Provider search gating (replace-only): native web search now stands in for
a client web_search def that survived the persona visibility filter — on
both OpenAI surfaces and both injection lanes (web_search_options, the
server_side_tools loop, and _convert_tools' capability lane). A scribe or
any envelope hiding web_search stays search-free on search-capable models;
coordinators and tool-less utility calls stop receiving search too.
Resume stamp discipline: resume() loads config and parses the target's
stamp BEFORE touching session identity/history, so a corrupt stamp raises
with the session intact instead of half-adopting and then 'repairing' the
target's stamp on the next config save. The MCP lever now follows the
stamp on mid-session adoption: an MCP-off stamp drops the live surface in
place (listeners deregistered, toolsets reset); adopting an MCP-on stamp
into a session whose persona gated the client off is refused loudly (the
surface cannot be rebuilt post-construction). The REPL /resume handler
reports these errors instead of crashing the CLI.
Fail-closed default lane: a FAILED default-persona lookup at create is a
503 (routes) / clear exit (CLI) instead of silently degrading to the
unstamped stock envelope; a clean 'no default configured' still creates
legacy. resolve_persona_for_kind reports storage-unavailable distinctly
from unknown-persona.
Soft-set governance: tool_search expansion under a persona visibility set
recomposes the system prompt so tool-gated policy segments land with the
tool they gate. MCP resource/prompt catalogs gate on read_resource /
use_prompt visibility. Spawn judge/audit projections carry persona (the
human approval header already did). Active-list rows carry persona like
their project_id twin.
RBAC catalogs: persona.{create,read,write} join _VALID_PERMISSIONS and
the roles-editor sections, making the documented grant-outward path real.
Storage hardening: default-persona invariants move to a shared _utils
helper (validate + demote) with a pg advisory xact lock serializing
promotions and a post-promote single-default assertion; create maps the
unique-name race to the same ValueError as the pre-check; reads validate
JSON shape loudly (naming the persona); serialize enforces size caps;
field validation runs before invariant checks so malformed input is a 400,
never a TypeError-500. org_id guards explicit null and caps at 64.
Also: base_override='' means 'no override' at the compose boundary;
persona tag flattened/capped before the spawn approval header; /creative
redirect resolves the writer persona before advertising it; memory-nudge
gating unified through _nudges_enabled.
Provider/row-shape tests updated to the new contracts (the old ones
pinned the injection hole and the pre-persona row shape).
- claude-fable-5 capability entry: 1M context / 128K output, adaptive
thinking (summarized display), effort low..max incl. xhigh, no
sampling params, web + tool search, vision, reasoning replay, native
mid-conversation system messages
- document the Fable 5 wire quirk at the capability table: an explicit
thinking={"type": "disabled"} is a 400 on this model; the adaptive
branch never emits "disabled", so adaptive-or-omitted is preserved
- widen the native mid-conversation-system comments from opus-4-8-only
to opus-4-8 + fable-5 (protocol, provider, tool_advisory, prompts,
session)
- raise the anthropic SDK floor 0.39 -> 0.108: 0.39 predates every
named kwarg the provider sends (output_config 0.77, top-level
cache_control 0.83, mid-conversation system blocks 0.105); 0.108
adds claude-fable-5
- tests: capability assertions for claude-fable-5 + dated-variant
prefix match
- ruff format on two test modules that had drifted (a stray blank line
and multi-line calls that now fit on one line) — restores a clean
`ruff format --check`.
- test_attachment_buffer: pull `buf.discard(...)` out of the `assert`
expressions into locals so the eviction still runs under `python -O`
(CodeQL: assert statement has a side effect).
The canonical-Turn migration left lowering's fold/drop/repair passes Turn-typed even
though they convert to dicts internally and feed dicts to the translators, so
_prepare_wire_messages round-tripped the whole history Turn->dict->Turn ~7-8x per send
(even on the no-op early-return paths). Make fold_system_turns / drop_empty_user_turns
/ repair_wire_messages dict-native (list[dict]->list[dict]); _prepare_wire_messages now
threads the dict projection _full_messages already produced straight through, with no
Turn round-trip. self.messages stays the canonical Turn trajectory. export.py is
simplified (it converted to dicts immediately after repair anyway). Equivalence-
preserving — test_wire_payload_golden stays byte-identical.
repair_wire_messages / fold_system_turns / drop_empty_user_turns take and
return list[Turn] — the neutral lowering layer (A representation + B validity)
now speaks the canonical type. Their intricate content-merge / orphan-detect
internals run over the dict projection (reading Turn content blocks would only
duplicate turn_to_dict's content logic), so each bridges
dicts_from_turns ↔ turns_from_dicts at its boundary; byte-identical.
ChatSession._prepare_wire_messages lifts the wire dicts into Turns, runs the
lowering passes, and lowers the result back to the dict projection the provider
translators (the C layer) consume — the dict bridge now lives in the wire layer,
not in _full_messages. Export runs the same repair, reordered before the
non-canonical reasoning-content attach (a key the Turn model does not carry).
The provider translators keep their dict input by design: they are the format
layer that emits provider bytes, the vLLM reasoning-attach is a non-canonical
wire concern that sits between lowering and the provider on dicts, and feeding
the converters the lowered projection is equivalent to — and simpler than —
threading Turn content through them. Wire harness byte-identical; full
non-live suite green (7130).
Synthesizing a cancellation result for an assistant tool_call with no
matching tool result was triplicated across the translators: Anthropic's
verbatim-replay (pc_tool_ids) and rebuild branches, and sanitize_messages
for the OpenAI-compatible lanes (Chat, Responses, Google). The Anthropic
pc_tool_ids branch was also the sole repairer of a native tool_use orphan.
Lift it to one neutral policy — lowering.repair_wire_messages — run once in
ChatSession._prepare_wire_messages before the translator. It reads tool_calls
only, which is sound because the native/tool_calls mirror is enforced at save
(normalize_native_for_save): a verbatim-replay orphan is caught via its
mirrored top-level call. The translators become pure format translation and
carry no orphan synthesis.
The neutral cancellation turn carries is_error=True; Anthropic renders it on
the tool_result block, the OpenAI-compatible tool message has no such field
so sanitize_messages drops it (the C-layer translation of the flag).
sanitize_messages keeps one orphan synth of its own: a back-filled empty-id
tool_call (local servers that omit ids) is id-less when the upstream repair
runs and so invisible to it, so that lane owns its cancellation — preserving
the pre-refactor behavior for local servers.
reconstruct's load-time strip and the runtime-cancel persist-synth are
unchanged. Proven byte-identical against the per-provider wire-payload golden
harness (including a new native_orphan fixture); the harness applies the same
send-side repair the session does.
Phase-2 follow-ups to the mid-conversation-system consolidation:
- user_interjection framing (known #2): a queued message that drains mid-turn is
re-framed via render_user_interjection ("The user sent … User message: …") so
the user's words keep USER authority, not operator authority — the regression
mattered most on the native path, where the turn enters as a real role=system
message. Empty/whitespace interjections (e.g. a bare "!!!") are dropped (bug-2).
- empty-content user turns dropped at the wire boundary after the fold
(known #3): the wake pipeline's synthetic empty send("") leaves an empty user
turn on the native path (the nudge stays inline); an empty user message is
invalid on every provider. The drop runs after the fold so the fold-path wake
turn, which the nudge fills, survives.
- leading-system guard (_anthropic): a turn that converts to nothing no longer
lets a system message become messages[0] (the API requires messages[0]=user).
Newly reachable now that the empty-turn drop can expose it on a fresh-session
native wake.
- refresh stale .msg.watch-result comments (the card was removed) to describe
the current operator-bubble rendering.
Replace the two operator-context hacks (the <tool_output>/<system-reminder> content envelope and the transient _reminders side-channel) with one persistent {role: system, _source} trajectory turn. Adds supports_mid_conversation_system (claude-opus-4-8): native models take the turn inline; all others fold it into the preceding turn as a nonce-delimited <system-reminder> block declared in the system prompt as the sole trusted marker. Producers (advisories, metacog nudges, user interjections, idle/watch) emit system turns; the envelope/_reminders machinery, escaping round-trip, replay parser, and reminder SSE events are removed. Eager 060 migration un-wraps legacy envelopes. Net -1662 lines.
Known follow-ups from review (unfixed here): (1) the 060 un-wrap heuristic can irreversibly mis-rewrite bare tool rows that resemble the envelope, so do not run the migration until it is tightened; (2) user_interjection turns lost the user-framing/priority preamble (a regression, and a native-path authority-framing concern); (3) native-path wake nudge can emit empty user content.
Drop the Tavily and DuckDuckGo (ddgs) web_search backends for a single
self-hosted SearxNG service bundled into the docker-compose stacks.
Core:
- New SearXNGClient + _format_searxng; rewrite resolve_web_search_client to
(backend, searxng_url, searxng_engines, ...). MCP backend + oauth_user guard
unchanged. _resolve_search_client follows storage -> toml -> env -> default
precedence (explicit "" disables, via ConfigStore.stored_keys()).
- Drop the Tavily-era topic=finance (no SearxNG category); topic is now
general/news.
Settings/config:
- Remove tools.tavily_api_key, get_tavily_key, $TAVILY_API_KEY, [api].tavily_key.
- Add tools.searxng_url (default http://searxng:8080) + tools.searxng_engines,
with get_searxng_url/get_searxng_engines.
Compose + bundled config:
- Internal-only searxng service (no published API port, :ro config, /healthz
healthcheck, persistent searxng-cache volume) in both stacks; bundle
turnstone/deploy/searxng/settings.yml (JSON output on, limiter off).
- Caddy serves the SearxNG web UI on :8444 (dev: localhost-only; prod: opt-in).
- bootstrap extractor + wheel packaging updated.
Deps: drop the ddg extra + ddgs mypy override (regenerates uv.lock, removing the
lxml/h2/brotli transitives).
Docs: tools/docker/architecture/openshell + diagrams + config example + CHANGELOG;
docs/docker.md carries the AGPL-3.0 §13 operator note.
BREAKING: tools.web_search_backend no longer accepts "tavily"/"ddg";
tools.tavily_api_key and the ddg extra are removed. Run the bundled SearxNG (ships
in the compose stacks) or set TURNSTONE_SEARXNG_URL to an external instance.
Closes#545
Register `claude-opus-4-8` in the Anthropic capability table. Opus 4.8
shares Opus 4.7's request/response surface exactly — adaptive-thinking
only (`budget_tokens` rejected), sampling params removed, the
low/medium/high/xhigh/max effort levels, `thinking.display` defaulting to
omitted, 1M context, and 128K output — so the entry is a verbatim copy of
the 4.7 row. `_lookup_capabilities` longest-prefix matching then resolves
date-suffixed ids (e.g. `claude-opus-4-8-20260601`) without colliding
with the 4.7 key.
No provider code paths change: the existing 4.7 handling already covers
all of 4.8's behavior. Models are selected via config.toml / the admin
ConfigStore UI, so there is no catalog or dropdown to update.
- _anthropic.py: new claude-opus-4-8 capability entry + effort comment
- tests/test_providers.py: opus 4.8 bare + dated capability tests
- turnstone.example.toml: bump the showcased model example to 4.8
* feat(providers): api_surface toggle + mistral medium reasoning fix
Mistral medium open-weights served by vLLM expects reasoning_effort via
the Responses API (`reasoning.effort`), not as a `chat_template_kwargs`
entry on Chat Completions. The session was unconditionally injecting
`{"reasoning_effort": ...}` into `chat_template_kwargs` for every
openai-compatible request, which corrupted the prompt rendering for any
backend whose chat template didn't consume that key (Mistral medium,
Mistral cloud, Groq, OpenRouter).
Changes:
- Add `api_surface` ("chat" | "responses") to `ModelConfig.server_compat`
and thread it through `create_provider` / `model_registry.get_provider`.
`openai-compatible` defaults to Chat Completions; operators can flip
individual aliases to Responses for endpoints that support it.
- New `vllm-mistral-medium` profile that pre-fills api_surface=responses
on Detect for known Mistral medium model ids.
- Drop the unconditional `reasoning_effort` injection into
`chat_template_kwargs`. Operators running gpt-oss-style local
templates that consume `reasoning_effort` from the chat template now
opt in via `server_compat.extra_body.chat_template_kwargs`.
- New "API Surface" select in the Models admin tab; allowlist-validated
server-side at create/update time; pre-filled by Detect via the
profile suggestion.
- Evict the cached provider singleton in `ModelRegistry.reload()` when
api_surface changes (previously only cfg.provider triggered eviction).
- Fix `_run_agent` fallback path to inherit the session's primary alias
for capability and server_compat resolution; previously the fallback
passed `alias=None`, which silently dropped per-model caps on the
agent path.
Tests: 5117 passed (-m "not live"); ruff + mypy clean.
* fix(providers): don't auto-suggest Responses for Mistral medium
vLLM's Responses API surface for Mistral medium open-weights doesn't
wire up the Mistral tool-call parser as of vLLM 0.x — tool calls leak
into the response as ``[TOOL_CALLS]<name>{...}`` text instead of
structured tool_calls. Chat Completions on the same engine handles
tools cleanly via ``--tool-call-parser mistral``, and reasoning can be
turned on via the vLLM CLI ``--reasoning-parser`` flag.
Drop the auto-suggest mapping so Detect falls back to the generic
``vllm`` profile. Keep the ``vllm-mistral-medium`` profile definition
in place so an operator who specifically wants per-request effort and
accepts the tool-calling limitation can still pick "Responses API"
manually in the admin UI.
* fix(providers): address Copilot review on PR #469
- providers/__init__.py: drop the redundant *_responses_provider /
*_chat_provider names; have create_provider use _openai_provider and
_openai_compat_provider directly so they're not flagged as unused
globals.
- console/server.py: tighten _validate_api_surface to a strict equality
match against the canonical {"chat", "responses"} set. The previous
strip().lower() membership check accepted ' Responses '/'CHAT' but
stored the raw string verbatim, which then failed to round-trip
through the admin <select>.
- console/static/admin.js: gate the entire server_compat block (server
type, api_surface, extra_body) on provider == "openai-compatible" at
save time so toggling provider away can't leave a stale hidden surface
selection in the persisted capabilities JSON.
- tests/test_session.py: splat the bad kwarg via **dict so CodeQL no
longer flags the call as a wrong-name keyword (the point of the test
is the runtime contract, not the static type).
- tests/test_admin_model_registry_refresh.py: add endpoint-level tests
for the api_surface validation on both create and update — covers the
bogus-value rejection, non-canonical-string rejection, and the happy
path persisting through to the refreshed registry.
User-channel metacognitive nudges (correction, denial, resume, start,
completion) used to be spliced into ``user_msg["content"]`` permanently,
which leaked the ``<system-reminder>`` envelope into every consumer of
``self.messages`` — UI replay (mitigated by a regex strip in /history),
compaction, title generation, and any future channel adapter that
echoes conversation context. The /history strip was a band-aid;
compaction and title-gen still saw the raw spliced text.
Switch to a side-channel: ``_attach_pending_user_reminders`` writes the
rendered reminder list to ``user_msg["_reminders"]`` (sibling key,
leading-underscore convention shared with ``_attachments_meta`` /
``_provider_content``). At the provider boundary, a new
``_apply_reminders_for_provider`` builds a transient shallow-copy with
the reminder spliced into ``content``; the original message dict
stays clean. ``sanitize_messages`` drops the sibling key on the wire.
Once-per-session-not-per-turn semantics for the wire: after stream
success the loop calls ``_mark_reminders_delivered``, which flips a
``_reminders_delivered`` flag on every user message that carried
reminders into that call. ``_apply_reminders_for_provider`` skips
already-delivered messages so the model sees each reminder exactly
once (the turn it advised). ``_build_history`` ignores the delivered
flag entirely, so reconnecting tabs render the same nudge bubble the
originating tab saw via the live ``user_reminder`` SSE event.
UI surface:
- ``SessionUIBase.on_user_reminder`` enqueues a
``{type: "user_reminder", reminders: [...]}`` SSE event with the
same shape ``_build_history`` surfaces.
- ``app.js`` renders a ``.msg.user-reminder`` bubble (yellow accent,
pill-styled) anchored above the user message it advises, both
live and on history replay.
- ``replayHistory`` renders ``addUserMessage`` before
``addUserReminder`` so the anchor lookup finds the just-rendered
turn (not a prior one).
- Multi-tab caveat documented inline: non-originating tabs receive
no ``user_message`` SSE event today, so a reminder may anchor to
a stale prior bubble until ``/history`` reload corrects it.
Pre-existing bug surfaced by the audit: cancel handlers
(``GenerationCancelled`` / ``KeyboardInterrupt`` / generic
``Exception``) in ``ChatSession.send`` cleared
``_pending_tool_advisories`` but not the user-channel buffer. Both
now drain through a shared ``_drain_pending_advisories`` helper.
Removed the ``/history`` regex strip — the side-channel approach
makes it redundant. Hoisted ``escape_wrapper_tags`` +
``render_system_reminder`` imports to module top (called 2-3× per
turn).
Tests:
- ``TestApplyRemindersForProvider`` — pass-through-by-reference,
string + list content splice, escape on user-typed wrapper tags,
multi-reminder ordering, source-untouched invariant, delivered
flag skip path, fallback for unexpected content shape.
- ``TestMarkRemindersDelivered`` — flag idempotency, no-reminders
no-flag, only marks user messages with reminders.
- ``TestUpdateTokenTableMsgsParam`` — calibration uses pre-built
msgs when provided, falls back when not.
- ``TestUserAdvisoryCancelClear`` — all three cancel branches drain
the user buffer.
- ``TestReminderSidechannelIsolation`` — compaction's
``_format_messages_for_summary`` and the title-gen extraction
loop cannot see reminders by construction.
- ``TestSessionUIBaseUserReminderHook`` — ``on_user_reminder``
enqueues the right SSE shape.
- ``TestBuildHistoryReminderPropagation`` — ``entry["reminders"]``
propagation, absent / empty / multi / coexist-with-attachments
cases, malformed input filtering, all-malformed elision.
- ``test_sanitize_messages_strips_underscore_sibling_keys`` covers
``_reminders`` and ``_reminders_delivered``.
* feat(providers): add gpt-5.5 and gpt-5.5-pro capability entries
OpenAI announced gpt-5.5 on 2026-04-23 (ChatGPT/Codex first, API
"very soon"). Mirror the gpt-5.4 / 5.4-pro capability shape: 1M
context, native tool search, vision, xhigh effort; pro is
always-reasoning with no temperature and medium/high/xhigh only.
No provider-logic changes needed — OpenAI announced no API-surface
changes vs 5.4. Cache retention already covers 5.5 via the existing
startswith("gpt-5") prefix rule.
* test(providers): cover gpt-5.4-pro and gpt-5.5-pro in cache retention test
Pro variants share the same gpt-5 prefix and should keep 24h
retention; explicit coverage guards against regressions if the
prefix rule narrows in the future.
- Add claude-opus-4-7 capability entry (1M ctx, 128K output, adaptive
thinking, supports_temperature=False, thinking_display=summarized)
- Suppress temperature param for Opus 4.7 (API returns 400)
- Add thinking display opt-in via new ModelCapabilities.thinking_display
field - Opus 4.7 omits thinking by default, always send summarized
- Add xhigh effort level to mapping and Opus 4.7 effort_levels
- Add xhigh/max options to skill template dropdowns in admin console
- Align reasoning effort label capitalization across all console dropdowns
- Update example config to reference claude-opus-4-7
- 10 new tests with regression guards for Opus 4.6 backward compat
Verified against live API: streaming and completion calls succeed.
* feat: pass resolved capabilities through to providers, add server compat layer
The LLMProvider protocol previously forced providers to re-derive
capabilities from static lookup tables, ignoring config overrides set
via the admin UI or config.toml (e.g. thinking_mode, token_param).
This adds an optional capabilities parameter to create_streaming and
create_completion so the session can pass its config-merged
ModelCapabilities through to providers.
On top of this, adds a server compatibility layer for local model
servers (vLLM, llama.cpp). Profiles suggest thinking mode and server
workarounds (skip_special_tokens for vLLM, reasoning_format for
llama.cpp) during model detection, with structured admin UI fields
for server type, thinking mode, and extra body params.
Verified against real vLLM (Gemma 4 31B) and llama.cpp (Gemma 4 E4B)
servers.
* fix: defensive copy in _finalize_extra_body, expose thinking_param in UI
Shallow-copy extra_params and its chat_template_kwargs in the provider
before _apply_thinking_mode mutates them, so callers that reuse the
same dict across models are safe.
Replace the hidden thinking_param input with a visible text field
that appears when thinking mode is enabled. Shows the default
"enable_thinking" and hints that Granite/DeepSeek use "thinking".
* fix: address Copilot review feedback on admin UI and server compat
- Preserve unrepresentable thinking_mode values (e.g. "adaptive") in
raw capabilities JSON instead of silently dropping on edit round-trip
- Validate capabilities and extra body JSON are plain objects, not
arrays or primitives
- Deep-merge chat_template_kwargs from extra_body instead of silently
dropping, so operators can extend/override template kwargs
* fix: hide server compat section for non-local providers
The Server Compatibility fields (server type, thinking mode, extra
body) only apply to openai-compatible (local model servers). Hide
the entire section when the provider is openai, anthropic, or google.
* fix: normalize capsObj to plain object on edit load
Defend against DB rows where capabilities is a JSON literal null,
an array, or a primitive — previous code would crash on the
capsObj.server_compat / capsObj.thinking_mode reads. Same defensive
check also applied to the server_compat nested value.
* refactor: extract _isPlainObject helper for JSON type checks
Consolidates the null/array/typeof check that was inlined at three
different call sites into a single helper. Keeps the intent obvious
at each use site and avoids the awkward multi-condition ternary.
* fix: universal tool_call/tool_result orphan detection for OpenAI-compat providers
The Anthropic provider had orphan detection for mismatched tool_call ↔
tool_result pairs, but OpenAI-compatible providers (Chat Completions,
Google, Responses API) had none. When an Anthropic model runs behind
an OpenAI-compat API (e.g. Azure) or cancellation creates orphans,
the API rejects the malformed request.
- Rewrite sanitize_messages() with orphan detection: synthesize error
tool results for unmatched tool_calls, drop tool results with no
matching tool_call, fill empty tool_call IDs with positional remap
- Call sanitize_messages() from Responses API _convert_messages()
* fix: address review feedback on orphan detection
- Track answered IDs per-turn (local_answered) instead of scanning
all of out, preventing false matches from reused IDs across turns
- Drop empty-ID tool results that have no remap entry instead of
passing them through with invalid empty tool_call_id
- Increment empty_result_idx for every empty result, not just remapped
- Remove dead result_ids peek-ahead code
- Add test for repeated tool_call IDs across turns
* fix: accurate token usage tracking for compaction across all providers
Anthropic's input_tokens excluded cached tokens, causing massive
under-reporting (e.g. 327 vs 9000 actual) when prompt caching was
active. This prevented auto-compaction from triggering.
- Normalize Anthropic prompt_tokens to total input (input_tokens +
cache_creation + cache_read), matching OpenAI semantics
- Reset _last_usage per API call so tool-chain iterations get fresh
usage instead of max()-merging with stale values
- Add mid-turn compaction check during tool chains to prevent context
overflow before end-of-turn
- Anchor _remaining_token_budget() on provider-reported prompt_tokens
with local estimates only for the delta since last API call
- Improve _msg_char_count() to include structural overhead (role,
tool_call_id, tool call IDs) and handle image tokens in calibration
- Emit status after every API call, not just end of turn
* fix: defensive null coercion and index clamping from review feedback
- Add `or 0` to all getattr calls for input_tokens/output_tokens in
Anthropic provider (streaming + non-streaming) to handle SDK nulls
- Use getattr for non-streaming input_tokens/output_tokens instead of
direct attribute access for consistency
- Clamp _calibrated_msg_count with min() in _remaining_token_budget()
to prevent stale state from over-slicing after compaction
Gemini's OpenAI-compat endpoint requires thought_signature to survive
the tool-call round-trip. Previously dropped because the Chat Completions
provider cherry-picks only standard fields (id, type, function).
Fix: GoogleProvider now captures raw tool-call dicts (including
thought_signature) via provider_blocks — the same fidelity lane the
Anthropic provider uses for signature round-tripping. On the next turn,
_prepare_messages reconstructs tool_calls from the stored raw data and
strips _provider_content so it never reaches the wire.
Changes:
- _openai_chat.py: add _prepare_messages and _extract_tool_calls hooks
- _google.py: override hooks + tap-pattern _iter_stream for streaming
- model_registry.py: auto-detect .googleapis.com → google provider
- session.py: read cancel_on_approval from ConfigStore
- console/server.py: add PUT/DELETE to proxy route methods
- server.py: fix fork naming (don't inherit source display name)
* fix: skip chat_template_kwargs for commercial OpenAI API
OpenAI rejects chat_template_kwargs as an unknown parameter — it's only
meaningful for local model servers (vLLM, llama.cpp, SGLang).
Split OpenAIProvider into separate singletons for "openai" vs
"openai-compatible" so _provider_extra_params can gate on provider_name
instead of inspecting base_url. Also deduplicates agent inline code into
the same method and fixes pre-existing test pollution where
get_capabilities was mutated on the singleton without cleanup.
* feat: add OpenAI Responses API provider for commercial models
Split the OpenAI provider into three concrete implementations behind the
LLMProvider protocol:
- _openai_chat.py: Chat Completions API for local model servers
(vLLM, llama.cpp, SGLang)
- _openai_responses.py: Responses API for commercial OpenAI
(GPT-5.x, O-series)
- _openai_common.py: shared capability table, temperature/reasoning
gating, cache retention, citations, usage extraction
The Responses API handles reasoning_effort as a {"effort": value} dict,
system messages as an instructions field, and tool format translation at
the provider boundary. ChatSession is unchanged — the provider abstracts
the API difference.
Also fixes diff_file direction when comparing against provided content.
* fix: Responses API input format and local model provider routing
- Assistant input messages use plain string content (not output_text)
- Tool call argument deltas match on item_id, not call_id
- Auto-detect openai-compatible provider for non-api.openai.com URLs
- Fix diff_file direction when comparing against provided content
* fix: resolve env vars before provider auto-detection in config.toml models
Config-file model entries using ${ENV_VAR} placeholders in base_url were
not resolving env vars before _resolve_openai_provider(), causing
commercial OpenAI configs to be misclassified as openai-compatible.
* fix: MCP tools not surfacing after Sync to Nodes, update Anthropic tool search
Three fixes:
1. session_factory closure captured mcp_client=None when no --mcp-config
was passed at startup. internal_mcp_reload created a new MCPClientManager
on app.state but the factory never saw it. New workstreams got 0 MCP tools.
Fix: mutable _mcp_ref list shared between factory and reload handler.
2. Anthropic dropped the date suffix from tool_search_tool_bm25_20251119
and now requires name == type. Updated constant and tool definition.
3. Add diagnostic logging around API errors (provider, model, base_url,
message counts, full exception chain) and workstream resume (pre/post
provider state, alias resolution warnings).
Also adds Node.js 24 LTS to Dockerfile via multi-stage copy for npx-based
MCP servers.
* fix: address Copilot review — set_storage on reload, sanitize log output
- Call mcp_mgr.set_storage(storage) when internal_mcp_reload creates a
new MCPClientManager so prompt sync works for post-startup servers
- Strip query params from base_url before logging (may contain API keys
in some vLLM deployments)
- Split API error logging: concise warning (type names only) + separate
debug with exc_info=True for full traceback when needed
* chore: remove DDG MCP sidecar, web_search uses built-in ddgs client
The DuckDuckGo MCP server container is redundant — the built-in
DuckDuckGoClient (via ddgs package, included in all extras) auto-detects
when no Tavily key is configured. Removes the ddg-search service,
ddgCluster profile, and mcp-ddg.json config file.
Claude 4.6 (Opus + Sonnet) unified on 1M token context windows.
Update capabilities table from 200K to 1M for both models. Remove
claude-opus-4 and claude-sonnet-4 entries (end of life). 4.5 models
remain at 200K. Default fallback stays at 200K for unknown models.
Addresses Copilot review feedback on #219:
1. Anthropic _convert_messages: collect tool_use IDs in order (list
not set), filter empty IDs, defer synthetic results until after
real tool results so _merge_consecutive produces correct ordering.
2. Universal repair in reconstruct_messages: synthesize tool results
for mid-conversation orphaned tool calls on DB load. Benefits all
providers (OpenAI is lenient today but may tighten).
3. Test improvements: assert on is_error flag instead of "cancelled"
substring, verify real-before-synthetic ordering in partial results.
When a cancel interrupts tool execution, the assistant message with
tool_use blocks is saved to DB before tools run, but
GenerationCancelled prevents tool results from being created. The
in-memory rollback removes the orphaned message, but the DB row
persists. On resume, Anthropic rejects the conversation with
"tool_use ids were found without tool_result blocks".
Fix: _convert_messages now peeks ahead after each assistant message
with tool_use blocks. If any tool_use IDs lack matching tool_result
messages, synthetic error results are injected (is_error: true,
"Tool execution was cancelled."). Transparent to all callers,
provider-specific (OpenAI is lenient about this).
5 new tests covering: single orphan, multiple orphans, partial
results, complete results (no synthesis), and trailing orphan.
plan_agent and task_agent fail on Anthropic models with "Streaming is
required for operations that may take longer than 10 minutes" from
the SDK. The non-streaming create_completion path used
client.messages.create() which the SDK rejects for thinking-enabled
models.
Fix: use client.messages.stream() internally and call
get_final_message() to get the same Message object. Transparent to
all callers — fixes sub-agents, title generation, summarization,
web fetch, and judge create_completion calls.
The streaming path captured thinking_delta events but not
signature_delta, leaving the signature empty on round-trip and
causing 400 errors on multi-turn conversations with thinking enabled.
Replayed or cancelled conversations could produce assistant messages
with content=None and no tool_calls, which OpenAI-compatible APIs
reject with a 400. Fix at three layers for defense in depth:
- session.py: use empty string instead of None when building assistant
messages (streaming + cancellation paths)
- _utils.py: normalise content on DB load in reconstruct_messages()
- _openai.py: add _sanitize_messages() catch-all at provider boundary
Closes#194
* feat: add vision/image support to read_file tool
read_file now detects image files (PNG, JPEG, GIF, WebP, BMP, TIFF, ICO)
and returns base64-encoded content parts for vision-capable models.
Non-vision models receive a text description instead. A new
supports_vision flag on ModelCapabilities gates the feature, with
config.toml [models.*.capabilities] overrides for local models
(vLLM, llama.cpp, NIM).
* fix: address PR review feedback
- Discard _read_files on no-vision OSError path, include exception detail
- Discard _read_files on oversized image error (not a successful read)
- Validate capabilities type from config.toml (reject non-dict)
- Clarify tool description re: vision behavior and offset/limit scope
- Remove unused os import in tests, fix import sort order
- Handle list content (image tool results) in eval.py tool result loop
* Add dynamic tool search with native defer_loading for Anthropic/OpenAI
When MCP tools push the total tool count past a configurable threshold
(default 20), tool definitions are deferred to reduce token overhead and
improve tool selection accuracy. Three-tier approach mirrors the existing
web search pattern:
- Anthropic (Claude 4.x): native defer_loading + server-side BM25 search
- OpenAI (GPT-5.4+): native defer_loading + hosted search
- vLLM/llama/NIM: client-side BM25 fallback via synthetic tool_search tool
New module turnstone/core/tool_search.py with BM25Index (pure-Python,
zero deps) and ToolSearchManager (session-scoped visibility, expansion,
server hint generation). Discovered tools persist for the session lifetime
so the model only searches once per capability needed.
Config: [tools] search/search_threshold/search_max_results
CLI: --tool-search {auto,on,off}, --tool-search-threshold, --tool-search-max-results
Agents (plan/task) exempt — their scoped tool sets are always small.
43 new tests (1253 total). All diagrams regenerated with PlantUML 1.2025.2.
* Fix Copilot review feedback on tool search
- Fix _MCP_PREFIX_RE to handle underscores in server names (non-greedy match)
- Use ordered dict for _expanded to preserve tool discovery order
- Avoid constructing ToolSearchManager when below threshold in auto mode
- Return empty string from _mcp_server_summary when no servers (not "none")
- Fix CLI help text to reference threshold generically, not hardcoded "20"
- Fix agent exemption docs to accurately describe scoped tool sets
- Fix README to not hardcode "30+" threshold number
* Add GPT-5.3, GPT-5.4, and pro model capabilities
Add capability entries for gpt-5-pro (272k output, high-only reasoning),
gpt-5.2-pro, gpt-5.3, gpt-5.4 (1.05M context), and gpt-5.4-pro.
* Validate reasoning_effort against model capabilities
_apply_model_params now falls back to caps.default_reasoning_effort when
the requested value is not in caps.reasoning_effort_values. Prevents
sending unsupported effort levels to models like gpt-5-pro (high only).
* Fix circuit breaker, rate limiter, and Anthropic web search correctness (#15)
Three tech debt items addressing correctness and security gaps:
Circuit breaker HALF_OPEN single-request permit:
- Rename should_allow_request property to acquire_request_permit() method
to make the side-effecting, non-idempotent nature explicit
- Add _half_open_permit flag: exactly one probe request in HALF_OPEN,
subsequent callers blocked until probe completes
- Explicitly reset permit on all state transitions (record_success,
record_failure) for clean state machine invariants
- Session uses BaseException catch to ensure record_failure always fires,
preventing permanent circuit deadlock on probe crash
Rate limiter X-Forwarded-For support:
- Add resolve_client_ip() with rightmost-untrusted XFF parsing
- Configurable trusted_proxies via --ratelimit-trusted-proxies CLI flag
and [ratelimit] trusted_proxies config (comma-separated CIDRs)
- IPv4-mapped IPv6 normalization (::ffff:x.x.x.x → IPv4) for dual-stack
- Clientless requests (request.client is None) pass through instead of
sharing a single "unknown" bucket
- Log warning for invalid CIDR entries in trusted_proxies config
- Show trusted proxies in startup log when enabled
Anthropic web search multi-turn encrypted content:
- Capture raw provider content blocks during streaming via _block_to_dict()
using model_dump(exclude_none=True) to avoid Anthropic API rejection
- Accumulate thinking_delta into raw_blocks (was silently empty on replay)
- Store _provider_content on assistant messages, pass through verbatim in
_convert_messages() so encrypted_content/encrypted_index survive turns
- Persist to SQLite via new provider_data column (auto-migrated)
- Add thinking/signature to _block_to_dict fallback attribute list
23 new tests (735 total), ruff + mypy clean.
* Fix Copilot PR #15 review issues: provider data, circuit breaker, IP normalization
- Persist assistant message when provider_data exists even if text
content is empty — prevents losing Anthropic web search encrypted
content needed for multi-turn replay (session.py)
- Re-raise KeyboardInterrupt/SystemExit immediately after recording
failure instead of attempting fallback models (session.py)
- Consume HALF_OPEN permit for the transition caller — prevents two
concurrent probe requests when only one should be allowed
(healthcheck.py)
- Normalize IPv4-mapped IPv6 addresses consistently in
resolve_client_ip() — prevents duplicate rate-limit buckets for
::ffff:x.x.x.x vs x.x.x.x (ratelimit.py)
* Add provider-native web search with Tavily fallback
Replace client-side Tavily web search with provider-native implementations:
- Anthropic: inject web_search_20250305 server-side tool, handle
server_tool_use / web_search_tool_result streaming blocks, emit
info_delta for search status display
- OpenAI: inject web_search_options for gpt-5-search-api, format
url_citation annotations as footnote sources
- Local/vLLM: preserve existing Tavily-based web_search tool as fallback
Add supports_web_search to ModelCapabilities and info_delta to StreamChunk.
Remove end-of-life GPT-4o model entries from capability tables.
Update docs, diagrams, and README. 88 provider tests (32 new).
* Fix Copilot PR #13 review: capture streaming url_citation annotations
Accumulate url_citation annotations during OpenAI streaming and emit
formatted citations as a final info_delta chunk after the stream ends.
Previously annotations were only captured in non-streaming mode, so
search model users in the interactive path never saw citation sources.
* Add multi-provider LLM adapter with model capability flags
Introduce a provider abstraction layer between ChatSession and LLM SDK
clients, enabling native support for Anthropic alongside OpenAI-compatible
APIs. Each provider translates at the API boundary while the internal
message format remains OpenAI-like throughout session history and persistence.
- LLMProvider protocol with StreamChunk/CompletionResult normalized types
- OpenAIProvider: GPT-4o, GPT-5.x, O-series capability tables with
conditional temperature, reasoning_effort, and token param handling
- AnthropicProvider: native streaming, message/tool format conversion,
adaptive vs manual thinking modes, effort parameter for 4.6 models
- ModelCapabilities per-model flags: temperature support, token param name,
thinking mode, effort levels, context window, max output
- Smart auto-detect: latest Opus for Anthropic, latest base GPT for OpenAI
- --provider CLI flag for both turnstone and turnstone-server
- anthropic SDK as optional dependency (pip install turnstone[anthropic])
- 56 new provider tests, 672 total passing
- Updated architecture docs and 4 PlantUML diagrams
* Fix Copilot PR #12 review: reasoning_effort gating, Anthropic thinking, provider factory
- Default reasoning_effort_values to () so unknown/local models don't
receive unsupported top-level reasoning_effort param. Models that need
it (GPT-5.x, search models) have explicit capability declarations.
- Fix Anthropic _reasoning_params: "none" and "" effort now return {}
instead of enabling thinking with 4096 budget.
- Use create_provider("openai") singleton instead of OpenAIProvider()
in ChatSession fallback for consistency with registry path.
- Add 8 parameter gating tests: unknown model no reasoning_effort,
GPT-5 no temperature, GPT-5.1 conditional temperature, O-series
no temperature, Anthropic none/empty/low effort.
Patrick Buckley