mirror of
https://github.com/turnstonelabs/turnstone.git
synced 2026-08-12 23:12:23 -06:00
110d44b07e8f0bfdb1fc3e980b60a76e5f418cbd
11 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Patrick Buckley
|
110d44b07e |
refactor(tools): remove man, math, and plan_agent built-in tools
`man` and `math` duplicated capabilities already reachable through `bash`; `plan_agent` is better expressed as a `task_agent` running a planning skill, and carried a large amount of special-case machinery (plan-review gate, refinement loop, per-kind model routing). Removing all three shrinks the tool surface and cuts per-call token cost. Also removed, as dead-once-the-tools-are-gone: - the `math` sandbox executor (`turnstone.core.sandbox`) and its `[sandbox]` extra; the eval analyst now runs bash-only - the read-only `AGENT_TOOLS` sub-agent tool set and the `agent` tool-metadata key (`task_agent`/`TASK_AGENT_TOOLS` retained) - the plan-review protocol end to end: the `on_plan_review` UI hook, `resolve_plan`, `POST /v1/api/plan` + `POST /v1/api/route/plan`, the `plan_review`/`plan_resolved` SSE events, and their Python SDK / TypeScript SDK / OpenAPI / frontend / Discord+Slack bindings - the `model.plan_alias` / `model.plan_effort` settings and the registry `plan_model` / `plan_effort` routing fields TOOLS 31->28, TASK_AGENT_TOOLS 13->11; COORDINATOR_TOOLS unchanged. BREAKING CHANGE: removes the `man`, `math`, `plan_agent` tools, the plan-review SSE/HTTP/SDK surface, and the plan_* model-routing settings from the experimental 1.6 line. |
||
|
Patrick Buckley
|
3233719856 |
feat(judge): output_guard LLM stage with capability gate (#560 mitigation #1)
Adds a second, LLM-driven stage to the output guard so domain-camouflaged prompt-injection payloads that the regex stage misses (arXiv:2605.22001 — Llama 3.1 8B evades the existing regex set on ~90% of camouflaged prompts) get caught before the tool output lands in the assistant's context. ## Surface * New `OutputGuardJudge` in `turnstone/core/output_guard_judge.py` — synchronous, single-shot LLM call. Inlines the alias-resolution + client-config + JSON-parsing helpers (copied verbatim from `IntentJudge` at `judge.py:917-969` / `1604-1659`) rather than going through a shared module — when `IntentJudge` lifts its own helpers, both copies move together. * JSON-in-content verdict with a 3-strategy parser (direct / markdown fence / balanced braces). `IntentJudge` ships a 4th regex-field fallback; OutputGuardJudge deliberately doesn't, because strategy-4 hits on broken LLM output can extract a "verdict" from the model's reasoning quote that lands in storage looking identical to a clean strategy-1 result. Failure of all three returns `error="unparseable_verdict"` and the heuristic stage stands. * `OutputJudgeVerdict` is a frozen dataclass with: `risk_level` (none/low/medium/high — normalises `critical`→`high` and `info[rmational]`→`low` for IntentJudge-echo safety), `flags: tuple[str, ...]`, `reasoning`, `confidence: float` (0.0-1.0, parsed + clamped from the LLM's self-report; pass-through to audit, no threshold gating), `judge_model`, `latency_ms`, `error`. * Real wall-clock timeout via `ThreadPoolExecutor.shutdown(wait=False, cancel_futures=True)` on the timeout/cancel path — `with ... as ex:` would block return until the worker drained. 1s `cancel_event` poll mirrors `IntentJudge._run_judge` at `judge.py:1117-1118`. * HTTP client lazy-init + reuse for the judge instance's lifetime. Session-side model swap drops the entire judge, dropping the client with it. * Untrusted tool output wrapped in per-call random-nonced `<tool_output_NONCE>...</tool_output_NONCE>` fence. Closing-tag substrings in the raw text are case-insensitively backslash-escaped first (`</tool_output` → `<\/tool_output`) so an attacker can't break out even if they guess the nonce. System prompt classifies the fenced region as UNTRUSTED DATA so directives inside are evaluated as content, not obeyed. * Judge user prompt carries the heuristic verdict (risk + flags + annotations), the tool description (looked up from the session's tools registry), and the tool args (truncated to 500 chars, also classified UNTRUSTED in the system prompt since they may be caller-supplied). Lets the judge defer to the regex on credential leaks and focus on injection signals the regex set misses; also enables output-vs-request plausibility reasoning. ## Session integration * `_evaluate_output(call_id, output, func_name, *, tool_args="")` — heuristic always runs; LLM stage runs when `judge.output_guard_llm` is enabled. When the LLM produces a usable verdict and the heuristic didn't detect credentials, the LLM verdict is acted on; otherwise the heuristic stands. * Credential redaction is a regex-only signal. When `heuristic. sanitized` is non-None, the heuristic owns the acted assessment regardless of what the LLM said — an LLM asked about prompt- injection can correctly label a credential-bearing output as "none" risk for injection, but the secret still needs redaction. * `_batch_evaluate_outputs` runs the per-tool guard concurrently (4-worker pool) when LLM is enabled and there are ≥2 string outputs — collapses N×LLM-latency to ⌈N/4⌉×latency on the common 5-20 tool-calls-per-turn turn. * Per-session `TokenBucket(rate=1.0, burst=60)` caps adversarial LLM-fan-out cost at 60 calls/min/session. * Pre-truncation: the per-tool loop truncates output before the judge sees it, so the judge evaluates exactly what enters the assistant's context (no wasted tokens on text that won't land). * Both heuristic and LLM tier rows persisted to `output_assessments` when the LLM ran (audit completeness); heuristic-only rows skip when matched-clean to keep the table focused. ## Storage Migration 057 extends `output_assessments` with five LLM-tier columns: `tier` (`heuristic` / `llm`, backfilled to `heuristic`), `reasoning`, `judge_model`, `latency_ms`, `confidence`. Tie-break on `(created DESC, tier='llm' first)` so downstream consumers see the acted verdict first when the two rows tie at second resolution. `StorageBackend.record_output_assessment` + sqlite/pg implementations + `SessionUIBase.record_output_assessment` + `SessionUI` protocol + the test stub overrides (cli, eval, 9 test files) all take the new LLM-tier kwargs. ## Config surface Three new judge.* settings in `settings_registry`: * `judge.output_guard_llm` (bool, default False) — capability gate. Default off; operators opt in once a small/fast model is pointed at `output_guard_model`. * `judge.output_guard_model` (str, default "") — alias for the LLM stage. Empty inherits the session model (same fallback shape as `judge.model`). * `judge.output_guard_llm_timeout` (float, default 30.0, min 1.0) — wall-clock budget per call. Both `server.py` and `console/session_factory.py` wire these into the `JudgeConfig` they hand to `ChatSession`. ## Notes * No backwards-compatibility shims — the LLM stage is purely additive. * No reasoning/threshold gating on confidence; it rides as an audit-only signal per maintainer direction. Surface it in the `on_output_warning` dict so live UI / cluster broadcast can sort flagged outputs by judge certainty. * Tests: 392 lines of judge-only coverage (`test_output_guard_judge. py`) + 629 lines of session-integration coverage in `test_session. py`, plus the storage and stub-shape updates. |
||
|
Patrick Buckley
|
29b850919f |
feat(sse): refresh-resume for mid-stream page reloads
Refreshing a coordinator or interactive workstream pane while the LLM is mid-stream now restores the partial assistant text + reasoning immediately and flips the composer back to stop-mode, instead of showing nothing until the response completes. Per-turn inflight buffers (`_ws_inflight_content`, `_ws_inflight_reasoning`, `_ws_inflight_seq`) on `SessionUIBase` are kept separate from the existing multi-turn `_ws_turn_content` buffer that drives the dashboard's IDLE-piggyback payload. New `on_turn_start` (top of send-loop, defensive) and `on_turn_committed` (right after `messages.append(assistant_msg)`, primary) lifecycle hooks reset inflight at turn boundaries. The seq counter is monotonic across turns so a long-lived subscriber's `snap_seq` cutoff stays valid for the lifetime of the connection — resetting per-turn would silently drop turn N+1's first M tokens (M = whatever was streamed pre-snapshot in turn N). `snapshot_and_consume_state_payload` also drains inflight at idle/error so cancel and exception paths don't leak stale text. New `register_listener_with_in_progress_snapshot` atomically registers a listener and snapshots the inflight buffers; `make_events_handler` emits a `state_change` event (so the JS busy machine flips to stop-mode) followed by a one-shot `in_progress_snapshot` after the kind-specific replay, then strips the internal `_seq` field from yielded live events while filtering against `snap_seq`. A per-listener shallow `dict` copy in the live drain prevents the multi-tab race where one listener's `del event["_seq"]` would corrupt another listener's filter view. `_synthesize_cancelled_results` now emits synthetic `on_tool_result` events for each cancelled tool so live coord tabs can drop the newly-additive `coord-tool-batch--running` indicator cleanly. The indicator now coexists with `--auto`/`--approved` (applied on `tool_info` and `approval_resolved` approved; removed when every row in the batch has a result), making live tool execution visually parallel to the replay-time orphan rendering. Frontend handlers in `app.js` (interactive) and `coordinator.js` (coord) absorb EventSource auto-reconnect re-replays via a length-based prefix check on the in-progress buffer. New `InProgressSnapshotEvent` + `StateChangeEvent` dataclasses in the Python and TypeScript SDKs with type guards. `_MAX_TURN_CONTENT_CHARS` lifted 256 KiB → 512 KiB (single constant for both buffers — headroom for current commercial models). Regression tests cover race-free composition under concurrent writers, seq-filter dedup invariants, the cross-turn seq monotonic invariant, idle/error inflight drain, synthesized `on_tool_result` on cancel (including UI-hook failure isolation), and the multi-listener shared-dict invariant. |
||
|
Patrick Buckley
|
7d6b31e18a |
fix(coord): close gaps an operator's harness shakedown surfaced (#444)
* fix(coord): close gaps an operator's harness shakedown surfaced
Operator-driven shakedown of the coordinator tool surface flagged
five issues; this commit addresses all of them plus the review
findings against the initial fix.
1. Cancelled-mid-stream partial assistant content now carries a
"[generation cancelled before completion]" marker. Without it,
``inspect_workstream`` / ``wait_for_workstream`` callers and the
next coord-LLM turn read the truncated text as a complete answer.
``_cancelled_partial_msg`` no longer ships ``_provider_content``
(Anthropic would otherwise read that lane verbatim and bypass the
marker; partial tool_use blocks could also leak through).
2. ``spawn_workstream`` / ``spawn_batch`` no longer surface the
routing-proxy ``status`` field (always HTTP 200 on the success
path). The tool description claimed it was "lifecycle state at
creation"; code that did ``if result["status"] == "idle"``
silently never matched. Lifecycle state lives on the workstream
row — ``inspect_workstream`` is the read. Tool JSON descriptions
plus docs/coordinator-skills.md and docs/bulk-endpoints.md
examples updated to match.
3. ``inspect_workstream`` not-found error string is bare ("workstream
not found"); the structured ``ws_id`` field carries the queried
id. Pre-fix the error STRING echoed the id back at the caller
who just sent it — redundant and out of step with the rest of the
surface. Cross-tenant + missing rows still return the same shape,
preserving the existence-leak guarantee.
4. ``tasks(...)`` is now rejected when called in a parallel tool
batch. The prior shape relied on a docstring warning ("a list
paralleled with writes can reflect pre-write state") that put
cognitive overhead on every model invocation; turning the silent
footgun into an explicit error means the model only thinks about
the rule the moment it actually breaks it. Warning dropped from
the tasks tool description. ``_PARALLEL_INCOMPATIBLE_TOOLS``
constant in session.py is the extension point for any future
tool with the same read-after-write hazard.
Plus the multi-stage code review's findings against the initial
fix (q-1 / q-2 docs drift, q-3 idiom, q-4 keys-assertion, q-5
duplicate guard) — all addressed in the same pass.
Tests: 4752 pass, +6 net since the pre-fix baseline. Ruff + mypy
clean. Three new tests pin the parallel-batch-rejection behaviour
on tasks (rejected when batched, runs alone, sibling tools
unaffected); existing cancel + spawn + inspect tests updated to
match the new shape.
* fix(coord): close two copilot review gaps on PR 444
Copilot review on PR 444 flagged two follow-ups:
1. Empty-content cancel divergence — when ``GenerationCancelled``
races BEFORE the first content token, the prior shape skipped
``save_message`` and only appended an empty-content msg in
memory. In-memory and storage diverged: a rehydrate would see
nothing in storage but the session would carry an empty
assistant turn. Both branches now persist; on the empty-content
shape the marker becomes the entire message
("[generation cancelled before completion]") so storage matches
the in-memory history.
2. Test stub cleanup — three new tests injected ``ui.approve_tools``
via ad-hoc ``lambda + type: ignore[attr-defined]``. Replaced
with a permissive ``approve_tools`` method on ``_StubUI`` so the
stub matches the SessionUI surface the dispatcher actually
reads. Tests that exercise approval pathways can still override
per-instance.
Tests: 4752 pass. Ruff + mypy clean.
|
||
|
Patrick Buckley
|
7a32c51a1c |
fix: synthesize cancelled tool results instead of stripping turns (#237)
* fix: synthesize cancelled tool results instead of stripping turns When a user cancels during tool execution, the model previously lost all context about what was attempted (assistant message + tool_calls stripped entirely). Now synthesizes tool_result messages with is_error=true and "Cancelled by user." content for any tool_calls that lack matching results. This keeps the conversation valid for both providers while preserving the full tool call structure so the model knows what was tried. Also applies to KeyboardInterrupt with "Interrupted by user." text. * fix: persist synthesized cancel results to DB, assert is_error in test Copilot review: synthesized tool messages were in-memory only, creating a mismatch with DB that could break rewind/retry. Now calls save_message() for each synthesized result. Also adds is_error=True assertion to the cancel test. |
||
|
Patrick Buckley
|
f74aa2264e |
refactor: add is_error to on_tool_result protocol, remove text heuris… (#207)
* refactor: add is_error to on_tool_result protocol, remove text heuristics Add is_error keyword arg to SessionUI.on_tool_result() so tools report errors structurally. Server and JS client no longer guess from output text prefixes — each tool sets the flag at the source. Bash tool: exit code >= 2 is error, exit code 1 is ambiguous (grep no-match). History reconstruction keeps text heuristic as fallback for pre-migration data. Update SDKs (Python + TypeScript), test mocks, docs, and diagrams. * fix: infinite recursion in _report_tool_result, signal exits, stale docs * fix: add _tool_error_flags to test_load_skill ChatSession stubs |
||
|
Patrick Buckley
|
4198b59a0f |
fix: eager cancel_ref registration, SDK type drift, force-cancel tests (#203)
Providers now register the SDK stream handle eagerly (before returning the iterator) instead of lazily inside a generator body. This closes the window where cancel() couldn't abort a blocked HTTP read because the stream handle wasn't populated yet. - OpenAI: yield from → return (HTTP call + cancel_ref happen eagerly) - Anthropic: split into eager __enter__ + _iter_with_cleanup generator with defensive __exit__ on __enter__ failure - Console SDK: delete_setting() return type fixed from StatusResponse to DeleteSettingResponse (matching actual endpoint response) - 2 new threaded force-cancel integration tests verifying orphaned threads don't mutate messages and new generations succeed - Updated _CancelRef docstring, test robustness (assert on wait) |
||
|
Patrick Buckley
|
4f6ef13ce9 |
fix: cancel button race condition with stream abort and force cancel (#202)
The cancel endpoint emitted a 'cancelled' SSE event before the worker thread terminated. The frontend transitioned to "send" mode prematurely, so the next send got rejected with "Already processing a request." Backend: - Providers expose SDK stream handle via cancel_ref parameter so cancel() can close the HTTP connection and unblock iteration - Generation counter prevents orphaned threads from mutating messages or clearing cancel state after force cancel - _check_cancelled() added between retry attempts in _try_stream - Server polls (async, non-blocking) for cancelled worker to exit - Force cancel (force:true) abandons stuck worker, keeps cancel event set so subprocesses are killed, guards against spurious SSE events Frontend: - 'cancelled' shows "Cancelling..." then escalates to "Force Stop" after 2s for a harder cancel that abandons the worker immediately - 10s safety timeout auto-recovers if stream_end never arrives - busy_error re-enables stop button instead of showing send - Timeout cleanup in disconnectSSE, stream_end, and force .then() - Layout shift prevention (min-width, white-space: nowrap) - aria-label updates for accessibility Tests: - 7 new tests: stream close, error suppression, cancel_ref population, transport error conversion, non-cancel exception propagation, retry cancellation check |
||
|
Patrick Buckley
|
5378b33641 |
feat: output guard — evaluate tool results before they enter context (#109)
* feat: output guard — evaluate tool results before they enter context Add turnstone/core/output_guard.py — a time-budgeted heuristic that evaluates tool execution results after execution but before they enter the conversation context window. Priority-ordered detection (5s budget, highest priority first): 1. Prompt injection: override phrases, role injection, instruction override markers, meta-injection patterns 2. Credential leakage: API keys (OpenAI/GitHub/AWS/Google), PEM private key blocks, connection strings, .env secret format 3. Encoded payloads: script data URIs, hex shellcode sequences 4. Adversarial URLs: cloud metadata endpoints, credential query params 5. System info disclosure: private IPs, sensitive file paths Annotates and optionally redacts (credentials → [REDACTED:<type>]). Does NOT gate — surfaces warnings via on_output_warning callback. Integration: - Wired into session.py tool result loop via _evaluate_output() - JudgeConfig gains output_guard + redact_secrets fields (both default true) - SessionUI protocol gains on_output_warning callback - 25 compiled regex patterns, pure function, no I/O 29 tests covering all detection categories, benign output false positive checks, credential redaction, and time budget behavior. * fix: address PR #109 review — protocol, config, and guard fixes Copilot review feedback: - Replace _CLEAN singleton with _clean() factory to prevent mutable shared state (OutputAssessment has list fields) - Remove redundant second _CREDENTIAL_PATTERNS loop in _check_credentials - Evaluate text parts of list outputs (images) not just string outputs - Wire output_guard + redact_secrets through ConfigStore settings registry and _build_judge_config() so operators can configure via admin Settings tab - Remove --no-output-guard CLI flag claim from docs (use Settings tab) Typecheck fix: - Add on_output_warning to all SessionUI implementations: NullUI (eval, 5 test files), WebUI (server — emits SSE event), TerminalUI (CLI — ANSI colored warning), RecordingUI, FakeUI |
||
|
Patrick Buckley
|
28cb3a5c51 |
fix: approval timeout UI state and content flush before tool calls (#42)
* fix: approval timeout UI state and content flush before tool calls Two bug fixes: 1. Approval timeout now shows denied state in UI — resolve_approval() emits an approval_resolved SSE event so the browser transitions from pending to denied (red border + badge). Also fixes the cancel- during-approval path. Frontend resolveInlineApproval() gains a skipPost parameter to avoid redundant POST when server-initiated. ApprovalResolvedEvent added to Python and TypeScript SDKs. 2. Content streaming flushes pending buffer before tool call deltas — _stream_response() held up to 13 trailing chars in the pending buffer (for <think> tag detection) when transitioning to tool calls. Now flushed eagerly when tool_call_deltas arrive, before clearing in_think so reasoning text is correctly categorized. * fix: address Copilot review feedback on PR #42 Patch _execute_tools in stream flush test to prevent real bash execution, simplify confusing nested comprehension, and update resolve_approval() docstring to reflect cancel/timeout call paths. |
||
|
Patrick Buckley
|
fd507c6a3c |
feat: generation cancellation — stop button, cancel API, cooperative … (#40)
* feat: generation cancellation — stop button, cancel API, cooperative cancel Add cooperative cancellation via threading.Event on ChatSession. The cancel signal is set from outside the worker thread (HTTP handler, MQ bridge, or Escape key) and checked at defined checkpoints: per streaming chunk, before tool execution, inside bash commands, and at each sub-agent turn. Core: GenerationCancelled(BaseException) exception, cancel()/_check_cancelled() methods, partial content preservation in _stream_response, clean rollback in send() with idle state emission (no re-raise). Server: POST /v1/api/cancel endpoint, CancelledEvent SSE emission, worker thread safety net. Frontend: Stop button (■ Stop) with send/stop swap via setBusy(), Escape key shortcut, cancelled event handler. Accessible: aria-label, focus-visible override, light theme contrast, non-color differentiation. MQ: CancelMessage inbound type, bridge _handle_cancel routed handler. SDK: cancel() on Python async+sync clients, CancelledEvent in Python+TypeScript event registries, isCancelledEvent type guard. OpenAPI: CancelRequest schema + endpoint spec. Docs: API reference, architecture, SDK docs updated. Diagrams: conversation turn, tool pipeline, MQ protocol, workstream states, SDK architecture. * fix: address PR #40 review feedback - setBusy() now resets stopBtn.disabled so stop button is re-enabled on next generation after a successful cancel - Gate cancel side effects (resolve_approval, resolve_plan, cancelled SSE event) on worker_thread.is_alive() to avoid spurious events when idle - Add /v1/api/cancel endpoint and CancelRequest schema to TypeScript openapi-server.json to keep it in sync with Python-generated spec |