "E6 unconsented C: AUD_C not set")
|
||||
|
||||
# E7 — cache flush → re-mint.
|
||||
store.delete_user_token(USER, "kc-a")
|
||||
posts_before = client.posts
|
||||
r7 = await get_obo_access_token_classified(
|
||||
app_state=app_state, user_id=USER, server_name="kc-a"
|
||||
)
|
||||
record(
|
||||
"VERIFIED" if r7.kind == "token" and client.posts > posts_before else "FAILED",
|
||||
f"E7 flush→re-mint: kind={r7.kind} kc_calls={client.posts - posts_before} (want >=1)",
|
||||
)
|
||||
finally:
|
||||
await inner.aclose()
|
||||
|
||||
|
||||
def main() -> int:
|
||||
required = [
|
||||
"KC_TOKEN_ENDPOINT",
|
||||
"KC_ISSUER",
|
||||
"KC_CLIENT_ID",
|
||||
"KC_CLIENT_SECRET",
|
||||
"KC_USER",
|
||||
"KC_PASSWORD",
|
||||
"AUD_A",
|
||||
"AUD_B",
|
||||
]
|
||||
cfg = {k: os.environ[k] for k in os.environ if k.startswith(("KC_", "AUD_", "SCOPE_"))}
|
||||
missing = [k for k in required if not cfg.get(k)]
|
||||
if missing:
|
||||
print(f"Missing env: {', '.join(missing)} — run via keycloak_e2e.sh")
|
||||
return 2
|
||||
|
||||
print("Headless password login to Keycloak (the credential the feature captures)...")
|
||||
refresh_token = _password_login(cfg)
|
||||
|
||||
asyncio.run(_run(cfg, refresh_token))
|
||||
|
||||
print("\n=== summary ===")
|
||||
for status, msg in RESULTS:
|
||||
print(f" {status:>8} {msg}")
|
||||
return 0 if all(s in ("VERIFIED", "SKIPPED") for s, _ in RESULTS) else 1
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
sys.exit(main())
|
||||
@@ -1,65 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
# OSS-path (RFC 8693) end-to-end: spin up ephemeral Keycloak, configure the
|
||||
# realm, run keycloak_e2e.py against the REAL Turnstone mint engine, tear down.
|
||||
# Fully headless — no browser. Manual test tooling, not run in CI.
|
||||
set -euo pipefail
|
||||
cd "$(dirname "$0")/../.." # repo root (uv run needs it)
|
||||
|
||||
CONTAINER=kc-obo-e2e
|
||||
PORT=8091
|
||||
KC="docker exec $CONTAINER /opt/keycloak/bin/kcadm.sh"
|
||||
|
||||
cleanup() { docker rm -f "$CONTAINER" >/dev/null 2>&1 || true; }
|
||||
trap cleanup EXIT
|
||||
cleanup
|
||||
|
||||
echo ">> starting Keycloak 26.3 (ephemeral)..."
|
||||
docker run -d --name "$CONTAINER" -p "127.0.0.1:${PORT}:8080" \
|
||||
-e KC_BOOTSTRAP_ADMIN_USERNAME=admin -e KC_BOOTSTRAP_ADMIN_PASSWORD=admin \
|
||||
quay.io/keycloak/keycloak:26.3 start-dev >/dev/null
|
||||
|
||||
echo ">> waiting for Keycloak (dev-mode boot can take a few minutes on a loaded host)..."
|
||||
# Wait on kcadm auth succeeding directly — more reliable than the host HTTP port,
|
||||
# and generous enough for a resource-starved boot (up to ~6 min).
|
||||
ready=""
|
||||
for _ in $(seq 1 90); do
|
||||
if $KC config credentials --server http://localhost:8080 --realm master \
|
||||
--user admin --password admin >/dev/null 2>&1; then
|
||||
ready=1
|
||||
break
|
||||
fi
|
||||
sleep 4
|
||||
done
|
||||
[ -n "$ready" ] || { echo "Keycloak did not become ready in time"; docker logs "$CONTAINER" 2>&1 | tail -15; exit 1; }
|
||||
|
||||
echo ">> configuring realm 'spike'..."
|
||||
$KC create realms -s realm=spike -s enabled=true >/dev/null
|
||||
# Confidential client with standard token exchange (the RFC 8693 leg) + direct
|
||||
# access grant (headless password login to fetch the user's refresh token).
|
||||
$KC create clients -r spike -s clientId=turnstone -s enabled=true -s publicClient=false \
|
||||
-s secret=spike-secret -s directAccessGrantsEnabled=true \
|
||||
-s 'attributes={"standard.token.exchange.enabled":"true"}' >/dev/null
|
||||
for t in mcp-a mcp-b mcp-c; do
|
||||
$KC create clients -r spike -s clientId=$t -s enabled=true -s publicClient=false -s secret=x >/dev/null
|
||||
done
|
||||
$KC create users -r spike -s username=e2e-user -s enabled=true -s email=e2e@spike.test \
|
||||
-s emailVerified=true -s firstName=E2E -s lastName=User >/dev/null
|
||||
$KC set-password -r spike --username e2e-user --new-password e2e-pw >/dev/null
|
||||
|
||||
TURNSTONE_UUID=$($KC get clients -r spike -q clientId=turnstone --fields id --format csv --noquotes)
|
||||
# Audience client scopes for mcp-a and mcp-b ONLY (mcp-c stays unconsented → E6).
|
||||
for t in mcp-a mcp-b; do
|
||||
SID=$($KC create client-scopes -r spike -s name=aud-$t -s protocol=openid-connect -i)
|
||||
$KC create "client-scopes/$SID/protocol-mappers/models" -r spike -s name=aud-$t \
|
||||
-s protocol=openid-connect -s protocolMapper=oidc-audience-mapper \
|
||||
-s "config={\"included.client.audience\":\"$t\",\"access.token.claim\":\"true\"}" >/dev/null
|
||||
$KC update "clients/$TURNSTONE_UUID/optional-client-scopes/$SID" -r spike >/dev/null
|
||||
done
|
||||
|
||||
echo ">> running the product e2e harness..."
|
||||
export KC_TOKEN_ENDPOINT="http://127.0.0.1:${PORT}/realms/spike/protocol/openid-connect/token"
|
||||
export KC_ISSUER="http://127.0.0.1:${PORT}/realms/spike"
|
||||
export KC_CLIENT_ID=turnstone KC_CLIENT_SECRET=spike-secret
|
||||
export KC_USER=e2e-user KC_PASSWORD=e2e-pw
|
||||
export AUD_A=mcp-a SCOPE_A=aud-mcp-a AUD_B=mcp-b SCOPE_B=aud-mcp-b AUD_C=mcp-c
|
||||
uv run python scripts/obo-e2e/keycloak_e2e.py
|
||||
@@ -1,962 +0,0 @@
|
||||
#!/usr/bin/env python3
|
||||
"""Browser-level SSE recovery livepass — boots the REAL interactive.js
|
||||
``InteractivePane`` against a REAL Turnstone node and drives the two
|
||||
headline recovery scenarios through headless Chrome over CDP, stamping
|
||||
``document.title`` verdicts (``RECOVERY-READY-*`` / ``RECOVERY-FAILED-*``)
|
||||
the livepass convention.
|
||||
|
||||
Unlike ``scripts/livepass.py`` (which stubs ``window.authFetch`` with
|
||||
canned fixtures), this page uses the REAL auth + REAL EventSource against a
|
||||
REAL node: the page is served same-origin by the node itself (so cookie
|
||||
auth and EventSource just work), the node runs a scripted-provider
|
||||
workstream so a REST ``/send`` drives a real bash storm, and the pane's
|
||||
own state machine (interactive.js) does the recovery.
|
||||
|
||||
Usage::
|
||||
|
||||
python3 scripts/recovery_e2e.py # run all three scenarios
|
||||
python3 scripts/recovery_e2e.py --scenario storm
|
||||
python3 scripts/recovery_e2e.py --scenario restart
|
||||
python3 scripts/recovery_e2e.py --scenario coord-restart
|
||||
python3 scripts/recovery_e2e.py --scenario both # A+B only (legacy)
|
||||
python3 scripts/recovery_e2e.py --keep-open 8971 # serve the storm page
|
||||
# for manual inspection
|
||||
|
||||
Scenario A (storm): the page connects, POSTs ``/send`` on stream-open (so
|
||||
the listener is registered first), the node runs a 4-parallel-bash
|
||||
``seq 1 500`` storm plus a task_agent whose sub-tools are chatty bashes;
|
||||
the page asserts the final DOM has the expected top-level tool rows, the
|
||||
task_agent card nests its sub-tool rows (NO child escaped to the top
|
||||
level), and the composer settles idle. Stamps ``RECOVERY-READY-STORM-<n>``.
|
||||
|
||||
Scenario B (hide mid-turn -> restart -> show): the runner hides the tab
|
||||
the moment the first streamed line paints (freezing the pane's cursor at
|
||||
a mid-turn event id — the MessageEvent ``lastEventId`` capture is what
|
||||
makes that cursor real; the pre-2026-07 object-form read left it null and
|
||||
this whole path unassertable), lets the turn and a follow-up text commit
|
||||
while hidden, restarts the node on the SAME port (fresh empty ring,
|
||||
storage-seeded counter), then shows the tab. The show-edge reconnect
|
||||
presents the stale cursor, MUST draw ``replay_truncated`` (asserted:
|
||||
trunc>=1), the truncated resync rebuilds from /history, and the turns
|
||||
committed during the hide window MUST be present afterwards (asserted:
|
||||
``healed`` — the 'turn disappeared' field symptom). Stamps
|
||||
``RECOVERY-READY-RESTART-rows<n>-trunc<n>``. The exact ``lost_count``
|
||||
arithmetic and the failed-resync retry stay at the server-contract level
|
||||
in Tier 1's ``test_restart_truncated_honesty`` /
|
||||
``test_failed_resync_retries_via_truncation_record``.
|
||||
|
||||
A NOTE ON THE BROWSER OVERFLOW (server-side poison): a real listener-queue
|
||||
poison needs the browser to STOP reading the socket so TCP backpressure
|
||||
reaches the server. A backgrounded/CPU-throttled tab does NOT do this --
|
||||
Chrome's network stack keeps draining the socket regardless of JS
|
||||
throttling, and interactive.js deliberately CLOSES the stream on tab-hide
|
||||
rather than starving it. So the server-side overflow -> stream_overflow ->
|
||||
reconnect path is NOT reliably forcible from a real browser (which is why
|
||||
that field bug was subtle); it is proven at the server-contract level in
|
||||
``tests/test_sse_recovery_e2e.py::test_slow_consumer_overflow_then_lossless_reconnect``.
|
||||
Scenario A here proves the OTHER half at the browser level: fix-3's
|
||||
de-amplified storm renders correctly with no escaped sub-agent children.
|
||||
|
||||
MANUAL RUNBOOK (if Chrome/CDP is unavailable): run this with
|
||||
``--keep-open PORT`` to boot the node + serve the storm page, open the
|
||||
printed URL in a browser (the script prints the auth cookie to set), and
|
||||
watch ``document.title``. For the restart scenario, boot with a fixed
|
||||
port, load the restart page, background the tab, restart the node
|
||||
(``RecoveryServer`` on the same port), foreground the tab, and watch the
|
||||
title settle to ``RECOVERY-READY-RESTART``.
|
||||
|
||||
Scenario C (coord-restart): the REAL coordinator pane
|
||||
(console/static/coordinator/coordinator.js — the #882 parity port of the
|
||||
same truncated-recovery machinery) driven through the SAME hide -> restart
|
||||
-> show sequence as Scenario B. The coordinator only runs under the
|
||||
console app in production, and the console's coordinator subsystems build
|
||||
inside its server lifespan against a config-resolved model registry — no
|
||||
``create_app(prebuilt SessionManager)`` seam for this harness's scripted
|
||||
provider. So the scenario mounts the pane against the interactive
|
||||
recovery node instead: the node serves the console's coordinator static
|
||||
tree at ``/coord-static`` (a distinct prefix — the node's own ``/static``
|
||||
mount would swallow the console path) and a pane-only page at
|
||||
``/coord-recovery``; the pane's module imports are all absolute
|
||||
``/shared/*`` and resolve against the node. Fidelity caveats, all inert
|
||||
for the recovery machinery under test: the workstream is
|
||||
interactive-kind (no coordinator status events — the status bar keeps its
|
||||
placeholder), and ``/children`` + ``/tasks`` 404 here (the pane's loaders
|
||||
catch and render empty by design). What IS real: the full chrome
|
||||
(buildCoordChrome), cookie auth, EventSource + MessageEvent cursor
|
||||
capture, the connect chokepoint, the dead-stream resync
|
||||
(loadHistoryThenReconnect), the churn limiter, and the jitter. Asserted:
|
||||
the show-edge reconnect draws ``replay_truncated`` (trunc>=1, counted at
|
||||
the transport by a page-side EventSource wrapper — the coordinator's
|
||||
handleEvent, cursor, and even its SSE indicator are closure-private or
|
||||
deliberately absent from the chrome), the resync rebuilds from /history
|
||||
with the hidden-window turns present (``healed``), tool rows intact, the
|
||||
stream re-opened post-show, the status bar not stuck dim, and idle
|
||||
asserted server-side by the runner. Stamps
|
||||
``RECOVERY-READY-COORD-rows<n>-trunc<n>``.
|
||||
|
||||
MANUAL COORDINATOR RUNBOOK (real console topology, no CDP): boot a dev
|
||||
console + one node (docker-compose dev cluster), open a coordinator with
|
||||
running children, hide the tab mid-turn, restart the CONSOLE process (the
|
||||
coordinator ring lives there), show the tab, and verify: the pane draws
|
||||
one truncated full rebuild (no blank pane), the mid-run turn's tool rows
|
||||
re-appear inside their batch (no standalone top-level orphan bubbles),
|
||||
and turns committed while hidden are present.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import base64
|
||||
import contextlib
|
||||
import json
|
||||
import os
|
||||
import shutil
|
||||
import socket
|
||||
import struct
|
||||
import subprocess
|
||||
import sys
|
||||
import time
|
||||
import urllib.request
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
|
||||
|
||||
# Healed-gap sentinel for scenario B: injected as the scripted turn-2 text
|
||||
# AND threaded to the page via ``?healed=`` (read into ``healedSentinel``),
|
||||
# so the injected text and the DOM check share one definition. Must never
|
||||
# collide with rendered command/output text — the bash command row paints
|
||||
# its shell source verbatim, which contains the keyword ``done``.
|
||||
HEALED_SENTINEL = "HEALED-e5b1"
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# The recovery page — served same-origin by the node at /recovery.
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
PAGE_HTML = r"""<!doctype html>
|
||||
<html lang="en">
|
||||
<head>
|
||||
<meta charset="utf-8" />
|
||||
<title>recovery livepass</title>
|
||||
<link rel="stylesheet" href="/shared/base.css" />
|
||||
<link rel="stylesheet" href="/shared/ui-base.css" />
|
||||
<link rel="stylesheet" href="/shared/chat.css" />
|
||||
<link rel="stylesheet" href="/shared/conversation.css" />
|
||||
<link rel="stylesheet" href="/shared/cards.css" />
|
||||
<link rel="stylesheet" href="/shared/interactive.css" />
|
||||
<style>
|
||||
body { margin: 0; background: var(--bg); color: var(--ink); }
|
||||
#mount { height: 100vh; display: flex; }
|
||||
#mount > * { flex: 1; min-height: 0; }
|
||||
</style>
|
||||
</head>
|
||||
<body>
|
||||
<div id="header"><div id="status-bar"></div></div>
|
||||
<div id="mount"></div>
|
||||
<script>
|
||||
// Minimal globals interactive.js reads on the standalone path.
|
||||
window.showToast = function (m) { console.log("toast:", m); };
|
||||
window.showLogin = function () {};
|
||||
</script>
|
||||
<script type="module">
|
||||
import { InteractivePane } from "/shared/interactive.js";
|
||||
|
||||
const q = new URLSearchParams(location.search);
|
||||
const wsId = q.get("ws_id");
|
||||
const scenario = q.get("scenario") || "storm";
|
||||
const expectRows = parseInt(q.get("rows") || "4", 10);
|
||||
// Healed-gap sentinel, threaded from the runner (HEALED_SENTINEL)
|
||||
// so the injected turn text and this check cannot drift apart.
|
||||
const healedSentinel = q.get("healed") || "";
|
||||
|
||||
// REAL pane against THIS origin (base=""): real authFetch (cookie) and
|
||||
// real EventSource. The default host provides all SSE seams.
|
||||
const pane = new InteractivePane(wsId, { base: "" });
|
||||
document.getElementById("mount").appendChild(pane.el);
|
||||
pane.wsId = wsId;
|
||||
window.__pane = pane;
|
||||
|
||||
window.__hide = function () {
|
||||
Object.defineProperty(document, "hidden", { configurable: true, value: true });
|
||||
Object.defineProperty(document, "visibilityState", { configurable: true, value: "hidden" });
|
||||
document.dispatchEvent(new Event("visibilitychange"));
|
||||
};
|
||||
window.__show = function () {
|
||||
Object.defineProperty(document, "hidden", { configurable: true, value: false });
|
||||
Object.defineProperty(document, "visibilityState", { configurable: true, value: "visible" });
|
||||
document.dispatchEvent(new Event("visibilitychange"));
|
||||
};
|
||||
|
||||
// Count top-level tool rows and escaped sub-agent children.
|
||||
function domCounts() {
|
||||
const topRows = pane.messagesEl.querySelectorAll(
|
||||
".conv-batch > .conv-row[data-call-id]"
|
||||
);
|
||||
let topLevel = 0;
|
||||
let escapedChildren = 0;
|
||||
topRows.forEach((r) => {
|
||||
const cid = r.dataset.callId || "";
|
||||
if (cid.includes("::")) escapedChildren += 1; // a child at the top level
|
||||
else topLevel += 1;
|
||||
});
|
||||
const agentCard = pane.messagesEl.querySelector(".conv-agent");
|
||||
const nested = pane.messagesEl.querySelectorAll(
|
||||
".conv-agent .conv-row[data-call-id]"
|
||||
).length;
|
||||
return { topLevel, escapedChildren, agentCard: !!agentCard, nested };
|
||||
}
|
||||
|
||||
let sent = false;
|
||||
function sendOnce(msg) {
|
||||
if (sent) return;
|
||||
sent = true;
|
||||
// The pane's SSE is open (host.onStreamOpen fired), so the listener is
|
||||
// registered before this /send -- no missed events.
|
||||
window
|
||||
.authFetch("/v1/api/workstreams/" + encodeURIComponent(wsId) + "/send", {
|
||||
method: "POST",
|
||||
headers: { "Content-Type": "application/json" },
|
||||
body: JSON.stringify({ message: msg }),
|
||||
})
|
||||
.catch((e) => { document.title = "RECOVERY-FAILED-send-" + e; });
|
||||
}
|
||||
|
||||
// Drive /send once the stream is live (wrap the default host hook).
|
||||
const origOpen = pane._host.onStreamOpen.bind(pane._host);
|
||||
pane._host.onStreamOpen = function (p) {
|
||||
origOpen(p);
|
||||
window.__streamOpen = (window.__streamOpen || 0) + 1;
|
||||
if (scenario === "storm") sendOnce("run the storm");
|
||||
else if (scenario === "restart" && window.__streamOpen === 1) sendOnce("run a turn");
|
||||
};
|
||||
|
||||
// First paint the REAL way: /history then connect SSE.
|
||||
pane._loadHistoryThenConnect(wsId);
|
||||
|
||||
if (scenario === "storm") {
|
||||
const deadline = Date.now() + 40000;
|
||||
const poll = () => {
|
||||
const c = domCounts();
|
||||
const idle = !pane.busy;
|
||||
if (c.topLevel >= expectRows && c.agentCard && c.nested >= 2 && idle) {
|
||||
document.title = c.escapedChildren
|
||||
? "RECOVERY-FAILED-escaped-" + c.escapedChildren
|
||||
: "RECOVERY-READY-STORM-" + c.topLevel + "-nested-" + c.nested;
|
||||
return;
|
||||
}
|
||||
if (Date.now() > deadline) {
|
||||
document.title =
|
||||
"RECOVERY-FAILED-STORM-top" + c.topLevel + "-agent" + (c.agentCard ? 1 : 0) +
|
||||
"-nested" + c.nested + "-escaped" + c.escapedChildren + "-busy" + (pane.busy ? 1 : 0);
|
||||
return;
|
||||
}
|
||||
setTimeout(poll, 200);
|
||||
};
|
||||
setTimeout(poll, 400);
|
||||
} else if (scenario === "restart") {
|
||||
// The runner drives hide -> (restart node) -> show via window.__hide/
|
||||
// __show. We watch for the truncated-triggered rebuild + idle settle.
|
||||
window.__truncatedSeen = 0;
|
||||
const origHandle = pane.handleEvent.bind(pane);
|
||||
pane.handleEvent = function (ev) {
|
||||
if (ev && ev.type === "replay_truncated") window.__truncatedSeen += 1;
|
||||
return origHandle(ev);
|
||||
};
|
||||
window.__verifyRestart = function () {
|
||||
// Browser-level restart RECOVERY, full contract: the runner hid
|
||||
// the tab MID-turn (cursor frozen below the commits that land
|
||||
// while hidden), so the show-edge reconnect must present the
|
||||
// stale cursor and draw ``replay_truncated`` (REQUIRED since the
|
||||
// MessageEvent lastEventId capture fix — the pre-fix object-form
|
||||
// read left manual reconnects cursorless and this envelope
|
||||
// unreachable, which is why trunc used to report 0), the
|
||||
// truncated resync must rebuild from /history, and the turns
|
||||
// committed DURING the hide window must be present afterwards
|
||||
// (``healed`` — the 'turn disappeared' field symptom). Composer
|
||||
// idle, status bar not stuck disconnected.
|
||||
const c = domCounts();
|
||||
const idle = !pane.busy;
|
||||
const disc = document.querySelector(".ws-sb-disconnected") !== null;
|
||||
// Sentinel must be collision-proof against everything else the
|
||||
// transcript renders: the paced bash COMMAND row paints its
|
||||
// shell text verbatim (buildConvCmd), which contains the
|
||||
// keyword ``done`` — a plain-word sentinel is vacuously
|
||||
// present whether or not the hidden-window turn survived.
|
||||
// The value rides the ?healed= param (single source:
|
||||
// HEALED_SENTINEL in the runner).
|
||||
const healed =
|
||||
healedSentinel !== "" &&
|
||||
(pane.messagesEl.textContent || "").includes(healedSentinel);
|
||||
const ok =
|
||||
c.topLevel >= 1 &&
|
||||
idle &&
|
||||
!disc &&
|
||||
healed &&
|
||||
window.__truncatedSeen >= 1;
|
||||
document.title = ok
|
||||
? "RECOVERY-READY-RESTART-rows" + c.topLevel + "-trunc" + window.__truncatedSeen
|
||||
: "RECOVERY-FAILED-RESTART-rows" + c.topLevel +
|
||||
"-busy" + (pane.busy ? 1 : 0) + "-disc" + (disc ? 1 : 0) +
|
||||
"-healed" + (healed ? 1 : 0) + "-trunc" + window.__truncatedSeen;
|
||||
};
|
||||
}
|
||||
</script>
|
||||
</body>
|
||||
</html>
|
||||
"""
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# The coordinator recovery page — served same-origin by the node at
|
||||
# /coord-recovery. A near-clone of the production standalone page
|
||||
# (console/static/coordinator/index.html): the same /shared script
|
||||
# substrate (classic theme.js first, then the deferred module set), the
|
||||
# same createCoordinatorPane(document.body, wsId, {standalone:true}) +
|
||||
# connect() bootstrap — with the coordinator files imported from
|
||||
# /coord-static (see the module docstring) and Google-fonts dropped
|
||||
# (hermetic run). Scenario instrumentation reads only public chrome ids.
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
COORD_PAGE_HTML = r"""<!doctype html>
|
||||
<html lang="en">
|
||||
<head>
|
||||
<meta charset="utf-8" />
|
||||
<title>coord recovery livepass</title>
|
||||
<link rel="stylesheet" href="/shared/base.css" />
|
||||
<link rel="stylesheet" href="/shared/ui-base.css" />
|
||||
<link rel="stylesheet" href="/shared/chat.css" />
|
||||
<link rel="stylesheet" href="/shared/conversation.css" />
|
||||
<link rel="stylesheet" href="/shared/mcp_error.css" />
|
||||
<link rel="stylesheet" href="/coord-static/coordinator.css" />
|
||||
<link rel="stylesheet" href="/coord-static/coord-chrome.css" />
|
||||
<style>
|
||||
body { height: 100vh; margin: 0; }
|
||||
</style>
|
||||
</head>
|
||||
<body>
|
||||
<script>
|
||||
// Transport-level instrumentation: the coordinator's handleEvent and
|
||||
// cursor are closure-private (unlike interactive's class methods), and
|
||||
// its chrome deliberately builds NO header/SSE indicator — so every
|
||||
// scenario signal is read off the wire by wrapping EventSource BEFORE
|
||||
// any module loads (classic script = runs before the deferred module
|
||||
// set, so the pane's connectSSE always constructs the wrapper):
|
||||
// __truncatedSeen — replay_truncated frames (the envelope);
|
||||
// __esOpens — stream opens (drives the send; a listener is
|
||||
// registered before /send so no events are missed);
|
||||
// __idFrames — id-bearing frames, i.e. exactly the frames that
|
||||
// advance the pane's reconnect cursor (same
|
||||
// ``!= null && !== ""`` guard as the pane) — the
|
||||
// hide fires only after this proves a live mid-turn
|
||||
// cursor.
|
||||
window.__truncatedSeen = 0;
|
||||
window.__esOpens = 0;
|
||||
window.__idFrames = 0;
|
||||
(function () {
|
||||
const RealES = window.EventSource;
|
||||
function CountingES(url, opts) {
|
||||
const es = new RealES(url, opts);
|
||||
es.addEventListener("open", function () {
|
||||
window.__esOpens += 1;
|
||||
});
|
||||
es.addEventListener("message", function (e) {
|
||||
if (e.lastEventId != null && e.lastEventId !== "") {
|
||||
window.__idFrames += 1;
|
||||
}
|
||||
try {
|
||||
const d = JSON.parse(e.data);
|
||||
if (d && d.type === "replay_truncated") window.__truncatedSeen += 1;
|
||||
} catch (_) {}
|
||||
});
|
||||
return es;
|
||||
}
|
||||
CountingES.prototype = RealES.prototype;
|
||||
CountingES.CONNECTING = RealES.CONNECTING;
|
||||
CountingES.OPEN = RealES.OPEN;
|
||||
CountingES.CLOSED = RealES.CLOSED;
|
||||
window.EventSource = CountingES;
|
||||
})();
|
||||
</script>
|
||||
<script src="/shared/theme.js"></script>
|
||||
<script type="module" src="/shared/utils.js"></script>
|
||||
<script type="module" src="/shared/toast.js"></script>
|
||||
<script type="module" src="/shared/auth.js"></script>
|
||||
<script type="module" src="/shared/kb.js"></script>
|
||||
<script type="module" src="/shared/composer.js"></script>
|
||||
<script type="module" src="/shared/composer_attachments.js"></script>
|
||||
<script type="module" src="/shared/composer_queue.js"></script>
|
||||
<script type="module" src="/shared/status_bar.js"></script>
|
||||
<script type="module" src="/shared/renderer.js"></script>
|
||||
<script type="module">
|
||||
import { createCoordinatorPane } from "/coord-static/coordinator.js";
|
||||
|
||||
const q = new URLSearchParams(location.search);
|
||||
const wsId = q.get("ws_id");
|
||||
const healedSentinel = q.get("healed") || "";
|
||||
|
||||
const pane = createCoordinatorPane(document.body, wsId, {
|
||||
standalone: true,
|
||||
});
|
||||
window.__pane = pane;
|
||||
if (pane) pane.connect();
|
||||
|
||||
window.__hide = function () {
|
||||
Object.defineProperty(document, "hidden", { configurable: true, value: true });
|
||||
Object.defineProperty(document, "visibilityState", { configurable: true, value: "hidden" });
|
||||
document.dispatchEvent(new Event("visibilitychange"));
|
||||
};
|
||||
window.__show = function () {
|
||||
Object.defineProperty(document, "hidden", { configurable: true, value: false });
|
||||
Object.defineProperty(document, "visibilityState", { configurable: true, value: "visible" });
|
||||
document.dispatchEvent(new Event("visibilitychange"));
|
||||
};
|
||||
|
||||
// Drive /send once the stream has OPENED at the transport (__esOpens —
|
||||
// the pane's listener is registered by then, so no events are missed).
|
||||
// The chrome has no SSE pill to poll: the header was deliberately
|
||||
// dropped (see buildCoordChrome's comment).
|
||||
let sent = false;
|
||||
function sendOnce(msg) {
|
||||
if (sent) return;
|
||||
sent = true;
|
||||
window
|
||||
.authFetch("/v1/api/workstreams/" + encodeURIComponent(wsId) + "/send", {
|
||||
method: "POST",
|
||||
headers: { "Content-Type": "application/json" },
|
||||
body: JSON.stringify({ message: msg }),
|
||||
})
|
||||
.catch((e) => { document.title = "RECOVERY-FAILED-COORD-send-" + e; });
|
||||
}
|
||||
const sendPoll = setInterval(() => {
|
||||
if (window.__esOpens >= 1) {
|
||||
clearInterval(sendPoll);
|
||||
sendOnce("run a turn");
|
||||
}
|
||||
}, 100);
|
||||
|
||||
window.__verifyCoordRestart = function () {
|
||||
// Same contract as Scenario B, read off the coordinator's public
|
||||
// chrome + the transport wrapper (idle is asserted SERVER-side by
|
||||
// the runner — the chrome has no state text element): the show-edge
|
||||
// reconnect must present the frozen mid-turn cursor and draw
|
||||
// replay_truncated (trunc>=1), the dead-stream resync must rebuild
|
||||
// from /history with the hidden-window turns present (healed), the
|
||||
// stream must have re-opened after the show (__esOpens >= 2), the
|
||||
// status bar must not be stuck dim (.ws-sb-disconnected removed by
|
||||
// the post-recovery onopen), and the tool rows must be intact.
|
||||
const messages = document.getElementById("coord-messages");
|
||||
const rows = messages
|
||||
? messages.querySelectorAll(".conv-row[data-call-id]").length
|
||||
: 0;
|
||||
const reopened = window.__esOpens >= 2;
|
||||
const disc =
|
||||
document.querySelector("#coord-status-bar.ws-sb-disconnected") !== null;
|
||||
const healed =
|
||||
healedSentinel !== "" &&
|
||||
((messages && messages.textContent) || "").includes(healedSentinel);
|
||||
const ok =
|
||||
rows >= 1 && reopened && !disc && healed && window.__truncatedSeen >= 1;
|
||||
document.title = ok
|
||||
? "RECOVERY-READY-COORD-rows" + rows + "-trunc" + window.__truncatedSeen
|
||||
: "RECOVERY-FAILED-COORD-rows" + rows +
|
||||
"-reopened" + (reopened ? 1 : 0) +
|
||||
"-disc" + (disc ? 1 : 0) + "-healed" + (healed ? 1 : 0) +
|
||||
"-trunc" + window.__truncatedSeen;
|
||||
};
|
||||
</script>
|
||||
</body>
|
||||
</html>
|
||||
"""
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Minimal dependency-free CDP client (WebSocket over a raw socket).
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
class CDP:
|
||||
"""Just enough Chrome DevTools Protocol: navigate, evaluate, set cookie."""
|
||||
|
||||
def __init__(self, ws_url: str) -> None:
|
||||
from urllib.parse import urlsplit
|
||||
|
||||
u = urlsplit(ws_url)
|
||||
self._sock = socket.create_connection((u.hostname, u.port or 80), timeout=10)
|
||||
key = base64.b64encode(os.urandom(16)).decode()
|
||||
path = u.path + (f"?{u.query}" if u.query else "")
|
||||
handshake = (
|
||||
f"GET {path} HTTP/1.1\r\nHost: {u.hostname}:{u.port}\r\n"
|
||||
f"Upgrade: websocket\r\nConnection: Upgrade\r\n"
|
||||
f"Sec-WebSocket-Key: {key}\r\nSec-WebSocket-Version: 13\r\n\r\n"
|
||||
)
|
||||
self._sock.sendall(handshake.encode())
|
||||
resp = b""
|
||||
while b"\r\n\r\n" not in resp:
|
||||
resp += self._sock.recv(4096)
|
||||
if b" 101 " not in resp.split(b"\r\n", 1)[0]:
|
||||
raise RuntimeError(f"CDP websocket handshake failed: {resp[:80]!r}")
|
||||
self._id = 0
|
||||
self._rbuf = b""
|
||||
|
||||
def _send(self, payload: bytes) -> None:
|
||||
header = bytearray([0x81]) # FIN + text opcode
|
||||
mask = os.urandom(4)
|
||||
n = len(payload)
|
||||
if n < 126:
|
||||
header.append(0x80 | n)
|
||||
elif n < 65536:
|
||||
header.append(0x80 | 126)
|
||||
header += struct.pack(">H", n)
|
||||
else:
|
||||
header.append(0x80 | 127)
|
||||
header += struct.pack(">Q", n)
|
||||
header += mask
|
||||
self._sock.sendall(bytes(header) + bytes(b ^ mask[i % 4] for i, b in enumerate(payload)))
|
||||
|
||||
def _recv_exact(self, n: int) -> bytes:
|
||||
while len(self._rbuf) < n:
|
||||
chunk = self._sock.recv(65536)
|
||||
if not chunk:
|
||||
raise ConnectionError("CDP socket closed")
|
||||
self._rbuf += chunk
|
||||
out, self._rbuf = self._rbuf[:n], self._rbuf[n:]
|
||||
return out
|
||||
|
||||
def _recv_message(self) -> str:
|
||||
data = b""
|
||||
while True:
|
||||
b0, b1 = self._recv_exact(2)
|
||||
fin = b0 & 0x80
|
||||
length = b1 & 0x7F
|
||||
if length == 126:
|
||||
length = struct.unpack(">H", self._recv_exact(2))[0]
|
||||
elif length == 127:
|
||||
length = struct.unpack(">Q", self._recv_exact(8))[0]
|
||||
data += self._recv_exact(length)
|
||||
if fin:
|
||||
return data.decode("utf-8", "replace")
|
||||
|
||||
def cmd(self, method: str, params: dict[str, Any] | None = None, timeout: float = 15) -> Any:
|
||||
self._id += 1
|
||||
mid = self._id
|
||||
self._send(json.dumps({"id": mid, "method": method, "params": params or {}}).encode())
|
||||
deadline = time.monotonic() + timeout
|
||||
while time.monotonic() < deadline:
|
||||
self._sock.settimeout(max(0.1, deadline - time.monotonic()))
|
||||
obj = json.loads(self._recv_message())
|
||||
if obj.get("id") == mid:
|
||||
if "error" in obj:
|
||||
raise RuntimeError(f"{method}: {obj['error']}")
|
||||
return obj.get("result", {})
|
||||
raise TimeoutError(method)
|
||||
|
||||
def evaluate(self, expression: str) -> Any:
|
||||
r = self.cmd(
|
||||
"Runtime.evaluate",
|
||||
{"expression": expression, "returnByValue": True, "awaitPromise": True},
|
||||
)
|
||||
return r.get("result", {}).get("value")
|
||||
|
||||
def title(self) -> str:
|
||||
return str(self.evaluate("document.title") or "")
|
||||
|
||||
def close(self) -> None:
|
||||
with contextlib.suppress(Exception):
|
||||
self._sock.close()
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Chrome launch + node boot
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
def _find_chrome() -> str | None:
|
||||
for name in ("google-chrome", "google-chrome-stable", "chromium", "chromium-browser"):
|
||||
p = shutil.which(name)
|
||||
if p:
|
||||
return p
|
||||
return None
|
||||
|
||||
|
||||
def _free_port() -> int:
|
||||
with socket.socket() as s:
|
||||
s.bind(("127.0.0.1", 0))
|
||||
return int(s.getsockname()[1])
|
||||
|
||||
|
||||
def _launch_chrome(chrome: str, profile: Path) -> tuple[subprocess.Popen[bytes], int]:
|
||||
cdp_port = _free_port()
|
||||
proc = subprocess.Popen(
|
||||
[
|
||||
chrome,
|
||||
"--headless=new",
|
||||
"--disable-gpu",
|
||||
"--no-sandbox",
|
||||
"--no-first-run",
|
||||
"--disable-extensions",
|
||||
"--disable-background-timer-throttling",
|
||||
f"--remote-debugging-port={cdp_port}",
|
||||
f"--user-data-dir={profile}",
|
||||
"about:blank",
|
||||
],
|
||||
stdout=subprocess.DEVNULL,
|
||||
stderr=subprocess.DEVNULL,
|
||||
)
|
||||
return proc, cdp_port
|
||||
|
||||
|
||||
def _page_ws_url(cdp_port: int, timeout: float = 15) -> str:
|
||||
deadline = time.monotonic() + timeout
|
||||
while time.monotonic() < deadline:
|
||||
try:
|
||||
with urllib.request.urlopen(f"http://127.0.0.1:{cdp_port}/json", timeout=2) as r:
|
||||
targets = json.loads(r.read())
|
||||
for t in targets:
|
||||
if t.get("type") == "page" and t.get("webSocketDebuggerUrl"):
|
||||
return str(t["webSocketDebuggerUrl"])
|
||||
except Exception:
|
||||
pass
|
||||
time.sleep(0.2)
|
||||
raise TimeoutError("no CDP page target")
|
||||
|
||||
|
||||
def _page_route() -> Any:
|
||||
from starlette.responses import HTMLResponse
|
||||
from starlette.routing import Route
|
||||
|
||||
async def recovery_page(_request: Any) -> HTMLResponse:
|
||||
return HTMLResponse(PAGE_HTML)
|
||||
|
||||
return Route("/recovery", recovery_page)
|
||||
|
||||
|
||||
def _coord_routes() -> list[Any]:
|
||||
"""The coordinator scenario's same-origin extras: the pane page, and the
|
||||
console's coordinator static tree under the ``/coord-static`` prefix —
|
||||
a DISTINCT prefix because the node's own ``/static`` mount (ui/static)
|
||||
matches first and would 404 the console path from inside its own tree.
|
||||
coordinator.js's module imports are all absolute ``/shared/*``, which
|
||||
the node already serves."""
|
||||
from starlette.responses import HTMLResponse
|
||||
from starlette.routing import Mount, Route
|
||||
from starlette.staticfiles import StaticFiles
|
||||
|
||||
import turnstone
|
||||
|
||||
coord_dir = Path(turnstone.__file__).resolve().parent / "console" / "static" / "coordinator"
|
||||
|
||||
async def coord_recovery_page(_request: Any) -> HTMLResponse:
|
||||
return HTMLResponse(COORD_PAGE_HTML)
|
||||
|
||||
return [
|
||||
Route("/coord-recovery", coord_recovery_page),
|
||||
Mount("/coord-static", app=StaticFiles(directory=str(coord_dir)), name="coord-static"),
|
||||
]
|
||||
|
||||
|
||||
def _boot_node(port: int = 0) -> Any:
|
||||
from tests._sse_recovery_server import RecoveryServer
|
||||
from turnstone.core.storage import init_storage, reset_storage
|
||||
|
||||
# The page route must bypass auth on first load (the cookie is set by the
|
||||
# runner via CDP BEFORE navigation), so make it public by prefixing under
|
||||
# a public path is unavailable here; instead the runner sets the cookie so
|
||||
# /recovery passes the middleware. init storage per boot (shared singleton).
|
||||
reset_storage()
|
||||
init_storage("sqlite", path=os.path.join(_scratch(), "recovery_e2e.db"), run_migrations=True)
|
||||
return RecoveryServer(extra_routes=[_page_route(), *_coord_routes()], port=port)
|
||||
|
||||
|
||||
def _scratch() -> str:
|
||||
d = os.environ.get("RECOVERY_E2E_TMP") or "/tmp/recovery_e2e"
|
||||
os.makedirs(d, exist_ok=True)
|
||||
return d
|
||||
|
||||
|
||||
def _set_cookie_and_navigate(cdp: CDP, base_url: str, token: str, page_url: str) -> None:
|
||||
cdp.cmd("Page.enable")
|
||||
cdp.cmd("Runtime.enable")
|
||||
cdp.cmd("Network.enable")
|
||||
cdp.cmd(
|
||||
"Network.setCookie",
|
||||
{
|
||||
"name": "turnstone_auth_server",
|
||||
"value": token,
|
||||
"url": base_url,
|
||||
"path": "/",
|
||||
},
|
||||
)
|
||||
cdp.cmd("Page.navigate", {"url": page_url})
|
||||
|
||||
|
||||
def _poll_title(cdp: CDP, timeout: float) -> str:
|
||||
deadline = time.monotonic() + timeout
|
||||
last = ""
|
||||
while time.monotonic() < deadline:
|
||||
last = cdp.title()
|
||||
if last.startswith("RECOVERY-"):
|
||||
return last
|
||||
time.sleep(0.3)
|
||||
return last or "RECOVERY-FAILED-timeout"
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Scenarios
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
def _storm_scripts() -> tuple[Any, ...]:
|
||||
"""A parallel bash storm PLUS a task_agent whose sub-tools are chatty
|
||||
bashes (so the browser proves both fix-3 batching AND sub-agent nesting
|
||||
with no escaped children)."""
|
||||
from tests._sse_recovery_server import final_text_script, parallel_bash_script
|
||||
|
||||
storm = parallel_bash_script({f"call_{i}": "seq 1 500" for i in range(4)})
|
||||
task = dict(
|
||||
tool_calls=[
|
||||
{
|
||||
"id": "task1",
|
||||
"name": "task_agent",
|
||||
"arguments": json.dumps({"prompt": "sub tools"}),
|
||||
}
|
||||
],
|
||||
finish_reason="tool_calls",
|
||||
)
|
||||
sub = dict(
|
||||
tool_calls=[
|
||||
{"id": "s_a", "name": "bash", "arguments": json.dumps({"command": ": a; seq 1 200"})},
|
||||
{"id": "s_b", "name": "bash", "arguments": json.dumps({"command": ": b; seq 1 200"})},
|
||||
],
|
||||
finish_reason="tool_calls",
|
||||
)
|
||||
# Turn 1: the 4-bash storm; turn 2: a task_agent with 2 chatty sub-bashes.
|
||||
return (
|
||||
storm,
|
||||
final_text_script("storm done"),
|
||||
task,
|
||||
sub,
|
||||
final_text_script("sub done"),
|
||||
final_text_script("all done"),
|
||||
)
|
||||
|
||||
|
||||
def run_storm(chrome: str) -> str:
|
||||
node = _boot_node()
|
||||
ws_id = node.create_workstream(*_storm_scripts(), name="browser-storm")
|
||||
profile = Path(_scratch()) / "chrome-storm"
|
||||
proc, cdp_port = _launch_chrome(chrome, profile)
|
||||
cdp: CDP | None = None
|
||||
try:
|
||||
cdp = CDP(_page_ws_url(cdp_port))
|
||||
# The page POSTs the STORM turn on stream-open; the runner sends the
|
||||
# task_agent follow-up once the first turn settles so both land.
|
||||
url = f"{node.base_url}/recovery?ws_id={ws_id}&scenario=storm&rows=4"
|
||||
_set_cookie_and_navigate(cdp, node.base_url, node.token, url)
|
||||
# After the storm turn, trigger the task_agent turn via REST so the
|
||||
# page renders the nested sub-agent card.
|
||||
_wait_state(node, ws_id, "idle", 40)
|
||||
node.send(ws_id, "spawn the sub agent")
|
||||
return _poll_title(cdp, 45)
|
||||
finally:
|
||||
if cdp is not None:
|
||||
cdp.close()
|
||||
_kill(proc)
|
||||
node.stop()
|
||||
|
||||
|
||||
def run_restart(chrome: str) -> str:
|
||||
from tests._sse_recovery_server import final_text_script, parallel_bash_script
|
||||
|
||||
port = _free_port()
|
||||
node = _boot_node(port=port)
|
||||
# A PACED turn so the tab can hide MID-turn: the browser cursor
|
||||
# freezes at a mid-stream event id, the rest of turn 1 plus the
|
||||
# turn-2 text commit while hidden, and the restarted node's seeded
|
||||
# counter therefore sits ABOVE the frozen cursor -> the show-edge
|
||||
# reconnect draws ``replay_truncated`` and must heal the gap.
|
||||
paced = parallel_bash_script({"r0": "for i in $(seq 1 40); do echo r-$i; sleep 0.05; done"})
|
||||
# The turn-2 text is the healed-gap sentinel — it must be a token
|
||||
# that cannot appear in any rendered command/output (the bash
|
||||
# command row contains the shell keyword ``done``, so the obvious
|
||||
# word is vacuously present; see __verifyRestart). Single source:
|
||||
# the same constant is injected as the scripted turn text AND
|
||||
# threaded to the page via ?healed=, so the two sides cannot drift.
|
||||
ws_id = node.create_workstream(
|
||||
paced, final_text_script(HEALED_SENTINEL), name="browser-restart"
|
||||
)
|
||||
profile = Path(_scratch()) / "chrome-restart"
|
||||
proc, cdp_port = _launch_chrome(chrome, profile)
|
||||
cdp: CDP | None = None
|
||||
try:
|
||||
cdp = CDP(_page_ws_url(cdp_port))
|
||||
url = f"{node.base_url}/recovery?ws_id={ws_id}&scenario=restart&healed={HEALED_SENTINEL}"
|
||||
_set_cookie_and_navigate(cdp, node.base_url, node.token, url)
|
||||
# Hide as soon as the FIRST streamed line has painted (proof the
|
||||
# pane holds a live mid-turn cursor) — NOT after wait_turn, which
|
||||
# would leave the cursor at/above the committed counter and the
|
||||
# reconnect on the lossless replay_ok path (trunc0).
|
||||
deadline = time.monotonic() + 15
|
||||
while time.monotonic() < deadline:
|
||||
painted = cdp.evaluate("document.querySelector('.tool-output-stream') !== null")
|
||||
if painted:
|
||||
break
|
||||
time.sleep(0.2)
|
||||
else:
|
||||
raise AssertionError("restart scenario: first streamed line never painted")
|
||||
cdp.evaluate("window.__hide && window.__hide()")
|
||||
# The turn (and the follow-up text) commits while the tab is hidden.
|
||||
node.wait_turn(ws_id, timeout=30)
|
||||
# Restart the node on the SAME port (fresh empty ring, seeded counter).
|
||||
node.stop()
|
||||
node = _boot_node(port=port)
|
||||
node.open_workstream(ws_id)
|
||||
# Show the tab -> stale-cursor reconnect -> truncated -> jittered
|
||||
# resync (0-10s) -> /history rebuild. Settle past the worst-case
|
||||
# jitter before the verdict.
|
||||
cdp.evaluate("window.__show && window.__show()")
|
||||
time.sleep(12.0)
|
||||
cdp.evaluate("window.__verifyRestart && window.__verifyRestart()")
|
||||
return _poll_title(cdp, 20)
|
||||
finally:
|
||||
if cdp is not None:
|
||||
cdp.close()
|
||||
_kill(proc)
|
||||
node.stop()
|
||||
|
||||
|
||||
def run_coord_restart(chrome: str) -> str:
|
||||
from tests._sse_recovery_server import final_text_script, parallel_bash_script
|
||||
|
||||
port = _free_port()
|
||||
node = _boot_node(port=port)
|
||||
# Same shape as Scenario B: a PACED turn so the tab can hide MID-turn.
|
||||
# The coordinator renders no streamed tool output (no tool_output_chunk
|
||||
# case), but the chunk frames still advance the pane's cursor in
|
||||
# onmessage BEFORE dispatch — so the hide freezes a genuinely mid-turn
|
||||
# cursor even though the paint signal differs (see below). The closing
|
||||
# assistant text after the bash is the healed-gap sentinel, committed
|
||||
# while hidden.
|
||||
paced = parallel_bash_script({"c0": "for i in $(seq 1 40); do echo c-$i; sleep 0.05; done"})
|
||||
ws_id = node.create_workstream(
|
||||
paced, final_text_script(HEALED_SENTINEL), name="browser-coord-restart"
|
||||
)
|
||||
profile = Path(_scratch()) / "chrome-coord-restart"
|
||||
proc, cdp_port = _launch_chrome(chrome, profile)
|
||||
cdp: CDP | None = None
|
||||
try:
|
||||
cdp = CDP(_page_ws_url(cdp_port))
|
||||
url = f"{node.base_url}/coord-recovery?ws_id={ws_id}&healed={HEALED_SENTINEL}"
|
||||
_set_cookie_and_navigate(cdp, node.base_url, node.token, url)
|
||||
# Hide once the BROWSER has captured a live mid-turn cursor: the
|
||||
# coordinator chrome has no status/SSE text elements (the header was
|
||||
# deliberately dropped) and paints no streamed output line, so the
|
||||
# signal is transport-level — id-bearing frames received by the page
|
||||
# (__idFrames; exactly the frames that advance the pane's reconnect
|
||||
# cursor). The turn must also still be RUNNING server-side, or the
|
||||
# frozen cursor could sit at/above the committed counter and the
|
||||
# reconnect would take the lossless replay_ok path (trunc0). The
|
||||
# paced bash runs >=2s, so this lands mid-turn.
|
||||
deadline = time.monotonic() + 15
|
||||
while time.monotonic() < deadline:
|
||||
frames = cdp.evaluate("window.__idFrames || 0")
|
||||
if isinstance(frames, int) and frames >= 2 and node.ws_state(ws_id) == "running":
|
||||
break
|
||||
time.sleep(0.2)
|
||||
else:
|
||||
raise AssertionError(
|
||||
"coord-restart scenario: no mid-turn cursor captured "
|
||||
"(id frames never reached the page while running)"
|
||||
)
|
||||
time.sleep(0.5)
|
||||
cdp.evaluate("window.__hide && window.__hide()")
|
||||
# The turn (and the sentinel closing text) commits while hidden.
|
||||
node.wait_turn(ws_id, timeout=30)
|
||||
# Restart the node on the SAME port (fresh empty ring, seeded counter).
|
||||
node.stop()
|
||||
node = _boot_node(port=port)
|
||||
node.open_workstream(ws_id)
|
||||
# Show the tab -> stale-cursor reconnect -> truncated -> jittered
|
||||
# resync (0-10s) -> /history rebuild. Settle past the worst-case
|
||||
# jitter, assert idle SERVER-side (the chrome has no state text to
|
||||
# read), then take the in-page verdict.
|
||||
cdp.evaluate("window.__show && window.__show()")
|
||||
time.sleep(12.0)
|
||||
_wait_state(node, ws_id, "idle", 15)
|
||||
cdp.evaluate("window.__verifyCoordRestart && window.__verifyCoordRestart()")
|
||||
return _poll_title(cdp, 20)
|
||||
finally:
|
||||
if cdp is not None:
|
||||
cdp.close()
|
||||
_kill(proc)
|
||||
node.stop()
|
||||
|
||||
|
||||
def _wait_state(node: Any, ws_id: str, state: str, timeout: float) -> None:
|
||||
deadline = time.monotonic() + timeout
|
||||
while time.monotonic() < deadline:
|
||||
if node.ws_state(ws_id) == state:
|
||||
return
|
||||
time.sleep(0.1)
|
||||
|
||||
|
||||
def _kill(proc: subprocess.Popen[bytes]) -> None:
|
||||
if proc.poll() is None:
|
||||
proc.terminate()
|
||||
try:
|
||||
proc.wait(8)
|
||||
except subprocess.TimeoutExpired:
|
||||
proc.kill()
|
||||
|
||||
|
||||
def keep_open(port: int) -> None:
|
||||
"""Boot the node + storm ws and serve the page for manual inspection."""
|
||||
node = _boot_node(port=port)
|
||||
ws_id = node.create_workstream(*_storm_scripts(), name="manual-storm")
|
||||
print(f"node: {node.base_url}")
|
||||
print(f"cookie: turnstone_auth_server={node.token}")
|
||||
print(f"page: {node.base_url}/recovery?ws_id={ws_id}&scenario=storm&rows=4")
|
||||
print("set the cookie for this origin, then open the page. Ctrl+C to stop.")
|
||||
try:
|
||||
while True:
|
||||
time.sleep(1)
|
||||
except KeyboardInterrupt:
|
||||
node.stop()
|
||||
|
||||
|
||||
def main() -> None:
|
||||
ap = argparse.ArgumentParser(description=__doc__.splitlines()[0])
|
||||
ap.add_argument(
|
||||
"--scenario",
|
||||
choices=["storm", "restart", "coord-restart", "both", "all"],
|
||||
default="all",
|
||||
help="'both' = A+B (legacy alias); 'all' adds the coordinator scenario",
|
||||
)
|
||||
ap.add_argument("--keep-open", type=int, metavar="PORT", help="serve the storm page, no CDP")
|
||||
args = ap.parse_args()
|
||||
|
||||
if args.keep_open:
|
||||
keep_open(args.keep_open)
|
||||
return
|
||||
|
||||
chrome = _find_chrome()
|
||||
if chrome is None:
|
||||
print("recovery_e2e: no chrome/chromium on PATH — see the module docstring runbook")
|
||||
raise SystemExit(2)
|
||||
|
||||
failures = 0
|
||||
if args.scenario in ("storm", "both", "all"):
|
||||
verdict = run_storm(chrome)
|
||||
print(f"scenario A (storm): {verdict}")
|
||||
failures += 0 if verdict.startswith("RECOVERY-READY") else 1
|
||||
if args.scenario in ("restart", "both", "all"):
|
||||
verdict = run_restart(chrome)
|
||||
print(f"scenario B (restart): {verdict}")
|
||||
failures += 0 if verdict.startswith("RECOVERY-READY") else 1
|
||||
if args.scenario in ("coord-restart", "all"):
|
||||
verdict = run_coord_restart(chrome)
|
||||
print(f"scenario C (coord): {verdict}")
|
||||
failures += 0 if verdict.startswith("RECOVERY-READY") else 1
|
||||
raise SystemExit(1 if failures else 0)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -2,7 +2,7 @@
|
||||
"openapi": "3.1.0",
|
||||
"info": {
|
||||
"title": "turnstone Console API",
|
||||
"version": "1.8.0a2",
|
||||
"version": "1.7.0a2",
|
||||
"description": "Cluster-wide visibility and control across all turnstone nodes."
|
||||
},
|
||||
"paths": {
|
||||
@@ -4213,166 +4213,6 @@
|
||||
}
|
||||
}
|
||||
},
|
||||
"/v1/api/admin/personas": {
|
||||
"get": {
|
||||
"summary": "List all personas, archived included",
|
||||
"operationId": "v1_api_admin_personas_get",
|
||||
"tags": [
|
||||
"Admin"
|
||||
],
|
||||
"responses": {
|
||||
"200": {
|
||||
"description": "Success",
|
||||
"content": {
|
||||
"application/json": {
|
||||
"schema": {
|
||||
"$ref": "#/components/schemas/ListPersonasResponse"
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
"post": {
|
||||
"summary": "Create a persona",
|
||||
"operationId": "v1_api_admin_personas_post",
|
||||
"tags": [
|
||||
"Admin"
|
||||
],
|
||||
"requestBody": {
|
||||
"required": true,
|
||||
"content": {
|
||||
"application/json": {
|
||||
"schema": {
|
||||
"$ref": "#/components/schemas/CreatePersonaRequest"
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
"responses": {
|
||||
"200": {
|
||||
"description": "Success",
|
||||
"content": {
|
||||
"application/json": {
|
||||
"schema": {
|
||||
"$ref": "#/components/schemas/PersonaInfo"
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
"400": {
|
||||
"description": "Error 400",
|
||||
"content": {
|
||||
"application/json": {
|
||||
"schema": {
|
||||
"$ref": "#/components/schemas/ErrorResponse"
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
"/v1/api/admin/personas/{persona_id}": {
|
||||
"get": {
|
||||
"summary": "Get a single persona",
|
||||
"operationId": "v1_api_admin_personas_{persona_id}_get",
|
||||
"tags": [
|
||||
"Admin"
|
||||
],
|
||||
"parameters": [
|
||||
{
|
||||
"name": "persona_id",
|
||||
"in": "path",
|
||||
"required": true,
|
||||
"schema": {
|
||||
"type": "string"
|
||||
}
|
||||
}
|
||||
],
|
||||
"responses": {
|
||||
"200": {
|
||||
"description": "Success",
|
||||
"content": {
|
||||
"application/json": {
|
||||
"schema": {
|
||||
"$ref": "#/components/schemas/PersonaInfo"
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
"404": {
|
||||
"description": "Error 404",
|
||||
"content": {
|
||||
"application/json": {
|
||||
"schema": {
|
||||
"$ref": "#/components/schemas/ErrorResponse"
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
"patch": {
|
||||
"summary": "Update a persona (edit levers, archive/unarchive, flip default)",
|
||||
"operationId": "v1_api_admin_personas_{persona_id}_patch",
|
||||
"tags": [
|
||||
"Admin"
|
||||
],
|
||||
"parameters": [
|
||||
{
|
||||
"name": "persona_id",
|
||||
"in": "path",
|
||||
"required": true,
|
||||
"schema": {
|
||||
"type": "string"
|
||||
}
|
||||
}
|
||||
],
|
||||
"requestBody": {
|
||||
"required": true,
|
||||
"content": {
|
||||
"application/json": {
|
||||
"schema": {
|
||||
"$ref": "#/components/schemas/UpdatePersonaRequest"
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
"responses": {
|
||||
"200": {
|
||||
"description": "Success",
|
||||
"content": {
|
||||
"application/json": {
|
||||
"schema": {
|
||||
"$ref": "#/components/schemas/PersonaInfo"
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
"400": {
|
||||
"description": "Error 400",
|
||||
"content": {
|
||||
"application/json": {
|
||||
"schema": {
|
||||
"$ref": "#/components/schemas/ErrorResponse"
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
"404": {
|
||||
"description": "Error 404",
|
||||
"content": {
|
||||
"application/json": {
|
||||
"schema": {
|
||||
"$ref": "#/components/schemas/ErrorResponse"
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
"/v1/api/admin/node-metadata": {
|
||||
"get": {
|
||||
"summary": "Get metadata for all nodes (bulk)",
|
||||
@@ -6688,7 +6528,7 @@
|
||||
"tags": [
|
||||
"Coordinator"
|
||||
],
|
||||
"description": "Aggregates the persisted row, a best-effort live block from the owning node (or the in-process coordinator manager for ``kind=\"coordinator\"`` rows), and the tail of the message history. Gated on the ``admin.cluster.inspect`` permission (granted to ``builtin-admin`` via migration 040; revoke or reassign to a custom role for tighter control). A workstream attached to a *private* project stays confidential to its members: a permitted caller who isn't its owner / creator / project member gets a 404 (same masking as an unknown id). ``live`` is null on node unreachability / 5xx so callers can degrade gracefully.",
|
||||
"description": "Aggregates the persisted row, a best-effort live block from the owning node (or the in-process coordinator manager for ``kind=\"coordinator\"`` rows), and the tail of the message history. Gated on the ``admin.cluster.inspect`` permission (granted to ``builtin-admin`` via migration 040; revoke or reassign to a custom role for tighter control). ``live`` is null on node unreachability / 5xx so callers can degrade gracefully.",
|
||||
"parameters": [
|
||||
{
|
||||
"name": "ws_id",
|
||||
@@ -7785,18 +7625,6 @@
|
||||
"title": "Skill",
|
||||
"type": "string"
|
||||
},
|
||||
"persona": {
|
||||
"default": "",
|
||||
"description": "Persona slug; resolved and snapshotted at creation, empty = kind default",
|
||||
"title": "Persona",
|
||||
"type": "string"
|
||||
},
|
||||
"project_id": {
|
||||
"default": "",
|
||||
"description": "Project to attach the workstream to (validated against membership, empty = none)",
|
||||
"title": "Project Id",
|
||||
"type": "string"
|
||||
},
|
||||
"resume_ws": {
|
||||
"default": "",
|
||||
"description": "Workstream ID to resume (loads previous conversation)",
|
||||
@@ -8124,12 +7952,6 @@
|
||||
"description": "Optional skill name to apply to the coordinator session.",
|
||||
"title": "Skill"
|
||||
},
|
||||
"persona": {
|
||||
"default": "",
|
||||
"description": "Persona slug; resolved and snapshotted at creation, empty = kind default",
|
||||
"title": "Persona",
|
||||
"type": "string"
|
||||
},
|
||||
"initial_message": {
|
||||
"default": "",
|
||||
"description": "Optional first user message dispatched to the new coordinator session.",
|
||||
@@ -8502,18 +8324,6 @@
|
||||
"title": "Skill",
|
||||
"type": "string"
|
||||
},
|
||||
"persona": {
|
||||
"default": "",
|
||||
"description": "Persona slug (empty = kind default)",
|
||||
"title": "Persona",
|
||||
"type": "string"
|
||||
},
|
||||
"project_id": {
|
||||
"default": "",
|
||||
"description": "Project to attach the workstream to",
|
||||
"title": "Project Id",
|
||||
"type": "string"
|
||||
},
|
||||
"notify_targets": {
|
||||
"description": "Notification targets on completion (channel_type + channel_id/user_id)",
|
||||
"items": {
|
||||
@@ -8677,30 +8487,6 @@
|
||||
"default": null,
|
||||
"title": "Skill"
|
||||
},
|
||||
"persona": {
|
||||
"anyOf": [
|
||||
{
|
||||
"type": "string"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"title": "Persona"
|
||||
},
|
||||
"project_id": {
|
||||
"anyOf": [
|
||||
{
|
||||
"type": "string"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"title": "Project Id"
|
||||
},
|
||||
"notify_targets": {
|
||||
"anyOf": [
|
||||
{
|
||||
@@ -8796,16 +8582,6 @@
|
||||
"title": "Skill",
|
||||
"type": "string"
|
||||
},
|
||||
"persona": {
|
||||
"default": "",
|
||||
"title": "Persona",
|
||||
"type": "string"
|
||||
},
|
||||
"project_id": {
|
||||
"default": "",
|
||||
"title": "Project Id",
|
||||
"type": "string"
|
||||
},
|
||||
"notify_targets": {
|
||||
"items": {
|
||||
"additionalProperties": {
|
||||
@@ -11120,345 +10896,6 @@
|
||||
"title": "ListModelDefinitionsResponse",
|
||||
"type": "object"
|
||||
},
|
||||
"PersonaInfo": {
|
||||
"description": "Full persona row \u2014 the authoring shape (contrast PersonaChoice, the\npicker's display-only projection on the server surface).",
|
||||
"properties": {
|
||||
"persona_id": {
|
||||
"title": "Persona Id",
|
||||
"type": "string"
|
||||
},
|
||||
"name": {
|
||||
"title": "Name",
|
||||
"type": "string"
|
||||
},
|
||||
"display_name": {
|
||||
"default": "",
|
||||
"title": "Display Name",
|
||||
"type": "string"
|
||||
},
|
||||
"description": {
|
||||
"default": "",
|
||||
"title": "Description",
|
||||
"type": "string"
|
||||
},
|
||||
"base_prompt": {
|
||||
"anyOf": [
|
||||
{
|
||||
"type": "string"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"description": "BASE-module override; null = the kind's stock base",
|
||||
"title": "Base Prompt"
|
||||
},
|
||||
"tool_allowlist": {
|
||||
"anyOf": [
|
||||
{
|
||||
"items": {
|
||||
"type": "string"
|
||||
},
|
||||
"type": "array"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"description": "Tool visibility set: null = unrestricted, [] = no tools, [names] = exact set (include 'tool_search' to keep the set soft/expandable)",
|
||||
"title": "Tool Allowlist"
|
||||
},
|
||||
"mcp_enabled": {
|
||||
"default": true,
|
||||
"title": "Mcp Enabled",
|
||||
"type": "boolean"
|
||||
},
|
||||
"memory_enabled": {
|
||||
"default": true,
|
||||
"title": "Memory Enabled",
|
||||
"type": "boolean"
|
||||
},
|
||||
"applies_to_kinds": {
|
||||
"items": {
|
||||
"type": "string"
|
||||
},
|
||||
"title": "Applies To Kinds",
|
||||
"type": "array"
|
||||
},
|
||||
"is_default": {
|
||||
"default": false,
|
||||
"title": "Is Default",
|
||||
"type": "boolean"
|
||||
},
|
||||
"enabled": {
|
||||
"default": true,
|
||||
"description": "false = archived",
|
||||
"title": "Enabled",
|
||||
"type": "boolean"
|
||||
},
|
||||
"org_id": {
|
||||
"default": "",
|
||||
"title": "Org Id",
|
||||
"type": "string"
|
||||
},
|
||||
"created_by": {
|
||||
"default": "",
|
||||
"title": "Created By",
|
||||
"type": "string"
|
||||
},
|
||||
"created": {
|
||||
"default": "",
|
||||
"title": "Created",
|
||||
"type": "string"
|
||||
},
|
||||
"updated": {
|
||||
"default": "",
|
||||
"title": "Updated",
|
||||
"type": "string"
|
||||
}
|
||||
},
|
||||
"required": [
|
||||
"persona_id",
|
||||
"name"
|
||||
],
|
||||
"title": "PersonaInfo",
|
||||
"type": "object"
|
||||
},
|
||||
"CreatePersonaRequest": {
|
||||
"properties": {
|
||||
"name": {
|
||||
"description": "Immutable slug (lowercase: a-z, 0-9, '-', '_')",
|
||||
"title": "Name",
|
||||
"type": "string"
|
||||
},
|
||||
"display_name": {
|
||||
"default": "",
|
||||
"title": "Display Name",
|
||||
"type": "string"
|
||||
},
|
||||
"description": {
|
||||
"default": "",
|
||||
"title": "Description",
|
||||
"type": "string"
|
||||
},
|
||||
"base_prompt": {
|
||||
"anyOf": [
|
||||
{
|
||||
"type": "string"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"description": "Inline BASE override \u2014 required. Every persona must name a prompt source; built-in file-backed personas are seeded by migration, not created here, so an operator-created persona must supply base_prompt.",
|
||||
"title": "Base Prompt"
|
||||
},
|
||||
"tool_allowlist": {
|
||||
"anyOf": [
|
||||
{
|
||||
"items": {
|
||||
"type": "string"
|
||||
},
|
||||
"type": "array"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"title": "Tool Allowlist"
|
||||
},
|
||||
"mcp_enabled": {
|
||||
"default": true,
|
||||
"title": "Mcp Enabled",
|
||||
"type": "boolean"
|
||||
},
|
||||
"memory_enabled": {
|
||||
"default": true,
|
||||
"title": "Memory Enabled",
|
||||
"type": "boolean"
|
||||
},
|
||||
"applies_to_kinds": {
|
||||
"items": {
|
||||
"type": "string"
|
||||
},
|
||||
"title": "Applies To Kinds",
|
||||
"type": "array"
|
||||
},
|
||||
"is_default": {
|
||||
"default": false,
|
||||
"title": "Is Default",
|
||||
"type": "boolean"
|
||||
},
|
||||
"enabled": {
|
||||
"default": true,
|
||||
"title": "Enabled",
|
||||
"type": "boolean"
|
||||
},
|
||||
"org_id": {
|
||||
"default": "",
|
||||
"description": "Owning org (informational; capped at 64)",
|
||||
"title": "Org Id",
|
||||
"type": "string"
|
||||
}
|
||||
},
|
||||
"required": [
|
||||
"name"
|
||||
],
|
||||
"title": "CreatePersonaRequest",
|
||||
"type": "object"
|
||||
},
|
||||
"UpdatePersonaRequest": {
|
||||
"description": "PATCH body \u2014 absent fields are left unchanged.\n\nExplicit ``null`` resets ``tool_allowlist`` to unrestricted, and \u2014 on a\nBUILT-IN persona only \u2014 clears ``base_prompt`` (the operator override),\nreverting to that persona's file-backed prompt. An OPERATOR persona has no\nfallback source, so ``base_prompt: null`` on one is rejected: every persona\nmust name a prompt source. ``null`` on the boolean flags or\n``applies_to_kinds`` is ignored (treated as absent), so a client serializing\nunset optionals as null cannot archive a persona or flip levers by accident.\n\nArchive = ``{\"enabled\": false}``; default flip = ``{\"is_default\": true}``\non the successor (storage demotes the incumbent atomically). ``name``\nis immutable; existing workstreams are never affected by edits.",
|
||||
"properties": {
|
||||
"display_name": {
|
||||
"anyOf": [
|
||||
{
|
||||
"type": "string"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"title": "Display Name"
|
||||
},
|
||||
"description": {
|
||||
"anyOf": [
|
||||
{
|
||||
"type": "string"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"title": "Description"
|
||||
},
|
||||
"base_prompt": {
|
||||
"anyOf": [
|
||||
{
|
||||
"type": "string"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"title": "Base Prompt"
|
||||
},
|
||||
"tool_allowlist": {
|
||||
"anyOf": [
|
||||
{
|
||||
"items": {
|
||||
"type": "string"
|
||||
},
|
||||
"type": "array"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"title": "Tool Allowlist"
|
||||
},
|
||||
"mcp_enabled": {
|
||||
"anyOf": [
|
||||
{
|
||||
"type": "boolean"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"title": "Mcp Enabled"
|
||||
},
|
||||
"memory_enabled": {
|
||||
"anyOf": [
|
||||
{
|
||||
"type": "boolean"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"title": "Memory Enabled"
|
||||
},
|
||||
"applies_to_kinds": {
|
||||
"anyOf": [
|
||||
{
|
||||
"items": {
|
||||
"type": "string"
|
||||
},
|
||||
"type": "array"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"title": "Applies To Kinds"
|
||||
},
|
||||
"is_default": {
|
||||
"anyOf": [
|
||||
{
|
||||
"type": "boolean"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"title": "Is Default"
|
||||
},
|
||||
"enabled": {
|
||||
"anyOf": [
|
||||
{
|
||||
"type": "boolean"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"title": "Enabled"
|
||||
}
|
||||
},
|
||||
"title": "UpdatePersonaRequest",
|
||||
"type": "object"
|
||||
},
|
||||
"ListPersonasResponse": {
|
||||
"properties": {
|
||||
"personas": {
|
||||
"items": {
|
||||
"$ref": "#/components/schemas/PersonaInfo"
|
||||
},
|
||||
"title": "Personas",
|
||||
"type": "array"
|
||||
},
|
||||
"tool_inventory": {
|
||||
"additionalProperties": {
|
||||
"items": {
|
||||
"type": "string"
|
||||
},
|
||||
"type": "array"
|
||||
},
|
||||
"description": "Per-kind builtin tool names (plus the synthetic 'tool_search') for the visibility checklist \u2014 derived server-side so clients never hand-mirror the inventory",
|
||||
"title": "Tool Inventory",
|
||||
"type": "object"
|
||||
}
|
||||
},
|
||||
"required": [
|
||||
"personas"
|
||||
],
|
||||
"title": "ListPersonasResponse",
|
||||
"type": "object"
|
||||
},
|
||||
"ModelReloadResponse": {
|
||||
"properties": {
|
||||
"status": {
|
||||
@@ -13358,17 +12795,21 @@
|
||||
},
|
||||
"pending_approval": {
|
||||
"default": false,
|
||||
"description": "True when at least one approval cycle is live (a gate thread parked awaiting an operator approve/deny). Mirrors the same field on ``DashboardWorkstream`` / cluster live projections so a freshly-loaded chat tab can render the inline approval gate from the detail snapshot before SSE replay arrives.",
|
||||
"description": "True when the workstream is parked on ``_approval_event`` awaiting an operator approve/deny. Mirrors the same field on ``DashboardWorkstream`` / cluster live projections so a freshly-loaded chat tab can render the inline approval gate from the detail snapshot before SSE replay arrives.",
|
||||
"title": "Pending Approval",
|
||||
"type": "boolean"
|
||||
},
|
||||
"pending_approval_details": {
|
||||
"description": "Inline approval payloads, one per live cycle, oldest first \u2014 same shape as ``DashboardWorkstream.pending_approval_details``. Empty when no approval is pending. Lets a reload paint every action row + judge verdicts immediately instead of relying on the SSE approve_request replay timing window. Replaces 1.6's ``pending_approval_detail`` single-object field (breaking, 1.7).",
|
||||
"items": {
|
||||
"$ref": "#/components/schemas/PendingApprovalDetail"
|
||||
},
|
||||
"title": "Pending Approval Details",
|
||||
"type": "array"
|
||||
"pending_approval_detail": {
|
||||
"anyOf": [
|
||||
{
|
||||
"$ref": "#/components/schemas/PendingApprovalDetail"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"description": "Inline approval payload \u2014 same shape as ``DashboardWorkstream.pending_approval_detail``. ``None`` when no approval is pending. Lets a reload paint the action row + judge verdicts immediately instead of relying on the SSE approve_request replay timing window."
|
||||
}
|
||||
},
|
||||
"required": [
|
||||
@@ -13381,14 +12822,8 @@
|
||||
"type": "object"
|
||||
},
|
||||
"PendingApprovalDetail": {
|
||||
"description": "Inline approval payload merged into ``DashboardWorkstream``.\n\nOne entry per live approval CYCLE \u2014 a gate thread parked in\n``approve_tools`` awaiting the operator. Parallel task agents run\nconcurrent gates, so a workstream can have several of these at\nonce (``pending_approval_details``, oldest first). Cross-tenant\nexposure here follows the same trusted-team posture as\n``activity`` / ``tokens`` \u2014 see ``server.py``'s ``dashboard``\nhandler comment.",
|
||||
"description": "Inline approval payload merged into ``DashboardWorkstream``.\n\nSet when a workstream's ``approve_tools`` is parked on\n``_approval_event``; ``None`` (omitted) otherwise. Cross-tenant\nexposure here follows the same trusted-team posture as\n``activity`` / ``tokens`` \u2014 see ``server.py``'s ``dashboard``\nhandler comment.",
|
||||
"properties": {
|
||||
"cycle_id": {
|
||||
"default": "",
|
||||
"description": "Identity of this approval cycle. Echo it back on ``POST /v1/api/workstreams/{ws_id}/approve`` to resolve exactly this round \u2014 required for correctness when several cycles are live (parallel task agents).",
|
||||
"title": "Cycle Id",
|
||||
"type": "string"
|
||||
},
|
||||
"call_id": {
|
||||
"default": "",
|
||||
"description": "Primary call_id \u2014 first non-empty call_id in items list order. Matches the 409 ``current_call_id`` response from ``POST /v1/api/workstreams/{ws_id}/approve`` so the UI can render the same identifier the server reports as current.",
|
||||
@@ -13413,7 +12848,7 @@
|
||||
"type": "object"
|
||||
},
|
||||
"PendingApprovalItem": {
|
||||
"description": "One pending tool-call inside a ``PendingApprovalDetail`` envelope.\n\nMirrors the dict ``SessionUIBase.serialize_pending_approval_details``\nemits per item inside each cycle entry. ``heuristic_verdict`` / ``judge_verdict`` are kept\nloosely-typed because the underlying verdict shape varies by tier;\nconsumers that want the full structure can decode against\n:class:`turnstone.sdk.events.IntentVerdictEvent`.",
|
||||
"description": "One pending tool-call inside a ``PendingApprovalDetail`` envelope.\n\nMirrors the dict ``SessionUIBase.serialize_pending_approval_detail``\nemits per item. ``heuristic_verdict`` / ``judge_verdict`` are kept\nloosely-typed because the underlying verdict shape varies by tier;\nconsumers that want the full structure can decode against\n:class:`turnstone.sdk.events.IntentVerdictEvent`.",
|
||||
"properties": {
|
||||
"call_id": {
|
||||
"default": "",
|
||||
|
||||
@@ -2,7 +2,7 @@
|
||||
"openapi": "3.1.0",
|
||||
"info": {
|
||||
"title": "turnstone Server API",
|
||||
"version": "1.8.0a2",
|
||||
"version": "1.7.0a2",
|
||||
"description": "Single-node workstream management, chat interaction, and real-time streaming."
|
||||
},
|
||||
"paths": {
|
||||
@@ -228,16 +228,6 @@
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
"409": {
|
||||
"description": "Error 409",
|
||||
"content": {
|
||||
"application/json": {
|
||||
"schema": {
|
||||
"$ref": "#/components/schemas/ErrorResponse"
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
@@ -400,26 +390,6 @@
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
"409": {
|
||||
"description": "Error 409",
|
||||
"content": {
|
||||
"application/json": {
|
||||
"schema": {
|
||||
"$ref": "#/components/schemas/ErrorResponse"
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
"503": {
|
||||
"description": "Error 503",
|
||||
"content": {
|
||||
"application/json": {
|
||||
"schema": {
|
||||
"$ref": "#/components/schemas/ErrorResponse"
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
@@ -1473,27 +1443,6 @@
|
||||
}
|
||||
}
|
||||
},
|
||||
"/v1/api/personas": {
|
||||
"get": {
|
||||
"summary": "List enabled personas for the workstream-creation picker",
|
||||
"operationId": "v1_api_personas_get",
|
||||
"tags": [
|
||||
"Personas"
|
||||
],
|
||||
"responses": {
|
||||
"200": {
|
||||
"description": "Success",
|
||||
"content": {
|
||||
"application/json": {
|
||||
"schema": {
|
||||
"$ref": "#/components/schemas/ListPersonaChoicesResponse"
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
"/v1/api/models": {
|
||||
"get": {
|
||||
"summary": "List available model aliases",
|
||||
@@ -2290,23 +2239,16 @@
|
||||
"SendResponse": {
|
||||
"properties": {
|
||||
"status": {
|
||||
"description": "'ok' (fresh turn dispatched), 'queued' (folded into the live turn's interjection queue, or \u2014 when `deferred` is true \u2014 parked for dispatch after the current command window), 'queue_full', 'attachments_busy' (attachments can't ride a queued turn; retry when idle), or 'cross_user_interjection' (another participant's turn is in flight; carried on the 409 body).",
|
||||
"description": "'ok', 'busy', 'queued', or 'queue_full'",
|
||||
"examples": [
|
||||
"ok",
|
||||
"busy",
|
||||
"queued",
|
||||
"queue_full",
|
||||
"attachments_busy",
|
||||
"cross_user_interjection"
|
||||
"queue_full"
|
||||
],
|
||||
"title": "Status",
|
||||
"type": "string"
|
||||
},
|
||||
"deferred": {
|
||||
"default": false,
|
||||
"description": "Set on `queued` responses: the message is parked on the workstream's deferred-send list (a slash-command window holds the worker slot, or earlier deferred sends are still pending) and dispatches as an ordinary full-fidelity send afterwards \u2014 it is NOT in a live turn's interjection queue. `DELETE .../send` retracts it until dispatch. Node-local and in-memory: a node restart before dispatch drops it (at-most-once intake).",
|
||||
"title": "Deferred",
|
||||
"type": "boolean"
|
||||
},
|
||||
"attached_ids": {
|
||||
"description": "Attachment ids actually attached to this turn. Subset of the request's `attachment_ids` (or the auto-consumed pending set). Empty when the send carries no attachments.",
|
||||
"items": {
|
||||
@@ -2483,12 +2425,6 @@
|
||||
"title": "Skill",
|
||||
"type": "string"
|
||||
},
|
||||
"persona": {
|
||||
"default": "",
|
||||
"description": "Persona name (slug) to create the workstream with. Resolved and snapshotted at creation \u2014 later persona edits never affect this workstream. Empty selects the kind's default persona; on a database with no personas seeded the workstream is created with legacy (unrestricted) behavior.",
|
||||
"title": "Persona",
|
||||
"type": "string"
|
||||
},
|
||||
"notify_targets": {
|
||||
"anyOf": [
|
||||
{
|
||||
@@ -2601,23 +2537,6 @@
|
||||
},
|
||||
"title": "Attachment Ids",
|
||||
"type": "array"
|
||||
},
|
||||
"initial_message_status": {
|
||||
"anyOf": [
|
||||
{
|
||||
"enum": [
|
||||
"queue_full",
|
||||
"refused_closed"
|
||||
],
|
||||
"type": "string"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"description": "Present ONLY when the workstream was created but its initial_message could not be delivered: 'queue_full' (a raced live worker's interjection queue was at capacity \u2014 resend via /send; any uploads stay staged) or 'refused_closed' (the workstream was closed mid-create). Absent whenever the message was dispatched.",
|
||||
"title": "Initial Message Status"
|
||||
}
|
||||
},
|
||||
"required": [
|
||||
@@ -2746,17 +2665,21 @@
|
||||
},
|
||||
"pending_approval": {
|
||||
"default": false,
|
||||
"description": "True when at least one approval cycle is live (a gate thread parked awaiting an operator approve/deny). Mirrors the same field on ``DashboardWorkstream`` / cluster live projections so a freshly-loaded chat tab can render the inline approval gate from the detail snapshot before SSE replay arrives.",
|
||||
"description": "True when the workstream is parked on ``_approval_event`` awaiting an operator approve/deny. Mirrors the same field on ``DashboardWorkstream`` / cluster live projections so a freshly-loaded chat tab can render the inline approval gate from the detail snapshot before SSE replay arrives.",
|
||||
"title": "Pending Approval",
|
||||
"type": "boolean"
|
||||
},
|
||||
"pending_approval_details": {
|
||||
"description": "Inline approval payloads, one per live cycle, oldest first \u2014 same shape as ``DashboardWorkstream.pending_approval_details``. Empty when no approval is pending. Lets a reload paint every action row + judge verdicts immediately instead of relying on the SSE approve_request replay timing window. Replaces 1.6's ``pending_approval_detail`` single-object field (breaking, 1.7).",
|
||||
"items": {
|
||||
"$ref": "#/components/schemas/PendingApprovalDetail"
|
||||
},
|
||||
"title": "Pending Approval Details",
|
||||
"type": "array"
|
||||
"pending_approval_detail": {
|
||||
"anyOf": [
|
||||
{
|
||||
"$ref": "#/components/schemas/PendingApprovalDetail"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"description": "Inline approval payload \u2014 same shape as ``DashboardWorkstream.pending_approval_detail``. ``None`` when no approval is pending. Lets a reload paint the action row + judge verdicts immediately instead of relying on the SSE approve_request replay timing window."
|
||||
}
|
||||
},
|
||||
"required": [
|
||||
@@ -2769,14 +2692,8 @@
|
||||
"type": "object"
|
||||
},
|
||||
"PendingApprovalDetail": {
|
||||
"description": "Inline approval payload merged into ``DashboardWorkstream``.\n\nOne entry per live approval CYCLE \u2014 a gate thread parked in\n``approve_tools`` awaiting the operator. Parallel task agents run\nconcurrent gates, so a workstream can have several of these at\nonce (``pending_approval_details``, oldest first). Cross-tenant\nexposure here follows the same trusted-team posture as\n``activity`` / ``tokens`` \u2014 see ``server.py``'s ``dashboard``\nhandler comment.",
|
||||
"description": "Inline approval payload merged into ``DashboardWorkstream``.\n\nSet when a workstream's ``approve_tools`` is parked on\n``_approval_event``; ``None`` (omitted) otherwise. Cross-tenant\nexposure here follows the same trusted-team posture as\n``activity`` / ``tokens`` \u2014 see ``server.py``'s ``dashboard``\nhandler comment.",
|
||||
"properties": {
|
||||
"cycle_id": {
|
||||
"default": "",
|
||||
"description": "Identity of this approval cycle. Echo it back on ``POST /v1/api/workstreams/{ws_id}/approve`` to resolve exactly this round \u2014 required for correctness when several cycles are live (parallel task agents).",
|
||||
"title": "Cycle Id",
|
||||
"type": "string"
|
||||
},
|
||||
"call_id": {
|
||||
"default": "",
|
||||
"description": "Primary call_id \u2014 first non-empty call_id in items list order. Matches the 409 ``current_call_id`` response from ``POST /v1/api/workstreams/{ws_id}/approve`` so the UI can render the same identifier the server reports as current.",
|
||||
@@ -2801,7 +2718,7 @@
|
||||
"type": "object"
|
||||
},
|
||||
"PendingApprovalItem": {
|
||||
"description": "One pending tool-call inside a ``PendingApprovalDetail`` envelope.\n\nMirrors the dict ``SessionUIBase.serialize_pending_approval_details``\nemits per item inside each cycle entry. ``heuristic_verdict`` / ``judge_verdict`` are kept\nloosely-typed because the underlying verdict shape varies by tier;\nconsumers that want the full structure can decode against\n:class:`turnstone.sdk.events.IntentVerdictEvent`.",
|
||||
"description": "One pending tool-call inside a ``PendingApprovalDetail`` envelope.\n\nMirrors the dict ``SessionUIBase.serialize_pending_approval_detail``\nemits per item. ``heuristic_verdict`` / ``judge_verdict`` are kept\nloosely-typed because the underlying verdict shape varies by tier;\nconsumers that want the full structure can decode against\n:class:`turnstone.sdk.events.IntentVerdictEvent`.",
|
||||
"properties": {
|
||||
"call_id": {
|
||||
"default": "",
|
||||
@@ -3060,13 +2977,17 @@
|
||||
"default": null,
|
||||
"title": "Project Id"
|
||||
},
|
||||
"pending_approval_details": {
|
||||
"description": "Inline approval payload for the coordinator children-tree UI: EVERY live approval cycle, oldest first \u2014 parallel task agents gate concurrently, so a workstream can hold several prompts at once. Each entry carries the cycle's items + per-call_id LLM verdict cache so a coord can render approve/deny buttons + judge pill without a separate per-child round-trip; resolve each with its ``cycle_id``. Empty when no approval is pending. Also surfaced (verbatim) on ``GET /v1/api/cluster/ws/live`` via the ``_CLUSTER_WS_LIVE_KEYS`` projection. Replaces 1.6's ``pending_approval_detail`` single-object field (breaking, 1.7).",
|
||||
"items": {
|
||||
"$ref": "#/components/schemas/PendingApprovalDetail"
|
||||
},
|
||||
"title": "Pending Approval Details",
|
||||
"type": "array"
|
||||
"pending_approval_detail": {
|
||||
"anyOf": [
|
||||
{
|
||||
"$ref": "#/components/schemas/PendingApprovalDetail"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"description": "Inline approval payload for the coordinator children-tree UI. Carries the merged ``_pending_approval`` items list + per-call_id LLM verdict cache so a coord can render approve/deny buttons + judge pill without a separate per-child round-trip. ``None`` when no approval is pending. Also surfaced (verbatim) on ``GET /v1/api/cluster/ws/live`` via the ``_CLUSTER_WS_LIVE_KEYS`` projection."
|
||||
},
|
||||
"recent_auto_approvals": {
|
||||
"description": "Per-ws ring buffer (cap 10) of recent tool calls that bypassed the operator approval gate. Surfaces ``WebUI._recent_auto_approvals`` so the coord-tree row can render an 'auto-approved by ...' pill when the child's skill / blanket / admin-policy rules silently let a tool through. Also projected onto ``GET /v1/api/cluster/ws/live`` via ``_CLUSTER_WS_LIVE_KEYS``.",
|
||||
@@ -3229,30 +3150,6 @@
|
||||
"default": 0.0,
|
||||
"title": "Context Ratio",
|
||||
"type": "number"
|
||||
},
|
||||
"project_id": {
|
||||
"anyOf": [
|
||||
{
|
||||
"type": "string"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"title": "Project Id"
|
||||
},
|
||||
"persona": {
|
||||
"anyOf": [
|
||||
{
|
||||
"type": "string"
|
||||
},
|
||||
{
|
||||
"type": "null"
|
||||
}
|
||||
],
|
||||
"default": null,
|
||||
"title": "Persona"
|
||||
}
|
||||
},
|
||||
"required": [
|
||||
@@ -3841,65 +3738,6 @@
|
||||
"title": "ListSkillSummaryResponse",
|
||||
"type": "object"
|
||||
},
|
||||
"PersonaChoice": {
|
||||
"description": "Display fields for the creation picker \u2014 the persona's levers\n(prompt / tool set / toggles) deliberately stay server-side.",
|
||||
"properties": {
|
||||
"name": {
|
||||
"description": "Persona slug, the value to pass as CreateWorkstreamRequest.persona",
|
||||
"title": "Name",
|
||||
"type": "string"
|
||||
},
|
||||
"display_name": {
|
||||
"default": "",
|
||||
"description": "Human-readable name",
|
||||
"title": "Display Name",
|
||||
"type": "string"
|
||||
},
|
||||
"description": {
|
||||
"default": "",
|
||||
"description": "What this persona is for",
|
||||
"title": "Description",
|
||||
"type": "string"
|
||||
},
|
||||
"applies_to_kinds": {
|
||||
"description": "Workstream kinds this persona can be attached to",
|
||||
"items": {
|
||||
"type": "string"
|
||||
},
|
||||
"title": "Applies To Kinds",
|
||||
"type": "array"
|
||||
},
|
||||
"is_default": {
|
||||
"default": false,
|
||||
"description": "Whether an empty persona field resolves to this one",
|
||||
"title": "Is Default",
|
||||
"type": "boolean"
|
||||
}
|
||||
},
|
||||
"required": [
|
||||
"name"
|
||||
],
|
||||
"title": "PersonaChoice",
|
||||
"type": "object"
|
||||
},
|
||||
"ListPersonaChoicesResponse": {
|
||||
"properties": {
|
||||
"personas": {
|
||||
"items": {
|
||||
"$ref": "#/components/schemas/PersonaChoice"
|
||||
},
|
||||
"title": "Personas",
|
||||
"type": "array"
|
||||
},
|
||||
"total": {
|
||||
"default": 0,
|
||||
"title": "Total",
|
||||
"type": "integer"
|
||||
}
|
||||
},
|
||||
"title": "ListPersonaChoicesResponse",
|
||||
"type": "object"
|
||||
},
|
||||
"AvailableModelInfo": {
|
||||
"properties": {
|
||||
"alias": {
|
||||
|
||||
Generated
+190
-551
File diff suppressed because it is too large
Load Diff
@@ -32,7 +32,7 @@
|
||||
],
|
||||
"license": "Apache-2.0",
|
||||
"devDependencies": {
|
||||
"typescript": "^7.0.0",
|
||||
"typescript": "^6.0.0",
|
||||
"vitest": "^4.1"
|
||||
}
|
||||
}
|
||||
|
||||
@@ -75,35 +75,15 @@ export interface ToolInfoEvent {
|
||||
items: Array<Record<string, unknown>>;
|
||||
}
|
||||
|
||||
/** One approval CYCLE awaiting the operator. Several can be outstanding
|
||||
* at once (parallel task agents each gate their own tool calls) — key
|
||||
* prompt UI by `cycle_id` and echo it back on the approve POST.
|
||||
*
|
||||
* `cycle_id` is optional because it was added in 1.7: a pre-1.7 server
|
||||
* omits it on the wire, so a current SDK talking to an older node sees
|
||||
* `undefined`. Resolve those the legacy way (no selector → oldest
|
||||
* cycle). A current server always sends it. */
|
||||
export interface ApproveRequestEvent {
|
||||
type: "approve_request";
|
||||
cycle_id?: string;
|
||||
items: Array<Record<string, unknown>>;
|
||||
judge_pending?: boolean;
|
||||
}
|
||||
|
||||
/** A specific approval cycle resolved; `cycle_id`/`call_ids` identify
|
||||
* which prompt to dismiss.
|
||||
*
|
||||
* Both are optional for the same reason as `ApproveRequestEvent.cycle_id`
|
||||
* — a pre-1.7 server emits neither, so a bare "something resolved"
|
||||
* dismisses the sole tracked prompt (the legacy fallback the UI and
|
||||
* channel adapters keep). A current server always sends both. */
|
||||
export interface ApprovalResolvedEvent {
|
||||
type: "approval_resolved";
|
||||
approved: boolean;
|
||||
feedback: string;
|
||||
always?: boolean;
|
||||
cycle_id?: string;
|
||||
call_ids?: string[];
|
||||
}
|
||||
|
||||
export interface ToolResultEvent {
|
||||
@@ -157,49 +137,6 @@ export interface CancelledEvent {
|
||||
type: "cancelled";
|
||||
}
|
||||
|
||||
/**
|
||||
* Context-compaction lifecycle. `start` carries `trigger` ("manual"/"auto";
|
||||
* auto adds `where` + `pct`); `progress` carries chunked-summarization
|
||||
* `part`/`total`/`depth` (or `retry_in`/`error` for a retry wait); `end`
|
||||
* carries `ok` plus either `before_tokens`/`after_tokens`/`summary` or the
|
||||
* failure `reason`/`message`. The successful end's summary also replays from
|
||||
* `/history` as a `role: "system"`, `source: "compaction"` entry.
|
||||
*/
|
||||
export interface CompactionEvent {
|
||||
type: "compaction";
|
||||
phase: "start" | "progress" | "end";
|
||||
/** Correlates every event of one compaction run (0 from legacy emitters). */
|
||||
compaction_id?: number;
|
||||
/**
|
||||
* End events only: true marks a force-abandoned compaction retiring
|
||||
* after a successor generation took over — skip failure notices for
|
||||
* those (an OK end's result still stands; the history swap happened).
|
||||
*/
|
||||
superseded?: boolean;
|
||||
/**
|
||||
* Failed ends only: the emitter-computed display verdict — show
|
||||
* `message` only when true, instead of re-deriving suppression from
|
||||
* reason/trigger/superseded client-side.
|
||||
*/
|
||||
notice?: boolean;
|
||||
/** Present on start and on every end (ok or failed). */
|
||||
trigger?: "manual" | "auto";
|
||||
where?: string;
|
||||
pct?: number;
|
||||
part?: number;
|
||||
total?: number;
|
||||
depth?: number;
|
||||
retry_in?: number;
|
||||
error?: string;
|
||||
warning?: string;
|
||||
ok?: boolean;
|
||||
reason?: string;
|
||||
message?: string;
|
||||
before_tokens?: number;
|
||||
after_tokens?: number;
|
||||
summary?: string;
|
||||
}
|
||||
|
||||
// Global events
|
||||
|
||||
export interface WsStateEvent {
|
||||
@@ -255,7 +192,6 @@ export type ServerEvent =
|
||||
| BusyErrorEvent
|
||||
| ClearUiEvent
|
||||
| CancelledEvent
|
||||
| CompactionEvent
|
||||
| WsStateEvent
|
||||
| WsActivityEvent
|
||||
| WsRenameEvent
|
||||
|
||||
@@ -166,13 +166,6 @@ export class TurnstoneServer extends BaseClient {
|
||||
approved?: boolean;
|
||||
feedback?: string | null;
|
||||
always?: boolean;
|
||||
/** Resolve exactly this approval cycle (from ApproveRequestEvent.cycle_id).
|
||||
* Omitting it resolves the OLDEST live cycle — ambiguous when parallel
|
||||
* task agents have several prompts outstanding, so pass it whenever the
|
||||
* triggering event is known. */
|
||||
cycleId?: string;
|
||||
/** Alternative selector: any call_id inside the target cycle. */
|
||||
callId?: string;
|
||||
}): Promise<StatusResponse> {
|
||||
return this.request(
|
||||
"POST",
|
||||
@@ -182,8 +175,6 @@ export class TurnstoneServer extends BaseClient {
|
||||
approved: opts.approved ?? true,
|
||||
feedback: opts.feedback,
|
||||
always: opts.always,
|
||||
cycle_id: opts.cycleId,
|
||||
call_id: opts.callId,
|
||||
},
|
||||
},
|
||||
);
|
||||
|
||||
@@ -130,12 +130,6 @@ export interface CreateWorkstreamRequest {
|
||||
auto_approve?: boolean;
|
||||
resume_ws?: string;
|
||||
skill?: string;
|
||||
/**
|
||||
* Persona name (slug) to create the workstream with. Resolved and
|
||||
* snapshotted at creation — later persona edits never affect this
|
||||
* workstream. Empty selects the kind's default persona.
|
||||
*/
|
||||
persona?: string;
|
||||
/**
|
||||
* Optional project to attach this workstream to. Drives the shared
|
||||
* `project` memory scope; coordinator children inherit the parent's project.
|
||||
@@ -164,13 +158,6 @@ export interface CreateWorkstreamResponse {
|
||||
message_count?: number;
|
||||
/** Ids of attachments saved by this request (multipart variant only). */
|
||||
attachment_ids?: string[];
|
||||
/**
|
||||
* Present ONLY when the workstream was created but its initial_message
|
||||
* could not be delivered: "queue_full" (raced live worker's interjection
|
||||
* queue at capacity — resend via /send; uploads stay staged) or
|
||||
* "refused_closed" (workstream closed mid-create).
|
||||
*/
|
||||
initial_message_status?: "queue_full" | "refused_closed";
|
||||
}
|
||||
|
||||
export interface CloseWorkstreamRequest {
|
||||
@@ -269,8 +256,6 @@ export interface SavedWorkstreamInfo {
|
||||
child_count?: number;
|
||||
context_tokens?: number;
|
||||
context_ratio?: number;
|
||||
/** Persona slug the workstream was created with (empty/absent = pre-persona). */
|
||||
persona?: string | null;
|
||||
}
|
||||
|
||||
export interface ListSavedWorkstreamsResponse {
|
||||
@@ -539,8 +524,6 @@ export interface ConsoleCreateWsRequest {
|
||||
model?: string;
|
||||
initial_message?: string;
|
||||
skill?: string;
|
||||
/** Persona slug — resolved and snapshotted at creation. */
|
||||
persona?: string;
|
||||
resume_ws?: string;
|
||||
}
|
||||
|
||||
|
||||
@@ -104,6 +104,7 @@ describe("TurnstoneServer attachments", () => {
|
||||
const [, init] = (fetchFn as ReturnType<typeof vi.fn>).mock.calls[0];
|
||||
expect(JSON.parse(init.body)).toEqual({
|
||||
message: "hi",
|
||||
ws_id: "ws-X",
|
||||
attachment_ids: ["a1", "a2"],
|
||||
});
|
||||
});
|
||||
@@ -116,7 +117,7 @@ describe("TurnstoneServer attachments", () => {
|
||||
});
|
||||
await client.send("hi", "ws-X");
|
||||
const [, init] = (fetchFn as ReturnType<typeof vi.fn>).mock.calls[0];
|
||||
expect(JSON.parse(init.body)).toEqual({ message: "hi" });
|
||||
expect(JSON.parse(init.body)).toEqual({ message: "hi", ws_id: "ws-X" });
|
||||
});
|
||||
|
||||
it("createWorkstream with attachments sends multipart and auto-generates ws_id", async () => {
|
||||
|
||||
@@ -74,8 +74,8 @@ describe("TurnstoneServer", () => {
|
||||
await client.send("Hello", "ws1");
|
||||
|
||||
const [url, init] = (fetchFn as ReturnType<typeof vi.fn>).mock.calls[0];
|
||||
expect(url).toBe("http://test/v1/api/workstreams/ws1/send");
|
||||
expect(JSON.parse(init.body)).toEqual({ message: "Hello" });
|
||||
expect(url).toBe("http://test/v1/api/send");
|
||||
expect(JSON.parse(init.body)).toEqual({ message: "Hello", ws_id: "ws1" });
|
||||
});
|
||||
|
||||
it("injects auth header when token provided", async () => {
|
||||
|
||||
+1
-29
@@ -3,37 +3,9 @@ not fixtures, and several test files want to import them directly."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import time
|
||||
from typing import TYPE_CHECKING, Any
|
||||
from typing import Any
|
||||
from unittest.mock import MagicMock
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from collections.abc import Callable
|
||||
|
||||
|
||||
def wait_until(cond: Callable[[], bool], timeout: float = 5.0) -> None:
|
||||
"""Poll ``cond`` to True within ``timeout`` or fail the test.
|
||||
|
||||
The worker/wake tests can't join threads by identity:
|
||||
``session_worker.send`` assigns ``ws.worker_thread`` under the lock
|
||||
BEFORE ``t.start()``, so the instant a dispatching call returns, a
|
||||
fast worker may already have run its exit backstop and installed the
|
||||
(not-yet-started) wake thread — joining whatever ``ws.worker_thread``
|
||||
points at races ``RuntimeError: cannot join thread before it is
|
||||
started``. Poll outcomes instead.
|
||||
"""
|
||||
deadline = time.monotonic() + timeout
|
||||
while time.monotonic() < deadline:
|
||||
if cond():
|
||||
return
|
||||
time.sleep(0.005)
|
||||
if cond():
|
||||
# Final re-check: the condition can become true during the last
|
||||
# sleep (or a CI descheduling stall past the deadline) — failing
|
||||
# without re-looking makes the helper itself a flake source.
|
||||
return
|
||||
raise AssertionError("condition not met within timeout")
|
||||
|
||||
|
||||
def make_chat_session(**overrides: Any) -> Any:
|
||||
"""Build a minimal ``ChatSession`` with sane test defaults.
|
||||
|
||||
@@ -1,43 +0,0 @@
|
||||
"""Shared process/polling helpers for the bash + background-shell suites.
|
||||
|
||||
One copy instead of three: ``test_bash_tool_background_hang``,
|
||||
``test_background_shells`` and ``test_bash_background_tool`` all assert on
|
||||
process liveness and poll for asynchronous state. Leading underscore so
|
||||
pytest doesn't collect it.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import contextlib
|
||||
import os
|
||||
import signal
|
||||
import time
|
||||
|
||||
|
||||
def pid_alive(pid: int) -> bool:
|
||||
try:
|
||||
os.kill(pid, 0)
|
||||
except ProcessLookupError:
|
||||
return False
|
||||
except PermissionError:
|
||||
return True
|
||||
return True
|
||||
|
||||
|
||||
def kill_pid(pid: int) -> None:
|
||||
with contextlib.suppress(OSError):
|
||||
os.kill(pid, signal.SIGKILL)
|
||||
|
||||
|
||||
def poll_until(predicate, timeout=10.0, interval=0.05):
|
||||
"""Poll ``predicate`` until truthy or ``timeout``; RETURNS the last value
|
||||
(falsy on timeout — assert at the call site). Deliberately named apart
|
||||
from ``tests/_helpers.wait_until``, which RAISES on timeout: two
|
||||
same-named helpers with opposite failure semantics invite silently-green
|
||||
tests."""
|
||||
deadline = time.monotonic() + timeout
|
||||
value = predicate()
|
||||
while not value and time.monotonic() < deadline:
|
||||
time.sleep(interval)
|
||||
value = predicate()
|
||||
return value
|
||||
@@ -51,12 +51,6 @@ def make_replay_mocks(
|
||||
ui._ws_messages = 0
|
||||
for key, value in ui_overrides.items():
|
||||
setattr(ui, key, value)
|
||||
# Both replay paths read cycle cards via ``pending_approval_cards()``
|
||||
# (one card per concurrent approval cycle). Model it from the
|
||||
# single-slot ``_pending_approval`` override so tests keep seeding
|
||||
# the one field; a bare MagicMock here would iterate empty and
|
||||
# silently drop the approve_request from the replay.
|
||||
ui.pending_approval_cards = lambda: [ui._pending_approval] if ui._pending_approval else []
|
||||
ws = MagicMock()
|
||||
ws.session = session
|
||||
request = MagicMock()
|
||||
|
||||
@@ -14,12 +14,9 @@ collect it as a test file — it's an importable utility, not a test.
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
from types import SimpleNamespace
|
||||
from typing import Any
|
||||
from unittest.mock import MagicMock
|
||||
|
||||
from turnstone.core.providers import StreamChunk, ToolCallDelta
|
||||
from turnstone.core.session import ChatSession
|
||||
from turnstone.core.session_ui_base import SessionUIBase
|
||||
|
||||
@@ -46,304 +43,3 @@ def make_session(**kwargs: Any) -> ChatSession:
|
||||
}
|
||||
defaults.update(kwargs)
|
||||
return ChatSession(**defaults)
|
||||
|
||||
|
||||
def mock_completion_result(
|
||||
content: str = "",
|
||||
tool_calls: list[dict[str, Any]] | None = None,
|
||||
) -> MagicMock:
|
||||
"""A provider result shaped like ``CompletionResult``.
|
||||
|
||||
Callers that route through ``model_turn`` (judges, task agents, and
|
||||
every lane #827 migrates) hit its re-ingest, which iterates
|
||||
``tool_calls``/``provider_blocks`` and joins ``reasoning`` — a bare
|
||||
MagicMock attribute would TypeError deep inside the seam, so every
|
||||
field the re-ingest reads is pinned to a real value here. ONE shared
|
||||
definition: when the re-ingest starts reading a new CompletionResult
|
||||
field, add it here and every suite moves together.
|
||||
"""
|
||||
result = MagicMock()
|
||||
result.content = content
|
||||
result.tool_calls = tool_calls
|
||||
result.finish_reason = "stop"
|
||||
result.usage = None
|
||||
result.provider_blocks = []
|
||||
result.reasoning = ""
|
||||
return result
|
||||
|
||||
|
||||
def fake_chat_stream(
|
||||
*,
|
||||
content: str | None = None,
|
||||
tool_calls: list[dict[str, str]] | None = None,
|
||||
finish_reason: str = "stop",
|
||||
prompt_tokens: int = 10,
|
||||
completion_tokens: int = 5,
|
||||
reasoning_content: str | None = None,
|
||||
reasoning: str | None = None,
|
||||
) -> list[Any]:
|
||||
"""Fake OpenAI Chat Completions SSE chunks for driving the REAL
|
||||
``OpenAIChatCompletionsProvider`` through a fake SDK client::
|
||||
|
||||
client.chat.completions.create = lambda **kw: fake_chat_stream(...)
|
||||
|
||||
Exercises the adapter's ``_iter_stream`` plus ``drain_stream`` end to
|
||||
end (the highest-fidelity fake lane), unlike ``as_stream`` which fakes
|
||||
at the provider boundary. ``tool_calls`` entries are
|
||||
``{"id", "name", "arguments"}`` dicts. ``SimpleNamespace`` (not
|
||||
``MagicMock``) so absent SDK fields read as real ``None`` — an
|
||||
auto-created mock attribute would leak into ``len()``/string paths.
|
||||
|
||||
Emits the realistic three-phase shape: data chunk(s), a finish-reason
|
||||
chunk, then the ``stream_options.include_usage`` usage-only chunk with
|
||||
empty ``choices``.
|
||||
"""
|
||||
|
||||
def _delta(
|
||||
content_val: str | None = None,
|
||||
tcs: list[Any] | None = None,
|
||||
rc: str | None = None,
|
||||
rsn: str | None = None,
|
||||
) -> SimpleNamespace:
|
||||
return SimpleNamespace(
|
||||
content=content_val,
|
||||
tool_calls=tcs,
|
||||
reasoning=rsn,
|
||||
reasoning_content=rc,
|
||||
annotations=None,
|
||||
)
|
||||
|
||||
chunks: list[Any] = []
|
||||
if reasoning_content is not None or reasoning is not None:
|
||||
chunks.append(
|
||||
SimpleNamespace(
|
||||
choices=[
|
||||
SimpleNamespace(
|
||||
finish_reason=None, delta=_delta(rc=reasoning_content, rsn=reasoning)
|
||||
)
|
||||
],
|
||||
usage=None,
|
||||
)
|
||||
)
|
||||
if content is not None:
|
||||
chunks.append(
|
||||
SimpleNamespace(
|
||||
choices=[SimpleNamespace(finish_reason=None, delta=_delta(content))],
|
||||
usage=None,
|
||||
)
|
||||
)
|
||||
if tool_calls:
|
||||
tcs = [
|
||||
SimpleNamespace(
|
||||
index=i,
|
||||
id=tc.get("id", ""),
|
||||
function=SimpleNamespace(
|
||||
name=tc.get("name", ""), arguments=tc.get("arguments", "")
|
||||
),
|
||||
)
|
||||
for i, tc in enumerate(tool_calls)
|
||||
]
|
||||
chunks.append(
|
||||
SimpleNamespace(
|
||||
choices=[SimpleNamespace(finish_reason=None, delta=_delta(None, tcs))],
|
||||
usage=None,
|
||||
)
|
||||
)
|
||||
chunks.append(
|
||||
SimpleNamespace(
|
||||
choices=[SimpleNamespace(finish_reason=finish_reason, delta=_delta())],
|
||||
usage=None,
|
||||
)
|
||||
)
|
||||
chunks.append(
|
||||
SimpleNamespace(
|
||||
choices=[],
|
||||
usage=SimpleNamespace(
|
||||
prompt_tokens=prompt_tokens,
|
||||
completion_tokens=completion_tokens,
|
||||
total_tokens=prompt_tokens + completion_tokens,
|
||||
prompt_tokens_details=None,
|
||||
input_tokens_details=None,
|
||||
),
|
||||
)
|
||||
)
|
||||
return chunks
|
||||
|
||||
|
||||
class _ScriptedClient:
|
||||
"""Callable client-method fake following a script of stream builders.
|
||||
|
||||
Call N returns the stream described by ``scripts[N]``; the last script
|
||||
repeats for any further calls. Each script is a dict of kwargs for
|
||||
the bound stream builder, or a pre-built return value. Records every
|
||||
call's kwargs on ``.calls`` — read ``len(fn.calls)`` where a test
|
||||
previously kept its own counter cell, and ``fn.calls[i]["messages"]``
|
||||
where it captured request bodies.
|
||||
"""
|
||||
|
||||
def __init__(self, scripts: tuple[Any, ...], to_stream: Any) -> None:
|
||||
self._scripts = scripts
|
||||
self._to_stream = to_stream
|
||||
self.calls: list[dict[str, Any]] = []
|
||||
|
||||
def __call__(self, **kwargs: Any) -> Any:
|
||||
self.calls.append(kwargs)
|
||||
script = self._scripts[min(len(self.calls) - 1, len(self._scripts) - 1)]
|
||||
return self._to_stream(**script) if isinstance(script, dict) else script
|
||||
|
||||
|
||||
def scripted_chat_client(*scripts: Any) -> _ScriptedClient:
|
||||
"""A scripted ``client.chat.completions.create`` — dict scripts are
|
||||
:func:`fake_chat_stream` kwargs."""
|
||||
return _ScriptedClient(scripts, fake_chat_stream)
|
||||
|
||||
|
||||
def scripted_anthropic_client(*scripts: Any) -> _ScriptedClient:
|
||||
"""A scripted ``client.messages.stream`` — dict scripts are
|
||||
:func:`fake_anthropic_stream` kwargs (``blocks`` plus optional
|
||||
``stop_reason``/``usage``)."""
|
||||
return _ScriptedClient(scripts, fake_anthropic_stream)
|
||||
|
||||
|
||||
class FakeAnthropicBlock:
|
||||
"""A full-content Anthropic content-block fake for
|
||||
:func:`fake_anthropic_stream` — plain attributes plus the
|
||||
``model_dump()`` the provider's block capture reads."""
|
||||
|
||||
def __init__(self, **fields: Any) -> None:
|
||||
self._fields = fields
|
||||
for key, value in fields.items():
|
||||
setattr(self, key, value)
|
||||
|
||||
def model_dump(self, **_kw: Any) -> dict[str, Any]:
|
||||
return dict(self._fields)
|
||||
|
||||
|
||||
def fake_anthropic_stream(
|
||||
blocks: list[Any],
|
||||
*,
|
||||
stop_reason: str | None = "end_turn",
|
||||
usage: Any = None,
|
||||
) -> Any:
|
||||
"""Fake Anthropic SDK stream context manager for tests that drive the
|
||||
REAL ``AnthropicProvider`` through a fake client::
|
||||
|
||||
client.messages.stream = lambda **kw: fake_anthropic_stream(...)
|
||||
|
||||
Accepts the same full-content block fakes the pre-#831
|
||||
``get_final_message`` fixtures used (objects with ``.type`` + fields
|
||||
and ``model_dump()``) and synthesizes the real event grammar the
|
||||
streaming iterator consumes: ``content_block_start`` carries the block
|
||||
with its text/thinking/signature EMPTIED and ``input`` as ``{}`` (the
|
||||
SDK start shape), deltas carry the content, ``content_block_stop``
|
||||
finalizes tool input, and the closing ``message_delta`` carries
|
||||
``stop_reason`` (+ optional usage object). Without the stripping, the
|
||||
provider's raw-block accumulator would double every text/thinking
|
||||
field (start capture + delta append).
|
||||
|
||||
``stop_reason=None`` omits the closing ``message_delta`` entirely —
|
||||
the terminal-signal-less lax-gateway shape ``finish_reason_optional``
|
||||
exists for (content arrives, then the stream just ends).
|
||||
"""
|
||||
events: list[Any] = []
|
||||
for idx, block in enumerate(blocks):
|
||||
d = dict(block.model_dump()) if hasattr(block, "model_dump") else dict(vars(block))
|
||||
btype = d.get("type", "")
|
||||
start = dict(d)
|
||||
if btype == "text":
|
||||
start["text"] = ""
|
||||
elif btype == "thinking":
|
||||
start["thinking"] = ""
|
||||
start["signature"] = ""
|
||||
elif btype == "tool_use":
|
||||
start["input"] = {}
|
||||
events.append(
|
||||
SimpleNamespace(
|
||||
type="content_block_start", index=idx, content_block=SimpleNamespace(**start)
|
||||
)
|
||||
)
|
||||
if btype == "text" and d.get("text"):
|
||||
events.append(
|
||||
SimpleNamespace(
|
||||
type="content_block_delta",
|
||||
index=idx,
|
||||
delta=SimpleNamespace(type="text_delta", text=d["text"]),
|
||||
)
|
||||
)
|
||||
elif btype == "thinking":
|
||||
if d.get("thinking"):
|
||||
events.append(
|
||||
SimpleNamespace(
|
||||
type="content_block_delta",
|
||||
index=idx,
|
||||
delta=SimpleNamespace(type="thinking_delta", thinking=d["thinking"]),
|
||||
)
|
||||
)
|
||||
if d.get("signature"):
|
||||
events.append(
|
||||
SimpleNamespace(
|
||||
type="content_block_delta",
|
||||
index=idx,
|
||||
delta=SimpleNamespace(type="signature_delta", signature=d["signature"]),
|
||||
)
|
||||
)
|
||||
elif btype == "tool_use":
|
||||
events.append(
|
||||
SimpleNamespace(
|
||||
type="content_block_delta",
|
||||
index=idx,
|
||||
delta=SimpleNamespace(
|
||||
type="input_json_delta",
|
||||
partial_json=json.dumps(d.get("input", {})),
|
||||
),
|
||||
)
|
||||
)
|
||||
events.append(SimpleNamespace(type="content_block_stop", index=idx))
|
||||
if stop_reason is not None or usage is not None:
|
||||
events.append(
|
||||
SimpleNamespace(
|
||||
type="message_delta", usage=usage, delta=SimpleNamespace(stop_reason=stop_reason)
|
||||
)
|
||||
)
|
||||
|
||||
mgr = MagicMock()
|
||||
mgr.__enter__ = MagicMock(return_value=events)
|
||||
mgr.__exit__ = MagicMock(return_value=False)
|
||||
return mgr
|
||||
|
||||
|
||||
def as_stream(result: Any) -> list[StreamChunk]:
|
||||
"""Adapt a ``CompletionResult``-shaped fake to a ``create_streaming``
|
||||
return value (single terminal chunk).
|
||||
|
||||
The #831 transport collapse routes every single-shot lane through
|
||||
``drain_stream(provider.create_streaming(...))``, so provider fakes
|
||||
return chunk iterables now. Tests keep building result-shaped fakes
|
||||
(``mock_completion_result`` or hand-rolled) and wrap them at
|
||||
assignment: ``provider.create_streaming.return_value =
|
||||
as_stream(result)``. A list re-iterates on every call, so one
|
||||
``return_value`` serves repeated-call tests; convert AFTER mutating
|
||||
the fake's fields — the chunk snapshots them.
|
||||
|
||||
Multi-chunk accumulation semantics are exercised by the dedicated
|
||||
``drain_stream`` unit tests, not through this helper.
|
||||
"""
|
||||
deltas = [
|
||||
ToolCallDelta(
|
||||
index=i,
|
||||
id=tc.get("id", ""),
|
||||
name=tc.get("function", {}).get("name", ""),
|
||||
arguments_delta=tc.get("function", {}).get("arguments", ""),
|
||||
)
|
||||
for i, tc in enumerate(result.tool_calls or [])
|
||||
]
|
||||
return [
|
||||
StreamChunk(
|
||||
content_delta=result.content or "",
|
||||
reasoning_delta=getattr(result, "reasoning", "") or "",
|
||||
tool_call_deltas=deltas,
|
||||
usage=result.usage,
|
||||
finish_reason=result.finish_reason or "stop",
|
||||
provider_blocks=list(result.provider_blocks or []),
|
||||
)
|
||||
]
|
||||
|
||||
@@ -1,604 +0,0 @@
|
||||
"""Browser-fidelity SSE recovery harness helpers.
|
||||
|
||||
The load-bearing assembly for ``tests/test_sse_recovery_e2e.py``: a
|
||||
``BrowserlikeSSEClient`` that speaks the exact wire contract the real
|
||||
``turnstone/shared_static/interactive.js`` pane speaks, and the
|
||||
assertion helpers the scenarios share. The server boot machinery lives
|
||||
in ``_sse_recovery_server.py``.
|
||||
|
||||
Why a raw-socket SSE reader (and not ``httpx.stream``): the slow-consumer
|
||||
overflow scenario needs the consumer to STALL — stop reading the socket
|
||||
so the server's SSE generator blocks on ``await send`` and stops draining
|
||||
the per-UI listener queue, which then poisons at its cap. A faithful
|
||||
stall needs (a) precise control over when bytes are read and (b) a small
|
||||
``SO_RCVBUF`` so the in-flight backlog before poison stays bounded to
|
||||
~100 KB instead of the client kernel's multi-MB autotuned default (which
|
||||
would need tens of thousands of events to overflow). A raw socket gives
|
||||
both; httpx (used here only for the plain ``/history`` request/response)
|
||||
gives neither. This is ALSO closer to the browser: EventSource has a
|
||||
bounded receive buffer, not an unbounded one.
|
||||
|
||||
Client contract mirrored from interactive.js (line references are to
|
||||
that file on the ``fix/sse-truncated-resync`` branch):
|
||||
|
||||
- ``_last_event_id`` advances ONLY from SSE ``id:`` fields, and only
|
||||
ring-buffer events carry one — synthetic replay frames (connected /
|
||||
status / state_change / in_progress_snapshot / replay_truncated /
|
||||
stream_overflow) do not, exactly like ``EventSource.lastEventId``
|
||||
(interactive.js onmessage ~1378).
|
||||
- reconnect presents ``connectCursor = _truncatedFromCursor ??
|
||||
_lastEventId`` as ``?last_event_id=`` (manual path) or a
|
||||
``Last-Event-ID`` header (native EventSource auto-reconnect path)
|
||||
(interactive.js connectSSE ~1328).
|
||||
- on a ``replay_truncated`` envelope the client records the
|
||||
truncation-time cursor keep-oldest (``_truncatedFromCursor =
|
||||
_lastEventId`` only when null) and runs ``_loadHistoryThenConnect``
|
||||
(disconnect → /history → adopt cursor → reconnect); a FAILED
|
||||
/history leaves the record armed so the reconnect re-presents the
|
||||
truncation-time cursor and re-draws the envelope (interactive.js
|
||||
handleEvent replay_truncated ~2303, _loadHistoryThenConnect ~1604,
|
||||
_refetchHistory seedCursor ~1706).
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import contextlib
|
||||
import json
|
||||
import socket
|
||||
import threading
|
||||
import time
|
||||
from dataclasses import dataclass
|
||||
from typing import TYPE_CHECKING, Any
|
||||
from urllib.parse import urlsplit
|
||||
|
||||
import httpx
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from collections.abc import Callable
|
||||
|
||||
# Small client receive buffer so a stalled consumer's in-flight backlog
|
||||
# before the server-side poison stays bounded (~100 KB) instead of the
|
||||
# multi-MB autotuned default. Paired with the server's small SO_SNDBUF
|
||||
# (see _sse_recovery_server.build_recovery_server).
|
||||
_CLIENT_RCVBUF = 2048
|
||||
|
||||
|
||||
@dataclass
|
||||
class SSEFrame:
|
||||
"""One decoded SSE frame, tagged with the connection it arrived on.
|
||||
|
||||
``event_id`` is the ``id:`` field verbatim (a stringified integer,
|
||||
or ``None`` for id-less synthetic frames — the same string domain as
|
||||
``EventSource.lastEventId``). ``etype`` is the ``type`` field of the
|
||||
JSON ``data:`` payload (the application event type), distinct from
|
||||
any SSE ``event:`` field, which the server never uses.
|
||||
"""
|
||||
|
||||
conn_index: int
|
||||
event_id: str | None
|
||||
etype: str | None
|
||||
payload: dict[str, Any] | None
|
||||
raw: str
|
||||
|
||||
@property
|
||||
def event_id_int(self) -> int | None:
|
||||
if self.event_id is None:
|
||||
return None
|
||||
try:
|
||||
return int(self.event_id)
|
||||
except ValueError:
|
||||
return None
|
||||
|
||||
|
||||
class BrowserlikeSSEClient:
|
||||
"""A single interactive pane's SSE + /history state machine.
|
||||
|
||||
Not thread-safe against concurrent public calls; drive it from one
|
||||
test thread. Internally a per-connection reader thread decodes the
|
||||
stream; ``stall()`` / ``resume()`` gate that thread's socket reads so
|
||||
a test can build server-side backpressure without closing the
|
||||
connection (the slow-consumer → listener-queue-poison path).
|
||||
"""
|
||||
|
||||
def __init__(self, base_url: str, ws_id: str, token: str) -> None:
|
||||
parts = urlsplit(base_url)
|
||||
self._host = parts.hostname or "127.0.0.1"
|
||||
self._port = parts.port or 80
|
||||
self._ws_id = ws_id
|
||||
self._token = token
|
||||
self._auth = {"Authorization": f"Bearer {token}"}
|
||||
self._http = httpx.Client(
|
||||
base_url=f"http://{self._host}:{self._port}", timeout=httpx.Timeout(15.0)
|
||||
)
|
||||
|
||||
# EventSource-equivalent cursor state.
|
||||
self._last_event_id: str | None = None
|
||||
self._truncated_from_cursor: str | None = None
|
||||
|
||||
# Transcript. ``_all_frames`` is the cross-connection accumulation
|
||||
# (what "the client eventually saw"); ``_conn_frames`` keeps each
|
||||
# connection's slice for per-connection assertions (contiguity).
|
||||
self._all_frames: list[SSEFrame] = []
|
||||
self._conn_frames: list[list[SSEFrame]] = []
|
||||
self._frames_lock = threading.Lock()
|
||||
|
||||
# Reader plumbing.
|
||||
self._sock: socket.socket | None = None
|
||||
self._reader: threading.Thread | None = None
|
||||
self._stop = threading.Event()
|
||||
self._read_gate = threading.Event()
|
||||
self._read_gate.set() # reading permitted by default
|
||||
self._status: int | None = None
|
||||
self._headers_done = threading.Event()
|
||||
|
||||
# -- connection lifecycle ------------------------------------------------
|
||||
|
||||
def _events_path(self, cursor: str | None) -> str:
|
||||
path = f"/v1/api/workstreams/{self._ws_id}/events"
|
||||
if cursor is not None:
|
||||
path += f"?last_event_id={cursor}"
|
||||
return path
|
||||
|
||||
def connect(self, *, native: bool = False, rcvbuf: int | None = None) -> None:
|
||||
"""Open the SSE stream, presenting the client's current cursor.
|
||||
|
||||
``native=True`` models the browser's EventSource auto-reconnect:
|
||||
the cursor rides a ``Last-Event-ID`` HEADER and never appears in
|
||||
the URL. ``native=False`` models the manual ``new EventSource(url
|
||||
+ '?last_event_id=')`` path interactive.js uses when it must
|
||||
override the live cursor (the ``connectCursor`` chokepoint).
|
||||
|
||||
``rcvbuf`` shrinks this connection's ``SO_RCVBUF`` — pass
|
||||
``_CLIENT_RCVBUF`` on a connection the test will ``stall()`` so the
|
||||
in-flight backlog before the server-side poison stays bounded.
|
||||
Leave it ``None`` (OS default) on recovery reconnects so the ring
|
||||
replay is not throttled to a crawl.
|
||||
"""
|
||||
if self._reader is not None:
|
||||
raise RuntimeError("already connected; disconnect() first")
|
||||
connect_cursor = (
|
||||
self._truncated_from_cursor
|
||||
if self._truncated_from_cursor is not None
|
||||
else self._last_event_id
|
||||
)
|
||||
header_lines = [
|
||||
f"Host: {self._host}:{self._port}",
|
||||
f"Authorization: Bearer {self._token}",
|
||||
"Accept: text/event-stream",
|
||||
"Cache-Control: no-cache",
|
||||
]
|
||||
if native:
|
||||
path = self._events_path(None)
|
||||
if connect_cursor is not None:
|
||||
header_lines.append(f"Last-Event-ID: {connect_cursor}")
|
||||
else:
|
||||
path = self._events_path(connect_cursor)
|
||||
request = f"GET {path} HTTP/1.1\r\n" + "\r\n".join(header_lines) + "\r\n\r\n"
|
||||
|
||||
sock = socket.create_connection((self._host, self._port), timeout=10)
|
||||
if rcvbuf is not None:
|
||||
sock.setsockopt(socket.SOL_SOCKET, socket.SO_RCVBUF, rcvbuf)
|
||||
sock.setsockopt(socket.IPPROTO_TCP, socket.TCP_NODELAY, 1)
|
||||
sock.settimeout(None)
|
||||
sock.sendall(request.encode())
|
||||
self._sock = sock
|
||||
|
||||
self._stop.clear()
|
||||
self._read_gate.set()
|
||||
self._status = None
|
||||
self._headers_done.clear()
|
||||
conn_index = len(self._conn_frames)
|
||||
frames: list[SSEFrame] = []
|
||||
self._conn_frames.append(frames)
|
||||
self._reader = threading.Thread(
|
||||
target=self._read_loop,
|
||||
args=(sock, conn_index, frames),
|
||||
name=f"sse-reader-{self._ws_id[:6]}-{conn_index}",
|
||||
daemon=True,
|
||||
)
|
||||
self._reader.start()
|
||||
|
||||
# Surface a non-200 handshake to the caller (409 half-built UI,
|
||||
# 404 unknown ws, 401 auth) rather than silently reading nothing.
|
||||
if not self._headers_done.wait(timeout=10):
|
||||
self.disconnect()
|
||||
raise AssertionError("events connect: no HTTP response headers")
|
||||
if self._status != 200:
|
||||
status = self._status
|
||||
self.disconnect()
|
||||
raise AssertionError(f"events connect returned HTTP {status}")
|
||||
|
||||
def disconnect(self) -> None:
|
||||
"""Close the stream and join the reader (leak-guard clean)."""
|
||||
self._stop.set()
|
||||
self._read_gate.set() # release a stalled reader so it sees _stop
|
||||
sock = self._sock
|
||||
if sock is not None:
|
||||
with contextlib.suppress(OSError):
|
||||
sock.shutdown(socket.SHUT_RDWR) # interrupt a blocked recv
|
||||
reader = self._reader
|
||||
if reader is not None:
|
||||
reader.join(timeout=15)
|
||||
if reader.is_alive():
|
||||
raise AssertionError("SSE reader thread failed to stop")
|
||||
if sock is not None:
|
||||
with contextlib.suppress(OSError):
|
||||
sock.close()
|
||||
self._sock = None
|
||||
self._reader = None
|
||||
|
||||
def close(self) -> None:
|
||||
"""Full teardown: disconnect any live stream + close the HTTP client."""
|
||||
if self._reader is not None:
|
||||
self.disconnect()
|
||||
self._http.close()
|
||||
|
||||
# -- the stall gate (backpressure driver) --------------------------------
|
||||
|
||||
def stall(self) -> None:
|
||||
"""Stop reading the socket. The kernel + uvicorn send buffers fill,
|
||||
blocking the server's SSE generator on its ``await send``, so it
|
||||
stops draining the per-UI listener queue — which poisons at its cap.
|
||||
"""
|
||||
self._read_gate.clear()
|
||||
|
||||
def resume(self) -> None:
|
||||
"""Resume reading. A poisoned-and-closed stream delivers its
|
||||
``stream_overflow`` farewell frame once the backlog drains."""
|
||||
self._read_gate.set()
|
||||
|
||||
# -- reader --------------------------------------------------------------
|
||||
|
||||
def _read_loop(self, sock: socket.socket, conn_index: int, frames: list[SSEFrame]) -> None:
|
||||
raw = b"" # undecoded bytes (headers, then chunked framing)
|
||||
sse = b"" # decoded SSE byte stream
|
||||
headers_parsed = False
|
||||
chunked = False
|
||||
while not self._stop.is_set():
|
||||
# Backpressure gate: while stalled we do NOT read the socket, so
|
||||
# its receive buffer fills and TCP flow control stalls the server.
|
||||
if not self._read_gate.wait(timeout=0.1):
|
||||
continue
|
||||
if self._stop.is_set():
|
||||
break
|
||||
try:
|
||||
chunk = sock.recv(65536)
|
||||
except OSError:
|
||||
break
|
||||
if not chunk:
|
||||
break # server closed
|
||||
raw += chunk
|
||||
if not headers_parsed:
|
||||
if b"\r\n\r\n" not in raw:
|
||||
continue
|
||||
header_blob, raw = raw.split(b"\r\n\r\n", 1)
|
||||
self._parse_headers(header_blob)
|
||||
chunked = b"transfer-encoding: chunked" in header_blob.lower()
|
||||
headers_parsed = True
|
||||
self._headers_done.set()
|
||||
if chunked:
|
||||
decoded, raw = _dechunk(raw)
|
||||
sse += decoded
|
||||
else:
|
||||
sse += raw
|
||||
raw = b""
|
||||
sse = sse.replace(b"\r\n", b"\n")
|
||||
while b"\n\n" in sse:
|
||||
block, sse = sse.split(b"\n\n", 1)
|
||||
self._handle_block(block.decode("utf-8", "replace"), conn_index, frames)
|
||||
|
||||
def _parse_headers(self, header_blob: bytes) -> None:
|
||||
first_line = header_blob.split(b"\r\n", 1)[0].decode("latin-1")
|
||||
# "HTTP/1.1 200 OK"
|
||||
parts = first_line.split(" ", 2)
|
||||
if len(parts) >= 2 and parts[1].isdigit():
|
||||
self._status = int(parts[1])
|
||||
|
||||
def _handle_block(self, block_text: str, conn_index: int, frames: list[SSEFrame]) -> None:
|
||||
event_id: str | None = None
|
||||
data_parts: list[str] = []
|
||||
retry: str | None = None
|
||||
for line in block_text.split("\n"):
|
||||
if not line or line.startswith(":"):
|
||||
continue # blank or comment (ping)
|
||||
field_name, _, value = line.partition(":")
|
||||
if value.startswith(" "):
|
||||
value = value[1:] # SSE strips a single leading space
|
||||
if field_name == "id":
|
||||
event_id = value
|
||||
elif field_name == "data":
|
||||
data_parts.append(value)
|
||||
elif field_name == "retry":
|
||||
retry = value
|
||||
# EventSource semantics: an event carrying an ``id:`` sets the
|
||||
# last-event-id buffer; an event without one leaves it unchanged.
|
||||
if event_id is not None:
|
||||
self._last_event_id = event_id
|
||||
if not data_parts:
|
||||
if retry is not None:
|
||||
self._record(SSEFrame(conn_index, None, "retry", None, block_text), frames)
|
||||
return
|
||||
data_str = "\n".join(data_parts)
|
||||
payload: dict[str, Any] | None
|
||||
try:
|
||||
parsed = json.loads(data_str)
|
||||
payload = parsed if isinstance(parsed, dict) else None
|
||||
except ValueError:
|
||||
payload = None
|
||||
etype = payload.get("type") if payload is not None else None
|
||||
frame = SSEFrame(conn_index, event_id, etype, payload, data_str)
|
||||
self._record(frame, frames)
|
||||
# Mirror the pane: the FIRST replay_truncated for an unrepaired gap
|
||||
# records the truncation-time cursor (keep-oldest). Its consumer is
|
||||
# the reconnect chokepoint (see ``connect``).
|
||||
if etype == "replay_truncated" and self._truncated_from_cursor is None:
|
||||
self._truncated_from_cursor = self._last_event_id
|
||||
|
||||
def _record(self, frame: SSEFrame, frames: list[SSEFrame]) -> None:
|
||||
with self._frames_lock:
|
||||
frames.append(frame)
|
||||
self._all_frames.append(frame)
|
||||
|
||||
# -- /history + cursor flow ----------------------------------------------
|
||||
|
||||
def fetch_history(self) -> dict[str, Any]:
|
||||
"""GET /history and return the parsed JSON ({ws_id, messages, cursor})."""
|
||||
r = self._http.get(f"/v1/api/workstreams/{self._ws_id}/history", headers=self._auth)
|
||||
r.raise_for_status()
|
||||
result: dict[str, Any] = r.json()
|
||||
return result
|
||||
|
||||
def seed_from_history(self) -> dict[str, Any]:
|
||||
"""The seedCursor step: fetch /history, adopt a non-null resume
|
||||
cursor into ``_last_event_id``, and clear the truncation record on
|
||||
a successful render (replayHistory clears ``_truncatedFromCursor``).
|
||||
"""
|
||||
data = self.fetch_history()
|
||||
cursor = data.get("cursor")
|
||||
if cursor is not None:
|
||||
self._last_event_id = str(cursor)
|
||||
self._truncated_from_cursor = None # successful full render repairs the gap
|
||||
return data
|
||||
|
||||
def load_history_then_connect(
|
||||
self, *, fail_history: bool = False, native: bool = False
|
||||
) -> dict[str, Any] | None:
|
||||
"""Reproduce interactive.js ``_loadHistoryThenConnect``.
|
||||
|
||||
Disconnect first, drop the live cursor (``_last_event_id = None``)
|
||||
but KEEP ``_truncated_from_cursor`` armed, then fetch /history and
|
||||
reconnect. On success adopt the returned cursor and clear the
|
||||
truncation record; on a FAILED /history (``fail_history`` — the
|
||||
harness IS the client here, so a client-side simulated failure is
|
||||
faithful) leave the record armed so the reconnect re-presents the
|
||||
truncation-time cursor and re-draws ``replay_truncated``.
|
||||
|
||||
Returns the /history JSON, or ``None`` when the fetch failed.
|
||||
"""
|
||||
if self._reader is not None:
|
||||
self.disconnect()
|
||||
self._last_event_id = None
|
||||
data: dict[str, Any] | None
|
||||
if fail_history:
|
||||
data = None
|
||||
else:
|
||||
data = self.fetch_history()
|
||||
cursor = data.get("cursor")
|
||||
if cursor is not None:
|
||||
self._last_event_id = str(cursor)
|
||||
self._truncated_from_cursor = None
|
||||
self.connect(native=native)
|
||||
return data
|
||||
|
||||
# -- accessors + waits ---------------------------------------------------
|
||||
|
||||
@property
|
||||
def last_event_id(self) -> str | None:
|
||||
return self._last_event_id
|
||||
|
||||
@property
|
||||
def truncated_from_cursor(self) -> str | None:
|
||||
return self._truncated_from_cursor
|
||||
|
||||
def all_frames(self) -> list[SSEFrame]:
|
||||
with self._frames_lock:
|
||||
return list(self._all_frames)
|
||||
|
||||
def conn_frames(self, conn_index: int) -> list[SSEFrame]:
|
||||
with self._frames_lock:
|
||||
return list(self._conn_frames[conn_index])
|
||||
|
||||
def latest_conn_frames(self) -> list[SSEFrame]:
|
||||
with self._frames_lock:
|
||||
return list(self._conn_frames[-1]) if self._conn_frames else []
|
||||
|
||||
def num_connections(self) -> int:
|
||||
with self._frames_lock:
|
||||
return len(self._conn_frames)
|
||||
|
||||
def frames_of_type(self, etype: str) -> list[SSEFrame]:
|
||||
return [f for f in self.all_frames() if f.etype == etype]
|
||||
|
||||
def has_type(self, etype: str) -> bool:
|
||||
return any(f.etype == etype for f in self.all_frames())
|
||||
|
||||
def tool_output_by_call(self) -> dict[str, str]:
|
||||
"""Concatenate every ``tool_output_chunk`` payload per call_id, in
|
||||
arrival order — the reconstructed live stream for each call."""
|
||||
out: dict[str, str] = {}
|
||||
for f in self.all_frames():
|
||||
if f.etype == "tool_output_chunk" and f.payload is not None:
|
||||
cid = str(f.payload.get("call_id", ""))
|
||||
out[cid] = out.get(cid, "") + str(f.payload.get("chunk", ""))
|
||||
return out
|
||||
|
||||
def tool_results_by_call(self) -> dict[str, str]:
|
||||
"""The last ``tool_result`` output seen per call_id."""
|
||||
out: dict[str, str] = {}
|
||||
for f in self.all_frames():
|
||||
if f.etype == "tool_result" and f.payload is not None:
|
||||
out[str(f.payload.get("call_id", ""))] = str(f.payload.get("output", ""))
|
||||
return out
|
||||
|
||||
def wait_for_type(self, etype: str, *, timeout: float = 45.0) -> SSEFrame:
|
||||
"""Block until a frame of ``etype`` has arrived on ANY connection."""
|
||||
deadline = time.monotonic() + timeout
|
||||
while time.monotonic() < deadline:
|
||||
for f in self.all_frames():
|
||||
if f.etype == etype:
|
||||
return f
|
||||
time.sleep(0.05)
|
||||
raise AssertionError(f"timed out waiting for a {etype!r} frame")
|
||||
|
||||
def wait_for(
|
||||
self, predicate: Callable[[BrowserlikeSSEClient], bool], *, timeout: float = 45.0
|
||||
) -> None:
|
||||
deadline = time.monotonic() + timeout
|
||||
while time.monotonic() < deadline:
|
||||
if predicate(self):
|
||||
return
|
||||
time.sleep(0.05)
|
||||
raise AssertionError("timed out waiting for predicate")
|
||||
|
||||
def wait_for_call_result(self, call_id: str, *, timeout: float = 45.0) -> None:
|
||||
self.wait_for(lambda c: call_id in c.tool_results_by_call(), timeout=timeout)
|
||||
|
||||
|
||||
def _dechunk(buf: bytes) -> tuple[bytes, bytes]:
|
||||
"""Incrementally decode HTTP/1.1 chunked transfer-encoding.
|
||||
|
||||
Consumes as many COMPLETE chunks from ``buf`` as possible and returns
|
||||
``(decoded_bytes, remainder)`` where ``remainder`` is the trailing
|
||||
partial chunk to carry into the next read. A zero-length chunk (stream
|
||||
end) simply stops consumption; the reader's ``recv`` EOF handles close.
|
||||
"""
|
||||
decoded = b""
|
||||
while True:
|
||||
if b"\r\n" not in buf:
|
||||
break # incomplete size line
|
||||
size_line, rest = buf.split(b"\r\n", 1)
|
||||
try:
|
||||
n = int(size_line.strip() or b"z", 16)
|
||||
except ValueError:
|
||||
break # malformed / partial — wait for more bytes
|
||||
if n == 0:
|
||||
break # last chunk marker
|
||||
if len(rest) < n + 2: # need n data bytes + trailing CRLF
|
||||
break
|
||||
decoded += rest[:n]
|
||||
buf = rest[n + 2 :]
|
||||
return decoded, buf
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Assertion helpers (shared by the scenarios).
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
def assert_contiguous_ids(frames: list[SSEFrame]) -> None:
|
||||
"""Every id-bearing frame in a connection forms a gap-free, dup-free,
|
||||
strictly increasing run.
|
||||
|
||||
Holds for a connection that took no ``_seq``-filtered fresh path — a
|
||||
fresh connect made before any event (snap_seq == 0) and every
|
||||
``replay_ok`` reconnect (snap_seq == 0). The server stamps a fresh
|
||||
monotonic id per enqueue with no in-ring coalescing, so a
|
||||
non-filtered consumer sees consecutive ids.
|
||||
"""
|
||||
ids = [f.event_id_int for f in frames if f.event_id_int is not None]
|
||||
assert ids, "connection carried no id-bearing frames"
|
||||
assert len(set(ids)) == len(ids), f"duplicate SSE ids: {ids}"
|
||||
assert ids == sorted(ids), f"SSE ids not monotonic: {ids}"
|
||||
for prev, cur in zip(ids, ids[1:], strict=False):
|
||||
assert cur == prev + 1, f"gap in SSE ids between {prev} and {cur}: {ids}"
|
||||
|
||||
|
||||
def assert_ids_monotonic_no_dupes(frames: list[SSEFrame]) -> None:
|
||||
"""Weaker invariant that holds on EVERY connection (including
|
||||
``_seq``-filtered fresh/truncated paths, where gaps are legal): ids
|
||||
are strictly increasing with no duplicates."""
|
||||
ids = [f.event_id_int for f in frames if f.event_id_int is not None]
|
||||
assert len(set(ids)) == len(ids), f"duplicate SSE ids: {ids}"
|
||||
assert ids == sorted(ids), f"SSE ids not monotonic: {ids}"
|
||||
|
||||
|
||||
def assert_chunk_result_ordering(frames: list[SSEFrame]) -> None:
|
||||
"""Every ``tool_output_chunk`` for a call precedes that call's own
|
||||
``tool_result`` on the wire (the load-bearing ordering — the client
|
||||
removes the streaming <pre> when it renders the result)."""
|
||||
result_index: dict[str, int] = {}
|
||||
for i, f in enumerate(frames):
|
||||
if f.etype == "tool_result" and f.payload is not None:
|
||||
result_index[str(f.payload.get("call_id", ""))] = i
|
||||
for i, f in enumerate(frames):
|
||||
if f.etype == "tool_output_chunk" and f.payload is not None:
|
||||
cid = str(f.payload.get("call_id", ""))
|
||||
assert cid in result_index, f"chunk for call {cid} has no tool_result"
|
||||
assert i < result_index[cid], (
|
||||
f"chunk for call {cid} arrived AFTER its tool_result "
|
||||
f"(chunk idx {i} >= result idx {result_index[cid]})"
|
||||
)
|
||||
|
||||
|
||||
def assert_children_stamped(frames: list[SSEFrame], parent_call_id: str) -> None:
|
||||
"""Every sub-agent child tool event carries ``parent_call_id`` (stamped
|
||||
at the flush chokepoint). Sub-tool call_ids are minted
|
||||
``{parent}::r{run}s{step}::{provider_id}`` — the ``::`` segment is the
|
||||
identifying mark — and NONE may escape unstamped to the top level."""
|
||||
unstamped: list[tuple[str | None, str, Any]] = []
|
||||
stamped = 0
|
||||
for f in frames:
|
||||
if f.payload is None:
|
||||
continue
|
||||
items = f.payload.get("items")
|
||||
entries = items if isinstance(items, list) else [f.payload]
|
||||
for entry in entries:
|
||||
if not isinstance(entry, dict):
|
||||
continue
|
||||
cid = str(entry.get("call_id", ""))
|
||||
if "::" not in cid:
|
||||
continue
|
||||
if entry.get("parent_call_id") == parent_call_id:
|
||||
stamped += 1
|
||||
else:
|
||||
unstamped.append((f.etype, cid, entry.get("parent_call_id")))
|
||||
assert stamped > 0, f"no child events found for parent {parent_call_id}"
|
||||
assert not unstamped, f"child events escaped unstamped (parent {parent_call_id}): {unstamped}"
|
||||
|
||||
|
||||
def history_tool_outputs(history_json: dict[str, Any]) -> dict[str, str]:
|
||||
"""Extract {call_id: output} from a /history projection, however the
|
||||
projection surfaces results (a folded ``output`` on a tool_call, or a
|
||||
trailing ``role: tool`` row keyed by ``tool_call_id``)."""
|
||||
out: dict[str, str] = {}
|
||||
for msg in history_json.get("messages", []):
|
||||
if not isinstance(msg, dict):
|
||||
continue
|
||||
if msg.get("role") == "tool":
|
||||
cid = msg.get("tool_call_id") or msg.get("call_id")
|
||||
if cid is not None:
|
||||
out[str(cid)] = str(msg.get("content", ""))
|
||||
for tc in msg.get("tool_calls") or ():
|
||||
if not isinstance(tc, dict):
|
||||
continue
|
||||
cid = tc.get("id") or tc.get("call_id")
|
||||
if cid is not None and tc.get("output") is not None:
|
||||
out[str(cid)] = str(tc.get("output", ""))
|
||||
return out
|
||||
|
||||
|
||||
def assert_converged(client: BrowserlikeSSEClient, history_json: dict[str, Any]) -> None:
|
||||
"""Turn-level equivalence: every tool result the client assembled live
|
||||
is present, with the same output, in a fresh /history projection.
|
||||
|
||||
Compares by call_id so a reconnect that re-delivered a result can't
|
||||
hide a divergence, and asserts the /history side isn't empty (a
|
||||
silently-lost turn would leave the projection short)."""
|
||||
live = client.tool_results_by_call()
|
||||
hist = history_tool_outputs(history_json)
|
||||
assert hist, "fresh /history projected no tool results — a turn was lost"
|
||||
for call_id, output in live.items():
|
||||
assert call_id in hist, f"call {call_id} seen live but absent from /history: {sorted(hist)}"
|
||||
assert hist[call_id] == output, (
|
||||
f"call {call_id} output diverged: live={output!r} history={hist[call_id]!r}"
|
||||
)
|
||||
@@ -1,413 +0,0 @@
|
||||
"""Boot the REAL interactive Turnstone server for the SSE recovery e2e
|
||||
harness: real ``SessionManager`` + real ``ChatSession`` engine driven
|
||||
through a scripted chat-completions client at the SDK boundary, executing
|
||||
REAL bash tools, exposed over a real uvicorn socket.
|
||||
|
||||
The recipe (verified end-to-end) has four load-bearing pieces:
|
||||
|
||||
1. **Provider injection seam.** ``create_app`` takes a PRE-BUILT
|
||||
``SessionManager``, so the harness owns the ``session_factory``: it
|
||||
passes ``client=fake_client`` and OMITS the registry, so
|
||||
``ChatSession`` falls back to ``create_provider("openai-compatible")``
|
||||
== ``OpenAIChatCompletionsProvider`` — exactly what
|
||||
``tests._session_helpers.scripted_chat_client`` targets. No production
|
||||
monkeypatch of the engine.
|
||||
|
||||
2. **Auto-title suppression.** The first user message spawns a background
|
||||
``_generate_title`` LLM call that would consume the first scripted
|
||||
response (the tool call) and desync a positional script. Setting
|
||||
``session._title_generated = True`` before the first send disables it.
|
||||
|
||||
3. **Completion barrier.** ``/send`` returns immediately after spawning
|
||||
``ws.worker_thread``; joining that thread is the true "turn complete,
|
||||
every SSE event enqueued" barrier (``stream_end`` is per-LLM-call, not
|
||||
per-turn, so it is NOT a completion marker).
|
||||
|
||||
4. **Thread hygiene.** ``create_app``'s lifespan unconditionally starts
|
||||
two daemon fan-out threads (``_global_fanout_thread`` blocking on
|
||||
``global_queue.get()``, ``_aggregate_emitter_thread`` on a 10s loop)
|
||||
with no shutdown sentinel — they would trip conftest's leaked-thread
|
||||
guard. They serve the cluster/global lane, which the per-ws ``/events``
|
||||
path under test never touches, so the harness swaps them for no-ops
|
||||
before boot (restored on ``stop``). The result is a fully clean
|
||||
teardown — no ``allow_thread_leak`` needed — with the real per-ws SSE
|
||||
engine fully intact.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import contextlib
|
||||
import json
|
||||
import queue as _q
|
||||
import socket
|
||||
import threading
|
||||
import time
|
||||
from types import SimpleNamespace
|
||||
from typing import TYPE_CHECKING, Any
|
||||
|
||||
import httpx
|
||||
import uvicorn
|
||||
|
||||
import turnstone.server as tsrv
|
||||
from tests._session_helpers import scripted_chat_client
|
||||
from turnstone.core.adapters.interactive_adapter import InteractiveAdapter
|
||||
from turnstone.core.auth import JWT_AUD_SERVER, create_jwt
|
||||
from turnstone.core.session import ChatSession
|
||||
from turnstone.core.session_manager import SessionManager
|
||||
from turnstone.core.session_ui_base import SessionUIBase
|
||||
from turnstone.core.storage import get_storage
|
||||
from turnstone.core.workstream import WorkstreamKind
|
||||
from turnstone.prompts import ClientType
|
||||
from turnstone.server import WebUI, create_app
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from turnstone.core.workstream import Workstream
|
||||
|
||||
_JWT_SECRET = "sse-recovery-e2e-jwt-secret-minimum-32-chars!"
|
||||
# Small server send buffer so a stalled consumer's in-flight backlog before
|
||||
# the listener-queue poison stays bounded (paired with the client's small
|
||||
# SO_RCVBUF in _sse_recovery_helpers). Harmless for prompt readers.
|
||||
_DEFAULT_SNDBUF = 8192
|
||||
|
||||
|
||||
def _noop_thread(*_args: object, **_kwargs: object) -> None:
|
||||
"""Stand-in for the cluster-lane daemon threads (see module docstring)."""
|
||||
|
||||
|
||||
# The REAL daemon-thread factories, captured once at import so restore always
|
||||
# targets them regardless of how many servers neuter/restore in a run (the
|
||||
# restart scenarios build a second server before the run ends).
|
||||
_REAL_FANOUT = tsrv._global_fanout_thread
|
||||
_REAL_AGGREGATE = tsrv._aggregate_emitter_thread
|
||||
|
||||
|
||||
def _fake_client(scripts: tuple[Any, ...]) -> Any:
|
||||
"""An SDK-shaped fake whose ``chat.completions.create`` follows a
|
||||
positional script (each a :func:`fake_chat_stream` kwargs dict)."""
|
||||
create_fn = scripted_chat_client(*scripts)
|
||||
client = SimpleNamespace(chat=SimpleNamespace(completions=SimpleNamespace(create=create_fn)))
|
||||
client.calls = create_fn.calls
|
||||
return client
|
||||
|
||||
|
||||
class RecoveryServer:
|
||||
"""A booted interactive node the recovery scenarios drive."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
*,
|
||||
sndbuf: int = _DEFAULT_SNDBUF,
|
||||
listener_cap: int | None = None,
|
||||
extra_routes: list[Any] | None = None,
|
||||
port: int = 0,
|
||||
) -> None:
|
||||
self._global_queue: _q.Queue[dict[str, Any]] = _q.Queue(maxsize=100000)
|
||||
self._global_listeners: list[_q.Queue[dict[str, Any]]] = []
|
||||
self._global_listeners_lock = threading.Lock()
|
||||
# Per-ws scripted client, resolved at factory-call time.
|
||||
self._pending_client: Any = _fake_client((dict(content="ok", finish_reason="stop"),))
|
||||
self._clients: dict[str, Any] = {}
|
||||
|
||||
WebUI._global_queue = self._global_queue
|
||||
|
||||
def session_factory(
|
||||
ui: Any,
|
||||
model_alias: str | None = None,
|
||||
ws_id: str | None = None,
|
||||
*,
|
||||
skill: Any = None,
|
||||
client_type: str = "",
|
||||
kind: WorkstreamKind = WorkstreamKind.INTERACTIVE,
|
||||
parent_ws_id: str | None = None,
|
||||
project_id: str = "",
|
||||
**_extra: Any,
|
||||
) -> ChatSession:
|
||||
client = self._pending_client
|
||||
if ws_id is not None:
|
||||
self._clients[ws_id] = client
|
||||
return ChatSession(
|
||||
client=client,
|
||||
model="test-model",
|
||||
ui=ui,
|
||||
instructions=None,
|
||||
temperature=None,
|
||||
max_tokens=1024,
|
||||
tool_timeout=30,
|
||||
ws_id=ws_id,
|
||||
user_id="recovery-user",
|
||||
client_type=ClientType.WEB,
|
||||
kind=kind,
|
||||
# Don't truncate large tool outputs: the harness tests
|
||||
# recovery, not the tool-result truncation budget, and a
|
||||
# truncated /history would diverge from the full live event
|
||||
# and defeat the convergence assertions.
|
||||
tool_truncation=10_000_000,
|
||||
)
|
||||
|
||||
self._adapter = InteractiveAdapter(
|
||||
global_queue=self._global_queue,
|
||||
ui_factory=lambda ws: WebUI(
|
||||
ws_id=ws.id, user_id=ws.user_id, kind=ws.kind, parent_ws_id=ws.parent_ws_id
|
||||
),
|
||||
session_factory=session_factory,
|
||||
)
|
||||
self._manager = SessionManager(
|
||||
self._adapter, storage=get_storage(), max_active=32, node_id="recovery-node"
|
||||
)
|
||||
self._adapter.attach(self._manager)
|
||||
WebUI._workstream_mgr = self._manager
|
||||
|
||||
# Neuter the cluster-lane daemons for a clean teardown (see docstring).
|
||||
tsrv._global_fanout_thread = _noop_thread
|
||||
tsrv._aggregate_emitter_thread = _noop_thread
|
||||
|
||||
# Optional small listener-queue cap. The cap is a default arg on the
|
||||
# registration methods with no config/env override, so lower it by
|
||||
# patching their ``__defaults__`` (restored on stop). fix-3's
|
||||
# de-amplification makes a real 500-cap overflow need a pathological
|
||||
# storm; a small cap exercises the identical _ListenerOverflow ->
|
||||
# stream_overflow -> reconnect-replay path within a bounded storm.
|
||||
self._orig_defaults: list[tuple[Any, tuple[Any, ...] | None]] = []
|
||||
if listener_cap is not None:
|
||||
for meth in (
|
||||
SessionUIBase._register_listener,
|
||||
SessionUIBase.register_listener_with_in_progress_snapshot,
|
||||
SessionUIBase.register_listener_with_replay,
|
||||
):
|
||||
self._orig_defaults.append((meth, meth.__defaults__))
|
||||
meth.__defaults__ = (listener_cap,)
|
||||
|
||||
self._app = create_app(
|
||||
workstreams=self._manager,
|
||||
global_queue=self._global_queue,
|
||||
global_listeners=self._global_listeners,
|
||||
global_listeners_lock=self._global_listeners_lock,
|
||||
skip_permissions=True,
|
||||
jwt_secret=_JWT_SECRET,
|
||||
node_id="recovery-node",
|
||||
# /history + tenant checks read app.state.auth_storage.
|
||||
auth_storage=get_storage(),
|
||||
)
|
||||
# Same-origin extras (Tier 2 serves its recovery page here so the real
|
||||
# Pane's cookie auth + EventSource work without cross-origin plumbing).
|
||||
if extra_routes:
|
||||
self._app.router.routes.extend(extra_routes)
|
||||
|
||||
# Pre-bind a listening socket with a small SO_SNDBUF (accepted conns
|
||||
# inherit it), then hand it to uvicorn.
|
||||
self._sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
|
||||
self._sock.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
|
||||
self._sock.setsockopt(socket.SOL_SOCKET, socket.SO_SNDBUF, sndbuf)
|
||||
self._sock.bind(("127.0.0.1", port)) # port=0 -> ephemeral; fixed -> restart reuse
|
||||
self._port = int(self._sock.getsockname()[1])
|
||||
self._sock.listen(128)
|
||||
|
||||
self._server = uvicorn.Server(uvicorn.Config(self._app, log_level="warning", lifespan="on"))
|
||||
self._thread = threading.Thread(
|
||||
target=self._serve, name=f"uvicorn-recovery-{self._port}", daemon=True
|
||||
)
|
||||
self._thread.start()
|
||||
if not _tcp_ready(self._port, 10.0):
|
||||
self.stop()
|
||||
raise AssertionError("recovery server did not accept TCP")
|
||||
|
||||
self._token = create_jwt(
|
||||
user_id="recovery-user",
|
||||
scopes=frozenset({"read", "write", "approve", "service"}),
|
||||
source="recovery",
|
||||
secret=_JWT_SECRET,
|
||||
audience=JWT_AUD_SERVER,
|
||||
)
|
||||
self._http = httpx.Client(base_url=self.base_url, timeout=httpx.Timeout(30.0))
|
||||
|
||||
def _serve(self) -> None:
|
||||
loop = asyncio.new_event_loop()
|
||||
asyncio.set_event_loop(loop)
|
||||
try:
|
||||
loop.run_until_complete(self._server.serve(sockets=[self._sock]))
|
||||
finally:
|
||||
pending = asyncio.all_tasks(loop)
|
||||
for task in pending:
|
||||
task.cancel()
|
||||
if pending:
|
||||
with contextlib.suppress(Exception):
|
||||
loop.run_until_complete(asyncio.gather(*pending, return_exceptions=True))
|
||||
loop.close()
|
||||
|
||||
# -- properties ----------------------------------------------------------
|
||||
|
||||
@property
|
||||
def base_url(self) -> str:
|
||||
return f"http://127.0.0.1:{self._port}"
|
||||
|
||||
@property
|
||||
def token(self) -> str:
|
||||
return self._token
|
||||
|
||||
@property
|
||||
def manager(self) -> SessionManager:
|
||||
return self._manager
|
||||
|
||||
# -- workstream lifecycle ------------------------------------------------
|
||||
|
||||
def create_workstream(self, *scripts: Any, name: str = "recovery-ws") -> str:
|
||||
"""Create a ws whose scripted LLM follows ``scripts`` (positional
|
||||
:func:`fake_chat_stream` kwargs). Auto-approves tools and suppresses
|
||||
the auto-title call so the positional script stays in sync."""
|
||||
self._pending_client = _fake_client(scripts)
|
||||
ws = self._manager.create(user_id="recovery-user", name=name)
|
||||
self._prime_ws(ws)
|
||||
return ws.id
|
||||
|
||||
def open_workstream(self, ws_id: str, *scripts: Any) -> None:
|
||||
"""Rehydrate a persisted ws on THIS node (the restart path). Fresh
|
||||
UI → empty ring + storage-seeded ``_event_id``."""
|
||||
if scripts:
|
||||
self._pending_client = _fake_client(scripts)
|
||||
ws = self._manager.open(ws_id)
|
||||
if ws is None:
|
||||
raise AssertionError(f"open_workstream: ws {ws_id} not resurrectable")
|
||||
self._prime_ws(ws)
|
||||
|
||||
def _prime_ws(self, ws: Workstream) -> None:
|
||||
if isinstance(ws.ui, SessionUIBase):
|
||||
ws.ui.auto_approve = True # blanket tool auto-approval
|
||||
if ws.session is not None:
|
||||
ws.session._title_generated = True # suppress the auto-title LLM call
|
||||
|
||||
def send(self, ws_id: str, message: str = "go") -> None:
|
||||
"""POST /send — spawns the worker thread and returns immediately."""
|
||||
r = self._http.post(
|
||||
f"/v1/api/workstreams/{ws_id}/send",
|
||||
headers={"Authorization": f"Bearer {self._token}"},
|
||||
json={"message": message},
|
||||
)
|
||||
r.raise_for_status()
|
||||
|
||||
def wait_turn(self, ws_id: str, *, timeout: float = 45.0) -> None:
|
||||
"""Block until the turn's worker thread finishes (the true
|
||||
turn-complete barrier) and the ws is idle."""
|
||||
deadline = time.monotonic() + timeout
|
||||
worker: threading.Thread | None = None
|
||||
while time.monotonic() < deadline:
|
||||
ws = self._manager.get(ws_id)
|
||||
worker = ws.worker_thread if ws is not None else None
|
||||
if worker is not None:
|
||||
break
|
||||
time.sleep(0.02)
|
||||
if worker is not None:
|
||||
worker.join(timeout=max(0.5, deadline - time.monotonic()))
|
||||
if worker.is_alive():
|
||||
raise AssertionError(f"turn worker for {ws_id} did not finish in {timeout}s")
|
||||
|
||||
def get_ws(self, ws_id: str) -> Workstream | None:
|
||||
return self._manager.get(ws_id)
|
||||
|
||||
def ws_state(self, ws_id: str) -> str:
|
||||
ws = self._manager.get(ws_id)
|
||||
return ws.state.value if ws is not None else ""
|
||||
|
||||
def ring_span(self, ws_id: str) -> tuple[int | None, int]:
|
||||
"""(earliest retained ring event_id or None, latest counter) — lets a
|
||||
scenario wait for the ring to evict a specific cursor."""
|
||||
ws = self._manager.get(ws_id)
|
||||
ui = ws.ui if ws is not None else None
|
||||
if not isinstance(ui, SessionUIBase):
|
||||
return None, 0
|
||||
buf = ui._event_buffer
|
||||
earliest = buf[0][0] if buf else None
|
||||
return earliest, ui._event_id
|
||||
|
||||
def listener_poisoned(self, ws_id: str) -> bool:
|
||||
"""True once any live SSE listener on the ws has poisoned (overflow)."""
|
||||
ws = self._manager.get(ws_id)
|
||||
ui = ws.ui if ws is not None else None
|
||||
if not isinstance(ui, SessionUIBase):
|
||||
return False
|
||||
return any(getattr(q, "poisoned", False) for q in list(ui._listeners))
|
||||
|
||||
def max_event_id(self, ws_id: str) -> int | None:
|
||||
"""The storage high-water ``MAX(conversations.event_id)`` — what a
|
||||
restarted node's fresh UI seeds ``_event_id`` from."""
|
||||
result: int | None = get_storage().get_max_event_id(ws_id)
|
||||
return result
|
||||
|
||||
def fetch_history(self, ws_id: str) -> dict[str, Any]:
|
||||
r = self._http.get(
|
||||
f"/v1/api/workstreams/{ws_id}/history",
|
||||
headers={"Authorization": f"Bearer {self._token}"},
|
||||
)
|
||||
r.raise_for_status()
|
||||
result: dict[str, Any] = r.json()
|
||||
return result
|
||||
|
||||
# -- teardown ------------------------------------------------------------
|
||||
|
||||
def stop(self) -> None:
|
||||
with contextlib.suppress(Exception):
|
||||
for ws in list(self._manager.list_all()):
|
||||
with contextlib.suppress(Exception):
|
||||
self._manager.close(ws.id)
|
||||
self._server.should_exit = True
|
||||
self._thread.join(timeout=20)
|
||||
with contextlib.suppress(Exception):
|
||||
self._http.close()
|
||||
with contextlib.suppress(OSError):
|
||||
self._sock.close()
|
||||
# Restore the cluster-lane daemon factories + any patched cap defaults.
|
||||
tsrv._global_fanout_thread = _REAL_FANOUT
|
||||
tsrv._aggregate_emitter_thread = _REAL_AGGREGATE
|
||||
for meth, defaults in self._orig_defaults:
|
||||
meth.__defaults__ = defaults
|
||||
|
||||
|
||||
def _tcp_ready(port: int, timeout: float) -> bool:
|
||||
end = time.monotonic() + timeout
|
||||
while time.monotonic() < end:
|
||||
try:
|
||||
with socket.create_connection(("127.0.0.1", port), timeout=0.3):
|
||||
return True
|
||||
except OSError:
|
||||
time.sleep(0.05)
|
||||
return False
|
||||
|
||||
|
||||
def bash_toolcall_script(
|
||||
call_id: str, command: str, *, finish_reason: str = "tool_calls"
|
||||
) -> dict[str, Any]:
|
||||
"""A scripted assistant turn issuing ONE bash tool call."""
|
||||
return dict(
|
||||
tool_calls=[{"id": call_id, "name": "bash", "arguments": json.dumps({"command": command})}],
|
||||
finish_reason=finish_reason,
|
||||
)
|
||||
|
||||
|
||||
def parallel_bash_script(commands: dict[str, str]) -> dict[str, Any]:
|
||||
"""A scripted assistant turn issuing SEVERAL bash tool calls at once
|
||||
(the parallel-pool storm), ``{call_id: command}``.
|
||||
|
||||
Each command is prefixed with a no-op ``: <call_id>;`` so the tool
|
||||
ARGUMENTS are distinct per call while the OUTPUT is unchanged (``:``
|
||||
ignores its args and prints nothing). Identical-argument parallel
|
||||
calls otherwise trip the session's repeat-tool-call guard, which
|
||||
appends a warning to the PERSISTED result only (not the live event) —
|
||||
an orthogonal divergence that would mask the recovery behavior the
|
||||
convergence assertions test.
|
||||
"""
|
||||
return dict(
|
||||
tool_calls=[
|
||||
{
|
||||
"id": cid,
|
||||
"name": "bash",
|
||||
"arguments": json.dumps({"command": f": {cid}; {cmd}"}),
|
||||
}
|
||||
for cid, cmd in commands.items()
|
||||
],
|
||||
finish_reason="tool_calls",
|
||||
)
|
||||
|
||||
|
||||
def final_text_script(content: str = "done") -> dict[str, Any]:
|
||||
"""The scripted assistant turn that ends the agent loop (no tools)."""
|
||||
return dict(content=content, finish_reason="stop")
|
||||
@@ -1,76 +0,0 @@
|
||||
"""Recording fake SDK client — captures the kwargs at each provider's seam.
|
||||
|
||||
Every provider's ``create_streaming`` assembles its kwargs and calls the
|
||||
SDK *eagerly* before returning the stream iterator (Anthropic
|
||||
``client.messages.stream``, OpenAI ``client.chat.completions.create``,
|
||||
Responses ``client.responses.create/stream``), so driving a provider
|
||||
against a :class:`RecordingClient` captures the full composed request
|
||||
payload without a network round-trip.
|
||||
|
||||
Shared by the wire-payload golden harness (``test_wire_payload_golden``)
|
||||
and the effort-ladder parity harness (``test_effort_ladder_wire_parity``)
|
||||
so both assert against the same capture seam.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import TYPE_CHECKING, Any
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from collections.abc import Iterator
|
||||
|
||||
|
||||
class _EmptyStream:
|
||||
"""Stand-in for an SDK stream / stream-manager: empty iterable AND no-op CM."""
|
||||
|
||||
def __iter__(self) -> Iterator[Any]:
|
||||
return iter(())
|
||||
|
||||
def __enter__(self) -> _EmptyStream:
|
||||
return self
|
||||
|
||||
def __exit__(self, *exc: object) -> None:
|
||||
return None
|
||||
|
||||
|
||||
class _Seam:
|
||||
"""Records the kwargs of a single SDK call, returns an empty stream stub."""
|
||||
|
||||
def __init__(self, sink: dict[str, Any]) -> None:
|
||||
self._sink = sink
|
||||
|
||||
def __call__(self, **kwargs: Any) -> _EmptyStream:
|
||||
# Last write wins; only one seam is exercised per provider call.
|
||||
self._sink["payload"] = kwargs
|
||||
return _EmptyStream()
|
||||
|
||||
|
||||
class _Completions:
|
||||
def __init__(self, sink: dict[str, Any]) -> None:
|
||||
self.create = _Seam(sink)
|
||||
|
||||
|
||||
class _Chat:
|
||||
def __init__(self, sink: dict[str, Any]) -> None:
|
||||
self.completions = _Completions(sink)
|
||||
|
||||
|
||||
class _Messages:
|
||||
def __init__(self, sink: dict[str, Any]) -> None:
|
||||
self.stream = _Seam(sink)
|
||||
|
||||
|
||||
class _Responses:
|
||||
def __init__(self, sink: dict[str, Any]) -> None:
|
||||
self.create = _Seam(sink)
|
||||
self.stream = _Seam(sink)
|
||||
|
||||
|
||||
class RecordingClient:
|
||||
"""Fake SDK client exposing every provider's call seam, recording kwargs."""
|
||||
|
||||
def __init__(self) -> None:
|
||||
self.captured: dict[str, Any] = {}
|
||||
self.messages = _Messages(self.captured)
|
||||
self.chat = _Chat(self.captured)
|
||||
self.responses = _Responses(self.captured)
|
||||
+1
-198
@@ -4,9 +4,6 @@ import asyncio
|
||||
import contextlib
|
||||
import logging
|
||||
import os
|
||||
import socket
|
||||
import subprocess
|
||||
import sys
|
||||
import threading
|
||||
import time
|
||||
from typing import TYPE_CHECKING, Any
|
||||
@@ -55,76 +52,8 @@ def serve_until_exit(server: Any) -> None:
|
||||
loop.close()
|
||||
|
||||
|
||||
class _PendingResolver:
|
||||
"""Race-free drop-in for ``threading.Timer(delay, ui.resolve_approval)``.
|
||||
|
||||
``approve_tools`` runs ``_approval_event.clear()`` -> register
|
||||
``_pending_approval`` -> ``_approval_event.wait(_APPROVAL_WAIT_TIMEOUT)``
|
||||
(3600s). A *fixed-delay* timer can fire ``resolve_approval``
|
||||
(``_approval_event.set()``) BEFORE that ``.clear()`` on a slow/loaded
|
||||
runner, so the set is wiped by the clear and ``approve_tools`` blocks the
|
||||
full hour -- surfacing as a CI hang. This instead waits until the approval
|
||||
is actually registered (which happens *after* the clear), then resolves, so
|
||||
the wakeup can never be lost. ``start()`` / ``cancel()`` mirror
|
||||
``threading.Timer`` so it drops into existing scaffolding. ``cancel()``
|
||||
signals the worker to stop and joins it, so a test that errors *before* the
|
||||
approval registers can't leak the thread or resolve late into a finished
|
||||
test. ``before`` runs just before resolving -- e.g. to snapshot
|
||||
pending-state fields the test asserts on.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
ui: Any,
|
||||
*args: Any,
|
||||
before: Callable[[], None] | None = None,
|
||||
deadline: float = 10.0,
|
||||
**kwargs: Any,
|
||||
) -> None:
|
||||
self._ui = ui
|
||||
self._args = args
|
||||
self._kwargs = kwargs
|
||||
self._before = before
|
||||
self._deadline = deadline
|
||||
self._cancelled = threading.Event()
|
||||
self._started = False
|
||||
self._thread = threading.Thread(target=self._run, name="resolve-when-pending", daemon=True)
|
||||
|
||||
def _run(self) -> None:
|
||||
end = time.monotonic() + self._deadline
|
||||
while time.monotonic() < end:
|
||||
if self._cancelled.is_set():
|
||||
return
|
||||
# getattr (not a bare read) so a UI without _pending_approval can't
|
||||
# crash the worker into a silent death that leaves approve_tools
|
||||
# blocked for the full _APPROVAL_WAIT_TIMEOUT.
|
||||
if getattr(self._ui, "_pending_approval", None) is not None:
|
||||
if self._before is not None:
|
||||
self._before()
|
||||
self._ui.resolve_approval(*self._args, **self._kwargs)
|
||||
return
|
||||
time.sleep(0.001)
|
||||
# Deadline without registration: approve_tools isn't parked on the
|
||||
# approval event (returned early, or never reached it) -- don't resolve
|
||||
# into an unknown state; let the test's own assertions speak.
|
||||
|
||||
def start(self) -> None:
|
||||
self._started = True
|
||||
self._thread.start()
|
||||
|
||||
def cancel(self) -> None:
|
||||
self._cancelled.set()
|
||||
if self._started:
|
||||
self._thread.join(timeout=5)
|
||||
|
||||
|
||||
def resolve_when_pending(ui: Any, *args: Any, **kwargs: Any) -> _PendingResolver:
|
||||
"""Build a race-free approval resolver (see :class:`_PendingResolver`)."""
|
||||
return _PendingResolver(ui, *args, **kwargs)
|
||||
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from collections.abc import Callable, Iterator
|
||||
from collections.abc import Iterator
|
||||
|
||||
from turnstone.core.mcp_client import MCPClientManager, StaticServerState
|
||||
from turnstone.core.mcp_crypto import MCPTokenCipher
|
||||
@@ -219,98 +148,6 @@ def _seed_static_state(mgr: MCPClientManager, name: str, **overrides: Any) -> St
|
||||
return state
|
||||
|
||||
|
||||
def _run_on_loop(loop: asyncio.AbstractEventLoop, coro: Any, timeout: float = 10) -> Any:
|
||||
"""Submit *coro* to *loop*, wait for the result.
|
||||
|
||||
The ONE copy shared by the MCP test files — four hand-synced copies
|
||||
had already drifted on the timeout (5s hardcoded vs a 10s default).
|
||||
The timeout is an upper bound on waiting, not a behavior assertion,
|
||||
so the most generous variant won the merge.
|
||||
"""
|
||||
fut = asyncio.run_coroutine_threadsafe(coro, loop)
|
||||
return fut.result(timeout=timeout)
|
||||
|
||||
|
||||
def _drain_background(mgr: MCPClientManager, loop: asyncio.AbstractEventLoop) -> None:
|
||||
"""Deterministically await ``mgr``'s tracked background tasks.
|
||||
|
||||
Replaces fixed sleeps for synchronizing with scheduled dead-grant
|
||||
drops / spawned refreshes: exact, and immune to slow-runner flake.
|
||||
"""
|
||||
|
||||
async def _drain() -> None:
|
||||
tasks = [t for t in list(mgr._background_tasks) if not t.done()]
|
||||
if tasks:
|
||||
await asyncio.gather(*tasks, return_exceptions=True)
|
||||
|
||||
_run_on_loop(loop, _drain())
|
||||
|
||||
|
||||
def _poll_until(predicate: Callable[[], bool], timeout: float, interval: float = 0.05) -> bool:
|
||||
"""Poll *predicate* until true or *timeout* elapses — the ONE wait loop.
|
||||
|
||||
Shared by the live MCP smoke tests' condition helpers so the
|
||||
deadline/poll pattern doesn't accrete per-file hand-synced copies.
|
||||
"""
|
||||
deadline = time.monotonic() + timeout
|
||||
while time.monotonic() < deadline:
|
||||
if predicate():
|
||||
return True
|
||||
time.sleep(interval)
|
||||
return False
|
||||
|
||||
|
||||
def _free_port() -> int:
|
||||
"""Grab an ephemeral localhost port for a live-server subprocess.
|
||||
|
||||
Shared by the live MCP smoke tests (flaky-server, push-refresh) so
|
||||
the socket-probe helpers stay in one place instead of drifting per
|
||||
file.
|
||||
"""
|
||||
with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
|
||||
s.bind(("127.0.0.1", 0))
|
||||
return int(s.getsockname()[1])
|
||||
|
||||
|
||||
def _tcp_accepts(port: int) -> bool:
|
||||
try:
|
||||
with socket.create_connection(("127.0.0.1", port), timeout=0.3):
|
||||
return True
|
||||
except OSError:
|
||||
return False
|
||||
|
||||
|
||||
def _wait_tcp_ready(port: int, timeout: float) -> bool:
|
||||
"""Poll until something accepts TCP on 127.0.0.1:*port* (live tests)."""
|
||||
return _poll_until(lambda: _tcp_accepts(port), timeout)
|
||||
|
||||
|
||||
def _wait_session_live(mgr: MCPClientManager, name: str, timeout: float) -> bool:
|
||||
"""Poll until static server *name* has a live session (live tests)."""
|
||||
|
||||
def _live() -> bool:
|
||||
state = mgr._static_servers.get(name)
|
||||
return state is not None and state.session is not None
|
||||
|
||||
return _poll_until(_live, timeout)
|
||||
|
||||
|
||||
def _popen_mcp_server(script_path: Any, port: int) -> subprocess.Popen[bytes]:
|
||||
"""Start a FastMCP live-server subprocess, streams to DEVNULL.
|
||||
|
||||
The shared spawn primitive for the live MCP smoke tests
|
||||
(flaky-server flap loop, push-refresh) — the readiness wait and the
|
||||
skip-vs-raise-on-failure policy legitimately differ per test and
|
||||
stay at the call sites. ``sys.executable`` runs the same interpreter,
|
||||
so a server-side import gap surfaces as a failed TCP wait, not here.
|
||||
"""
|
||||
return subprocess.Popen(
|
||||
[sys.executable, str(script_path), str(port)],
|
||||
stdout=subprocess.DEVNULL,
|
||||
stderr=subprocess.DEVNULL,
|
||||
)
|
||||
|
||||
|
||||
def make_oidc_test_config(**overrides: Any) -> OIDCConfig:
|
||||
"""Build a test ``OIDCConfig`` with sensible defaults.
|
||||
|
||||
@@ -430,40 +267,6 @@ def mock_openai_client():
|
||||
return client
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def make_config_store():
|
||||
"""Factory for a lightweight ConfigStore double.
|
||||
|
||||
``make_config_store(**overrides)`` returns an object whose ``.get(key)``
|
||||
yields the override when present, else the registered SettingDef default —
|
||||
mirroring the real :meth:`ConfigStore.get` fail-open (a bool setting reads
|
||||
as its ``False`` default on a miss, never ``None``). Shared by the
|
||||
``server.require_project`` gate / advisory tests.
|
||||
"""
|
||||
|
||||
_unset = object()
|
||||
|
||||
def _make(**overrides: Any) -> Any:
|
||||
from turnstone.core.settings_registry import SETTINGS
|
||||
|
||||
class _ConfigStoreDouble:
|
||||
def get(self, key: str, default: Any = _unset) -> Any:
|
||||
# Mirror ConfigStore.get precedence exactly: cache (overrides)
|
||||
# first, then a caller-supplied default, then the registry
|
||||
# default, then None — so a reused caller passing an explicit
|
||||
# default for an unset key gets the same value production would.
|
||||
if key in overrides:
|
||||
return overrides[key]
|
||||
if default is not _unset:
|
||||
return default
|
||||
defn = SETTINGS.get(key)
|
||||
return defn.default if defn else None
|
||||
|
||||
return _ConfigStoreDouble()
|
||||
|
||||
return _make
|
||||
|
||||
|
||||
@pytest.fixture(autouse=True)
|
||||
def _clear_policy_cache():
|
||||
"""Drop the in-process tool-policy cache between tests.
|
||||
|
||||
@@ -1,77 +0,0 @@
|
||||
{
|
||||
"cache_control": {
|
||||
"type": "ephemeral"
|
||||
},
|
||||
"extra_body": {
|
||||
"chat_template_kwargs": {
|
||||
"enable_thinking": true,
|
||||
"reasoning_effort": "high"
|
||||
}
|
||||
},
|
||||
"max_tokens": 4096,
|
||||
"messages": [
|
||||
{
|
||||
"content": "Weather in Paris and London?",
|
||||
"role": "user"
|
||||
},
|
||||
{
|
||||
"content": [
|
||||
{
|
||||
"id": "call_1",
|
||||
"input": {
|
||||
"city": "Paris"
|
||||
},
|
||||
"name": "get_weather",
|
||||
"type": "tool_use"
|
||||
},
|
||||
{
|
||||
"id": "call_2",
|
||||
"input": {
|
||||
"city": "London"
|
||||
},
|
||||
"name": "get_weather",
|
||||
"type": "tool_use"
|
||||
}
|
||||
],
|
||||
"role": "assistant"
|
||||
},
|
||||
{
|
||||
"content": [
|
||||
{
|
||||
"content": "18C, clear.",
|
||||
"tool_use_id": "call_1",
|
||||
"type": "tool_result"
|
||||
},
|
||||
{
|
||||
"content": "Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
|
||||
"is_error": true,
|
||||
"tool_use_id": "call_2",
|
||||
"type": "tool_result"
|
||||
},
|
||||
{
|
||||
"text": "Actually, never mind London.",
|
||||
"type": "text"
|
||||
}
|
||||
],
|
||||
"role": "user"
|
||||
}
|
||||
],
|
||||
"model": "qwen3.6-27b",
|
||||
"tools": [
|
||||
{
|
||||
"description": "Look up the weather for a city.",
|
||||
"input_schema": {
|
||||
"properties": {
|
||||
"city": {
|
||||
"type": "string"
|
||||
}
|
||||
},
|
||||
"required": [
|
||||
"city"
|
||||
],
|
||||
"type": "object"
|
||||
},
|
||||
"name": "get_weather"
|
||||
}
|
||||
]
|
||||
}
|
||||
@@ -1,32 +0,0 @@
|
||||
{
|
||||
"cache_control": {
|
||||
"type": "ephemeral"
|
||||
},
|
||||
"extra_body": {
|
||||
"chat_template_kwargs": {
|
||||
"enable_thinking": true,
|
||||
"reasoning_effort": "high"
|
||||
}
|
||||
},
|
||||
"max_tokens": 4096,
|
||||
"messages": [
|
||||
{
|
||||
"content": [
|
||||
{
|
||||
"text": "What's in this image?",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
"source": {
|
||||
"data": "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mP8z8BQDwAEhQGAhKmMIQAAAABJRU5ErkJggg==",
|
||||
"media_type": "image/png",
|
||||
"type": "base64"
|
||||
},
|
||||
"type": "image"
|
||||
}
|
||||
],
|
||||
"role": "user"
|
||||
}
|
||||
],
|
||||
"model": "qwen3.6-27b"
|
||||
}
|
||||
@@ -1,69 +0,0 @@
|
||||
{
|
||||
"cache_control": {
|
||||
"type": "ephemeral"
|
||||
},
|
||||
"extra_body": {
|
||||
"chat_template_kwargs": {
|
||||
"enable_thinking": true,
|
||||
"reasoning_effort": "high"
|
||||
}
|
||||
},
|
||||
"max_tokens": 4096,
|
||||
"messages": [
|
||||
{
|
||||
"content": "Think about the weather.",
|
||||
"role": "user"
|
||||
},
|
||||
{
|
||||
"content": [
|
||||
{
|
||||
"signature": "sig-abc",
|
||||
"thinking": "The user wants weather.",
|
||||
"type": "thinking"
|
||||
},
|
||||
{
|
||||
"text": "Let me check.",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
"id": "call_1",
|
||||
"input": {
|
||||
"city": "Paris"
|
||||
},
|
||||
"name": "get_weather",
|
||||
"type": "tool_use"
|
||||
}
|
||||
],
|
||||
"role": "assistant"
|
||||
},
|
||||
{
|
||||
"content": [
|
||||
{
|
||||
"content": "Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
|
||||
"is_error": true,
|
||||
"tool_use_id": "call_1",
|
||||
"type": "tool_result"
|
||||
}
|
||||
],
|
||||
"role": "user"
|
||||
}
|
||||
],
|
||||
"model": "qwen3.6-27b",
|
||||
"tools": [
|
||||
{
|
||||
"description": "Look up the weather for a city.",
|
||||
"input_schema": {
|
||||
"properties": {
|
||||
"city": {
|
||||
"type": "string"
|
||||
}
|
||||
},
|
||||
"required": [
|
||||
"city"
|
||||
],
|
||||
"type": "object"
|
||||
},
|
||||
"name": "get_weather"
|
||||
}
|
||||
]
|
||||
}
|
||||
@@ -1,68 +0,0 @@
|
||||
{
|
||||
"cache_control": {
|
||||
"type": "ephemeral"
|
||||
},
|
||||
"extra_body": {
|
||||
"chat_template_kwargs": {
|
||||
"enable_thinking": true,
|
||||
"reasoning_effort": "high"
|
||||
}
|
||||
},
|
||||
"max_tokens": 4096,
|
||||
"messages": [
|
||||
{
|
||||
"content": "Think about the weather.",
|
||||
"role": "user"
|
||||
},
|
||||
{
|
||||
"content": [
|
||||
{
|
||||
"signature": "sig-abc",
|
||||
"thinking": "The user wants weather.",
|
||||
"type": "thinking"
|
||||
},
|
||||
{
|
||||
"text": "Let me check.",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
"id": "call_1",
|
||||
"input": {
|
||||
"city": "Paris"
|
||||
},
|
||||
"name": "get_weather",
|
||||
"type": "tool_use"
|
||||
}
|
||||
],
|
||||
"role": "assistant"
|
||||
},
|
||||
{
|
||||
"content": [
|
||||
{
|
||||
"content": "18C, clear.",
|
||||
"tool_use_id": "call_1",
|
||||
"type": "tool_result"
|
||||
}
|
||||
],
|
||||
"role": "user"
|
||||
}
|
||||
],
|
||||
"model": "qwen3.6-27b",
|
||||
"tools": [
|
||||
{
|
||||
"description": "Look up the weather for a city.",
|
||||
"input_schema": {
|
||||
"properties": {
|
||||
"city": {
|
||||
"type": "string"
|
||||
}
|
||||
},
|
||||
"required": [
|
||||
"city"
|
||||
],
|
||||
"type": "object"
|
||||
},
|
||||
"name": "get_weather"
|
||||
}
|
||||
]
|
||||
}
|
||||
@@ -1,62 +0,0 @@
|
||||
{
|
||||
"cache_control": {
|
||||
"type": "ephemeral"
|
||||
},
|
||||
"extra_body": {
|
||||
"chat_template_kwargs": {
|
||||
"enable_thinking": true,
|
||||
"reasoning_effort": "high"
|
||||
}
|
||||
},
|
||||
"max_tokens": 4096,
|
||||
"messages": [
|
||||
{
|
||||
"content": "Run the deploy.",
|
||||
"role": "user"
|
||||
},
|
||||
{
|
||||
"content": [
|
||||
{
|
||||
"id": "call_1",
|
||||
"input": {},
|
||||
"name": "deploy",
|
||||
"type": "tool_use"
|
||||
}
|
||||
],
|
||||
"role": "assistant"
|
||||
},
|
||||
{
|
||||
"content": [
|
||||
{
|
||||
"content": "deployed",
|
||||
"tool_use_id": "call_1",
|
||||
"type": "tool_result"
|
||||
},
|
||||
{
|
||||
"text": "Great, what's next?",
|
||||
"type": "text"
|
||||
}
|
||||
],
|
||||
"role": "user"
|
||||
}
|
||||
],
|
||||
"model": "qwen3.6-27b",
|
||||
"system": "Output-guard: deploy output looked clean.",
|
||||
"tools": [
|
||||
{
|
||||
"description": "Look up the weather for a city.",
|
||||
"input_schema": {
|
||||
"properties": {
|
||||
"city": {
|
||||
"type": "string"
|
||||
}
|
||||
},
|
||||
"required": [
|
||||
"city"
|
||||
],
|
||||
"type": "object"
|
||||
},
|
||||
"name": "get_weather"
|
||||
}
|
||||
]
|
||||
}
|
||||
@@ -1,32 +0,0 @@
|
||||
{
|
||||
"cache_control": {
|
||||
"type": "ephemeral"
|
||||
},
|
||||
"extra_body": {
|
||||
"chat_template_kwargs": {
|
||||
"enable_thinking": true,
|
||||
"reasoning_effort": "high"
|
||||
}
|
||||
},
|
||||
"max_tokens": 4096,
|
||||
"messages": [
|
||||
{
|
||||
"content": "Hi there.",
|
||||
"role": "user"
|
||||
},
|
||||
{
|
||||
"content": [
|
||||
{
|
||||
"text": "Hello! How can I help?",
|
||||
"type": "text"
|
||||
}
|
||||
],
|
||||
"role": "assistant"
|
||||
},
|
||||
{
|
||||
"content": "What's the weather in Paris?",
|
||||
"role": "user"
|
||||
}
|
||||
],
|
||||
"model": "qwen3.6-27b"
|
||||
}
|
||||
@@ -1,68 +0,0 @@
|
||||
{
|
||||
"cache_control": {
|
||||
"type": "ephemeral"
|
||||
},
|
||||
"extra_body": {
|
||||
"chat_template_kwargs": {
|
||||
"enable_thinking": true,
|
||||
"reasoning_effort": "high"
|
||||
}
|
||||
},
|
||||
"max_tokens": 4096,
|
||||
"messages": [
|
||||
{
|
||||
"content": "Weather in Paris?",
|
||||
"role": "user"
|
||||
},
|
||||
{
|
||||
"content": [
|
||||
{
|
||||
"id": "call_1",
|
||||
"input": {
|
||||
"city": "Paris"
|
||||
},
|
||||
"name": "get_weather",
|
||||
"type": "tool_use"
|
||||
}
|
||||
],
|
||||
"role": "assistant"
|
||||
},
|
||||
{
|
||||
"content": [
|
||||
{
|
||||
"content": "18C, clear.",
|
||||
"tool_use_id": "call_1",
|
||||
"type": "tool_result"
|
||||
}
|
||||
],
|
||||
"role": "user"
|
||||
},
|
||||
{
|
||||
"content": [
|
||||
{
|
||||
"text": "It's 18C and clear in Paris.",
|
||||
"type": "text"
|
||||
}
|
||||
],
|
||||
"role": "assistant"
|
||||
}
|
||||
],
|
||||
"model": "qwen3.6-27b",
|
||||
"tools": [
|
||||
{
|
||||
"description": "Look up the weather for a city.",
|
||||
"input_schema": {
|
||||
"properties": {
|
||||
"city": {
|
||||
"type": "string"
|
||||
}
|
||||
},
|
||||
"required": [
|
||||
"city"
|
||||
],
|
||||
"type": "object"
|
||||
},
|
||||
"name": "get_weather"
|
||||
}
|
||||
]
|
||||
}
|
||||
@@ -1,60 +0,0 @@
|
||||
{
|
||||
"cache_control": {
|
||||
"type": "ephemeral"
|
||||
},
|
||||
"extra_body": {
|
||||
"chat_template_kwargs": {
|
||||
"enable_thinking": true,
|
||||
"reasoning_effort": "high"
|
||||
}
|
||||
},
|
||||
"max_tokens": 4096,
|
||||
"messages": [
|
||||
{
|
||||
"content": "Weather in Paris?",
|
||||
"role": "user"
|
||||
},
|
||||
{
|
||||
"content": [
|
||||
{
|
||||
"id": "call_1",
|
||||
"input": {
|
||||
"city": "Paris"
|
||||
},
|
||||
"name": "get_weather",
|
||||
"type": "tool_use"
|
||||
}
|
||||
],
|
||||
"role": "assistant"
|
||||
},
|
||||
{
|
||||
"content": [
|
||||
{
|
||||
"content": "Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
|
||||
"is_error": true,
|
||||
"tool_use_id": "call_1",
|
||||
"type": "tool_result"
|
||||
}
|
||||
],
|
||||
"role": "user"
|
||||
}
|
||||
],
|
||||
"model": "qwen3.6-27b",
|
||||
"tools": [
|
||||
{
|
||||
"description": "Look up the weather for a city.",
|
||||
"input_schema": {
|
||||
"properties": {
|
||||
"city": {
|
||||
"type": "string"
|
||||
}
|
||||
},
|
||||
"required": [
|
||||
"city"
|
||||
],
|
||||
"type": "object"
|
||||
},
|
||||
"name": "get_weather"
|
||||
}
|
||||
]
|
||||
}
|
||||
@@ -51,6 +51,9 @@
|
||||
}
|
||||
],
|
||||
"model": "claude-sonnet-4-6",
|
||||
"output_config": {
|
||||
"effort": "medium"
|
||||
},
|
||||
"temperature": 1.0,
|
||||
"thinking": {
|
||||
"type": "adaptive"
|
||||
|
||||
@@ -23,6 +23,9 @@
|
||||
}
|
||||
],
|
||||
"model": "claude-sonnet-4-6",
|
||||
"output_config": {
|
||||
"effort": "medium"
|
||||
},
|
||||
"temperature": 1.0,
|
||||
"thinking": {
|
||||
"type": "adaptive"
|
||||
|
||||
@@ -43,6 +43,9 @@
|
||||
}
|
||||
],
|
||||
"model": "claude-sonnet-4-6",
|
||||
"output_config": {
|
||||
"effort": "medium"
|
||||
},
|
||||
"temperature": 1.0,
|
||||
"thinking": {
|
||||
"type": "adaptive"
|
||||
|
||||
@@ -42,6 +42,9 @@
|
||||
}
|
||||
],
|
||||
"model": "claude-sonnet-4-6",
|
||||
"output_config": {
|
||||
"effort": "medium"
|
||||
},
|
||||
"temperature": 1.0,
|
||||
"thinking": {
|
||||
"type": "adaptive"
|
||||
|
||||
@@ -35,6 +35,9 @@
|
||||
}
|
||||
],
|
||||
"model": "claude-sonnet-4-6",
|
||||
"output_config": {
|
||||
"effort": "medium"
|
||||
},
|
||||
"system": "Output-guard: deploy output looked clean.",
|
||||
"temperature": 1.0,
|
||||
"thinking": {
|
||||
|
||||
@@ -23,6 +23,9 @@
|
||||
}
|
||||
],
|
||||
"model": "claude-sonnet-4-6",
|
||||
"output_config": {
|
||||
"effort": "medium"
|
||||
},
|
||||
"temperature": 1.0,
|
||||
"thinking": {
|
||||
"type": "adaptive"
|
||||
|
||||
@@ -42,6 +42,9 @@
|
||||
}
|
||||
],
|
||||
"model": "claude-sonnet-4-6",
|
||||
"output_config": {
|
||||
"effort": "medium"
|
||||
},
|
||||
"temperature": 1.0,
|
||||
"thinking": {
|
||||
"type": "adaptive"
|
||||
|
||||
@@ -34,6 +34,9 @@
|
||||
}
|
||||
],
|
||||
"model": "claude-sonnet-4-6",
|
||||
"output_config": {
|
||||
"effort": "medium"
|
||||
},
|
||||
"temperature": 1.0,
|
||||
"thinking": {
|
||||
"type": "adaptive"
|
||||
|
||||
@@ -51,6 +51,9 @@
|
||||
}
|
||||
],
|
||||
"model": "claude-opus-4-8",
|
||||
"output_config": {
|
||||
"effort": "medium"
|
||||
},
|
||||
"thinking": {
|
||||
"display": "summarized",
|
||||
"type": "adaptive"
|
||||
|
||||
@@ -23,6 +23,9 @@
|
||||
}
|
||||
],
|
||||
"model": "claude-opus-4-8",
|
||||
"output_config": {
|
||||
"effort": "medium"
|
||||
},
|
||||
"thinking": {
|
||||
"display": "summarized",
|
||||
"type": "adaptive"
|
||||
|
||||
@@ -43,6 +43,9 @@
|
||||
}
|
||||
],
|
||||
"model": "claude-opus-4-8",
|
||||
"output_config": {
|
||||
"effort": "medium"
|
||||
},
|
||||
"thinking": {
|
||||
"display": "summarized",
|
||||
"type": "adaptive"
|
||||
|
||||
@@ -42,6 +42,9 @@
|
||||
}
|
||||
],
|
||||
"model": "claude-opus-4-8",
|
||||
"output_config": {
|
||||
"effort": "medium"
|
||||
},
|
||||
"thinking": {
|
||||
"display": "summarized",
|
||||
"type": "adaptive"
|
||||
|
||||
@@ -39,6 +39,9 @@
|
||||
}
|
||||
],
|
||||
"model": "claude-opus-4-8",
|
||||
"output_config": {
|
||||
"effort": "medium"
|
||||
},
|
||||
"thinking": {
|
||||
"display": "summarized",
|
||||
"type": "adaptive"
|
||||
|
||||
@@ -23,6 +23,9 @@
|
||||
}
|
||||
],
|
||||
"model": "claude-opus-4-8",
|
||||
"output_config": {
|
||||
"effort": "medium"
|
||||
},
|
||||
"thinking": {
|
||||
"display": "summarized",
|
||||
"type": "adaptive"
|
||||
|
||||
@@ -42,6 +42,9 @@
|
||||
}
|
||||
],
|
||||
"model": "claude-opus-4-8",
|
||||
"output_config": {
|
||||
"effort": "medium"
|
||||
},
|
||||
"thinking": {
|
||||
"display": "summarized",
|
||||
"type": "adaptive"
|
||||
|
||||
@@ -34,6 +34,9 @@
|
||||
}
|
||||
],
|
||||
"model": "claude-opus-4-8",
|
||||
"output_config": {
|
||||
"effort": "medium"
|
||||
},
|
||||
"thinking": {
|
||||
"display": "summarized",
|
||||
"type": "adaptive"
|
||||
|
||||
@@ -47,6 +47,7 @@
|
||||
"stream_options": {
|
||||
"include_usage": true
|
||||
},
|
||||
"temperature": 0.5,
|
||||
"tools": [
|
||||
{
|
||||
"function": {
|
||||
|
||||
@@ -21,5 +21,6 @@
|
||||
"stream": true,
|
||||
"stream_options": {
|
||||
"include_usage": true
|
||||
}
|
||||
},
|
||||
"temperature": 0.5
|
||||
}
|
||||
|
||||
@@ -30,6 +30,7 @@
|
||||
"stream_options": {
|
||||
"include_usage": true
|
||||
},
|
||||
"temperature": 0.5,
|
||||
"tools": [
|
||||
{
|
||||
"function": {
|
||||
|
||||
@@ -30,6 +30,7 @@
|
||||
"stream_options": {
|
||||
"include_usage": true
|
||||
},
|
||||
"temperature": 0.5,
|
||||
"tools": [
|
||||
{
|
||||
"function": {
|
||||
|
||||
@@ -38,6 +38,7 @@
|
||||
"stream_options": {
|
||||
"include_usage": true
|
||||
},
|
||||
"temperature": 0.5,
|
||||
"tools": [
|
||||
{
|
||||
"function": {
|
||||
|
||||
@@ -18,5 +18,6 @@
|
||||
"stream": true,
|
||||
"stream_options": {
|
||||
"include_usage": true
|
||||
}
|
||||
},
|
||||
"temperature": 0.5
|
||||
}
|
||||
|
||||
@@ -34,6 +34,7 @@
|
||||
"stream_options": {
|
||||
"include_usage": true
|
||||
},
|
||||
"temperature": 0.5,
|
||||
"tools": [
|
||||
{
|
||||
"function": {
|
||||
|
||||
@@ -30,6 +30,7 @@
|
||||
"stream_options": {
|
||||
"include_usage": true
|
||||
},
|
||||
"temperature": 0.5,
|
||||
"tools": [
|
||||
{
|
||||
"function": {
|
||||
|
||||
@@ -47,6 +47,7 @@
|
||||
"stream_options": {
|
||||
"include_usage": true
|
||||
},
|
||||
"temperature": 0.5,
|
||||
"tools": [
|
||||
{
|
||||
"function": {
|
||||
|
||||
@@ -21,5 +21,6 @@
|
||||
"stream": true,
|
||||
"stream_options": {
|
||||
"include_usage": true
|
||||
}
|
||||
},
|
||||
"temperature": 0.5
|
||||
}
|
||||
|
||||
@@ -30,6 +30,7 @@
|
||||
"stream_options": {
|
||||
"include_usage": true
|
||||
},
|
||||
"temperature": 0.5,
|
||||
"tools": [
|
||||
{
|
||||
"function": {
|
||||
|
||||
@@ -30,6 +30,7 @@
|
||||
"stream_options": {
|
||||
"include_usage": true
|
||||
},
|
||||
"temperature": 0.5,
|
||||
"tools": [
|
||||
{
|
||||
"function": {
|
||||
|
||||
@@ -38,6 +38,7 @@
|
||||
"stream_options": {
|
||||
"include_usage": true
|
||||
},
|
||||
"temperature": 0.5,
|
||||
"tools": [
|
||||
{
|
||||
"function": {
|
||||
|
||||
@@ -18,5 +18,6 @@
|
||||
"stream": true,
|
||||
"stream_options": {
|
||||
"include_usage": true
|
||||
}
|
||||
},
|
||||
"temperature": 0.5
|
||||
}
|
||||
|
||||
@@ -34,6 +34,7 @@
|
||||
"stream_options": {
|
||||
"include_usage": true
|
||||
},
|
||||
"temperature": 0.5,
|
||||
"tools": [
|
||||
{
|
||||
"function": {
|
||||
|
||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user