Each pipeline phase owns one llama-server slot (steward 0, orchestrator
1, synthesizer 2), carried as id_slot in extra_body through the same
mechanism tool_choice already uses, so the phase's stable prompt prefix
stays in that slot's KV cache and a turn re-prefills only its new
tokens. Off by default; a no-op on the Claude backend and ignored by
Ollama, so the flag is safe on any backend and the cutover itself stays
a pure env swap.
The merge helper preserves existing extra_body keys — mutation-checked
(dropping the merge fails exactly the test written for it).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three native-API touchpoints converted — steward /api/generate to
/v1/chat/completions, embeddings /api/embeddings to /v1/embeddings,
health /api/tags to /v1/models — so the backend behind OLLAMA_HOST is
swappable by env alone. This makes the serving plan's "tatlock needs
zero changes" claim true for the llama-server cutover and for forge
after it. EMBEDDING_HOST (default: OLLAMA_HOST) lets gen and embed
point at different servers, which the boilerroom stack needs.
The dead OllamaClient goes with it: a native-API client nothing
imported, whose presence would make the post-cutover "no native
endpoints" grep lie.
Contract tests rewritten to mirror the new requests and extended with
the embeddings shape (dim must match the configured Qdrant dimension).
All pass against live Ollama's /v1 — deployable before any cutover.
Both new assertions mutation-checked via env overrides (bogus model,
wrong dim: each fails). The Anthropic contract now skips on 401: the
configured key is deliberately revoked per workspace D-11, which is
"fallback disabled", not a boundary break. Embedding continuity across
backends was measured separately: same nomic bytes, cosine 1.0000.
663 unit tests pass; ruff clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
`generate_response` was `async def` with a `pass` body and no `yield`.
An async function that never yields is a coroutine, so the declared type
was Coroutine[..., AsyncGenerator[OutputItem, None]] — something a
caller must await before it can be iterated.
Nobody awaits it. Both implementations contain yields (TatlockAgent 5,
LoremTesterAgent 3), which makes them async generators directly, and
both call sites do `async for item in agent.generate_response(...)`.
The abstract method's own docstring says "Yields:" and its own example
iterates the call without awaiting. Implementations, consumers and prose
all agreed; only the declaration dissented.
Removing one word fixes it, and it is the declaration that was wrong
rather than the four places reporting it.
WHY NOTHING CAUGHT THIS. The abstract body is `pass` and nothing calls
super().generate_response — verified across the tree — so the wrong
declaration has no runtime consequence and cannot fail a test. It was
invisible by construction, and it presented as four unrelated errors in
four files (two override, two attr-defined), none of which named the
cause. Anyone fixing them where they appeared would have annotated the
implementations to match the interface and made the real defect
permanent.
tests/agents/test_agent_interface.py covers it going forward. The
load-bearing case is not "the interface is X" or "the implementation is
Y" separately — both could drift together and still pass — but that the
two AGREE about what kind of callable this is.
Mutation-checked: restoring `async` fails 3 of the 5 new tests, the two
that still pass being the implementation checks, which are correctly
unaffected. Anchor asserted unique before the mutation was written, and
the fix asserted back into place afterwards.
95 errors -> 71 across this branch; this commit accounts for 4 of them.
Suite 662 passed, 1 failed — that failure is the known LLM-nondeterministic
calculator test, which passed on the previous run of this same branch and
failed on this one, which is the clearest available evidence that it is
unrelated to any of this work.
Co-Authored-By: Claude <noreply@anthropic.com>
Twelve gain `-> None`, each confirmed by AST to contain no returning
`return` and no `yield` rather than by reading the name and assuming.
The three context-manager exits gain the canonical
type[BaseException]/BaseException/TracebackType argument triple.
Both files taking TracebackType needed the import, and inserting it
before the first import broke ruff's I001 — lint was exit 0 at the
baseline commit, verified by stashing this work and re-running, so that
breakage was mine. Fixed with `ruff check --fix` on the two files, which
placed the import in sorted position.
86 errors -> 75; no-untyped-def 29 -> 14.
Suite: 658 passed. The baseline was 657 passed with one failure in
test_tatlock_tool_call_logging_calculator, which asserts on the content
of a live model's reply. It passing here is nondeterminism, NOT evidence
this commit fixed anything, and it may fail again on the next run.
Co-Authored-By: Claude <noreply@anthropic.com>
Each element type is taken from how the container is used rather than
guessed: kept_items is returned from trim_to_fit, whose signature is
already list[Any]; traces collects the dicts built at the append site;
expert_results and tool_outputs are keyed by tool_name (str) and hold
ToolReturnPart.content.
tracing_router.py needed `from typing import Any` added — it had no
import for it, so annotating without that would have traded a
var-annotated error for a name-defined one. Function-body variable
annotations are not evaluated at runtime, so this would not have raised;
mypy was the only thing that would have caught it.
90 errors -> 86.
Co-Authored-By: Claude <noreply@anthropic.com>
mypy's warn_unused_ignores is on, so these were reported as errors in
their own right — a suppression that no longer suppresses is a claim
that something is broken when it is not, and it silently widens to
cover a real error if one later appears on that line.
Two carried a "Forward reference" note that is still accurate; the note
is kept and only the ignore removed.
95 errors -> 90. Comment-only, so no runtime behaviour can have changed
and the suite was not re-run for this commit. The `# type: ignore` count
across the tree drops from 6 to 1, which is the anchor for the rest of
this work: clearing a type error by suppressing it would push that
number the other way.
Co-Authored-By: Claude <noreply@anthropic.com>
Partial work on the typecheck gate: 103 mypy errors down to 95, and the two
shared roots in agents/tatlock.py are gone. The rest is genuine per-function
annotation work and is not attempted here.
Five conversation lists were declared bare. mypy infers the element type from
the first append, which is a ModelRequest, and then rejects every ModelResponse
that follows — five errors from five lists that all hold the same thing: a
conversation, which is both kinds of message. Annotated as list[ModelMessage],
which is pydantic_ai's own union for exactly this.
The agent had no deps type. It is built as Agent(model, system_prompt=...),
inferred Agent[None, str], while every tool it registers takes
RunContext[ToolCallTracker] and run() is called with a tracker. The declaration
now says what was already happening: Agent[ToolCallTracker, str]. Note this is a
runtime-visible change — pydantic_ai is now told the deps type it was being
handed anyway — so it was verified against the suite rather than reasoned about:
658 passed.
_register_tools carries an assert rather than a None check. It is called from
_ensure_agent immediately after the agent is constructed, so a None there is a
broken invariant, not a case to handle; an `if is None: return` would silently
register no tools.
Two corrections to my own work in this commit. Declaring `_agent: Agent | None`
first made things worse, not better — resolving the bare Agent to Agent[None, str]
surfaced four new argument-type errors that the Any had been hiding, which is
how the missing deps type became visible at all. And an import fix I thought I
had made was a no-op: the target was a multi-line import, my replace matched
nothing, and I had asserted the precondition without asserting the result. Ruff
caught it. That is the same mistake as a changelog edit earlier today, so the
assert now checks what landed.
Co-Authored-By: Claude <noreply@anthropic.com>
The 21 the automatic pass could not make on its own. `ruff check` and
`ruff format --check` are both clean now; typecheck is still red and is next.
`in_reasoning` in chat/service.py was a complete state machine that nothing read:
initialised False, set True when a reasoning delta arrived, set False when the
summary ended — three assignments, zero reads. Ruff reported one at a time, and
removing each revealed the next, so what looked like a single stray variable took
three passes to bottom out. The branches themselves do real work and are
untouched; only the flag is gone.
Four `raise HTTPException` inside `except` blocks now chain with `from e`. Until
now a failure while handling an error was indistinguishable from the error, which
matters most in exactly the situation where the traceback is all you have.
In biographer/tools.py the binding was unused but the call is not: MemoryType()
is called for the ValueError it raises on an invalid name. The binding is gone
and the call and its comment stay, because dropping the line would have removed
the validation.
The rest are unused bindings in tests where the assertions are on something else
(call_args, mostly), plus three unused loop variables and an isinstance tuple.
One correction to my own work: removing a dead comprehension in
test_error_handling.py left an `if` block with nothing but comments in it, which
is a SyntaxError. Ruff caught it immediately. The block now says what the test
actually pins — that the stream parses without crashing, which reaching that line
demonstrates — rather than computing a list nobody asserts on.
`make test` is intermittent here, and it is not this change.
test_tatlock_tool_call_logging_calculator failed in two of five full runs across
both HEAD and this branch, and passes in the other three; it also fails in
isolation at HEAD while passing in isolation here. Order- or timing-dependent.
Recorded rather than chased, since tests are not gated in this repo yet.
Co-Authored-By: Claude <noreply@anthropic.com>
Mechanical only, and separated from the judgment calls that follow so the
reviewable changes are not buried in a 98-file whitespace diff.
227 automatic fixes: 60 blank lines carrying whitespace, 60 unsorted import
blocks, 34 Optional[X] to X | None, 28 unused imports, 16 deprecated typing
imports, 12 datetime.timezone.utc to datetime.UTC, and assorted smaller
modernisations. Then `ruff format` over src and tests: 98 files reformatted,
35 already conforming.
No file among the unused-import findings defines __all__ or is an __init__.py,
so nothing here removes a re-export.
`make test`: 658 passed, unchanged from HEAD.
Two things observed while verifying, neither addressed here:
`pytest tests/` cannot collect — tests/e2e/test_orchestration_e2e.py uses an
`e2e` marker that is not registered, and the config is strict about markers.
This fails identically at HEAD, so it predates this change; `make test` passes
because it ignores tests/e2e, tests/integration and tests/contracts.
test_tatlock_tool_call_logging_calculator is flaky. It failed once in a full run
with these changes and passed on the next, passes in isolation with them, and
fails in isolation at HEAD. It is order- or timing-dependent, not a regression
from this commit — established by running the full suite both ways rather than
by reasoning about which change could have caused it.
Co-Authored-By: Claude <noreply@anthropic.com>
The prompt tells the Steward to state its choice on a DELEGATE line and
to explain itself on REASON, COMPLEXITY and CONTEXT lines. Extraction
ignored that structure and substring-matched capability domains across
the entire response, so ordinary English in the explanation selected
agents: "description" contains the housekeeper domain "script",
"discover" contains "cover", "acknowledge" contains "knowledge" and
"know", "economy" contains the biographer domain "my".
Every one of those was a real delegation. A spurious librarian is a
multi-second web call on a query that asked for arithmetic.
It also made prose length a routing input, which would have quietly
corrupted the thinking benchmark this was found during: anything that
shortened the Steward's output reduces accidental substring hits and so
reads as improved routing.
Resolution is now layered, most explicit first — a DELEGATE line opening
with a capability name, then a capability named anywhere on that line,
then a domain on that line. With no DELEGATE line at all the response is
matched on capability names only, never domains, so the conversational
path still answers with no capabilities. Matching is whole-word
throughout.
Co-Authored-By: Claude <noreply@anthropic.com>
The homelab is retiring *.schweitz.internal and will rebind host
ports to loopback; container-to-container traffic must use container
names on docker-dataplane.
- SEARXNG_HOST: http://localhost:8087 -> http://searxng:8080
(SearXNG's internal port is 8080; 8087 was the host-published port)
- LIBRARY_DESK_HOST: http://localhost:8089 -> http://library-desk:8089
- CORE_API_HOST: http://localhost:8090 -> http://core-api:8083
(8090 is the Scheduler's host port; Core-API serves 8083 internally,
confirmed by the housekeeper client and test suite hitting :8083)
- scripts/test_housekeeper.sh: reach Core-API via localhost:8083
instead of the LAN IP, which will refuse after loopback rebinding
Local development against host-published ports keeps working via .env
overrides (.env.example unchanged; localhost stays valid on the host).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
search_wiki printed ordinally numbered results with no page ID while
get_wiki_page demands 'the page ID from search results' - the model
passed the list position (page 1) and 404'd. Results now carry
page_id and drop the ordinals.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
gemma thinking-only assistant turns carry content: null with no
tool_calls, slipping past the tool-call-only sanitizer and 400ing the
whole agent run ('invalid message content type: <nil>'). Null content is
now blanked for any role.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The scholarly persona prompt reproduced the exact pathology
TATLOCK_ORCHESTRATION_PROMPT fixed for the butler: gemma4 answered in
character ('please provide your request') without calling a single tool.
The research phase now uses a tool-discipline prompt; Tatlock's synthesis
supplies the voice. Anti-fabrication rules kept verbatim.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Post-coordination-removal sweep: the coordination wire protocol
(AgentRequest, AgentResponse, DelegationIntent, CoordinationResult,
DelegationReason, TaskComplexity, ToolCallRecord, AgentTimeoutError,
AgentUnavailableError, DelegationError) had zero importers left in
src/ - only its own test module. AgentError stays (raised by
run_librarian, mapped to user-safe failures by delegation.py).
Also drops the stale coordination.py line from the README tree.
Import-cycle sanity: python -c 'import src.main' passes.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QbFZyDvYksazX6nYQYZ67L
Phase A review minors:
- Coverage note: source_status (when present) is now used exclusively;
the source_counts-absence fallback only considers the optional legs
the request explicitly enabled (web/documents/volatile). library-desk
computes source_counts from the final top-N fused results only, so
absence of the always-on vector/graph legs is normal ranking behavior
- the old heuristic warned on virtually every healthy search
- update_wiki_page: the empty-list tags sentinel (leave unchanged) made
clearing all tags impossible; pass exactly ["__CLEAR__"] to send an
empty tag list, documented in the docstring for the local model
- Text-delegation parallel fallback: zip(..., strict=True) with an
explicit count-mismatch guard so results can never be silently
attributed to the wrong agent
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QbFZyDvYksazX6nYQYZ67L
One delegation implementation remains (src/agents/delegation.py).
Removed, after verifying zero live importers post-Phase-A/B:
- src/agents/coordination.py: CoordinationEngine, duplicate
delegate_to_librarian, AGENT_EXECUTORS/AGENT_STREAM_EXECUTORS
(only importer was its own test module)
- run_librarian_stream: documented-broken path (Ollama streaming +
tool call bug, PydanticAI #1292/#2256), only called by the deleted
coordination engine
- stream_delegate_to_* wrappers + STREAMING_DELEGATION_WRAPPERS and
the never-parsed __DELEGATION_RESULT__ marker in delegation.py
- HouseholdRegistry.get_streaming_delegation_tools() (no callers)
- tests/agents/test_coordination.py and the wrapper/stream tests
Note: the STREAMING_DELEGATION_WRAPPERS import in
src/responses/streaming.py was already removed by Phase A (7ce1c1a);
nothing to delete there.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QbFZyDvYksazX6nYQYZ67L
LibraryDeskClient._resolve_user only enforced non-empty: an explicit
user argument to any tenant-scoped method bypassed tatlock's tenant
guard entirely and went straight to library-desk, and padded values
were sent un-stripped on the wire.
Route the explicit-arg path through the same apply_tenant_guard() used
by context resolution and strip whitespace before the empty check, so
a non-production environment can never send the production tenant (or
a sanitization-collision variant) to library-desk, regardless of how
the user was supplied. Defense in depth - no in-repo caller passes an
explicit user today.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The request-level tenant guard compared the raw user string exactly
(user == PRODUCTION_TENANT), but all local namespaces (Qdrant
collections, Redis keys) are derived through sanitize_user_id(), which
lowercases and strips/maps punctuation. Case or punctuation variants
("JPMSchweitzer", "jpmschweitzer.", " jpmschweitzer") therefore passed
the guard yet resolved to the production namespaces, letting a dev
instance on the shared services read/write production tenant data.
- context.py: compare sanitize_user_id(user) against the sanitized
production tenant; expose the guard as public apply_tenant_guard()
- config.py: startup refusal validator uses the same sanitized
comparison, so a colliding DEFAULT_USER refuses startup loudly
instead of relying on the allowlist fallback
- tests: variant matrix at both config and request-context level,
plus a non-colliding passthrough case
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Library-desk is removing its server-side default user, so a request
without an explicit tenant will 422 after its next deploy:
- New client-level _resolve_user() resolves the tenant (explicit arg
or request context) and raises ValueError on an empty/whitespace
value BEFORE any bytes hit the wire; all 15 tenant-scoped methods
use it
- extract_content / extract_content_batch now accept and send the
user (query param), matching the rest of the API surface
- search_web no longer falls back to a phantom "tatlock-librarian"
tenant; it sends the resolved user
- health_check stays user-less (public, not tenant-scoped)
Tests: parametrized sweep pins the wire contract (user present in
params or payload) for every tenant-scoped method, for both context
and explicit users; empty-tenant calls are asserted to fail without
any HTTP call; the recorded-fixture hybrid contract test now pins
user as an explicit query param.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Non-production environments (development/testing) now force the
effective tenant to the reserved test tenant "llm_tester" (or a
test_-prefixed override) regardless of DEFAULT_USER misconfiguration:
- Config.effective_default_user only honors DEFAULT_USER outside
production when it is llm_tester or test_-prefixed; anything else
is forced to llm_tester (tenant_forced flags the override)
- Config refuses startup (validation error) when a non-production
environment is explicitly configured with the production tenant
jpmschweitzer
- get_user() applies the same guard at request-context resolution,
so an explicit request for the production tenant in dev/test is
forced to llm_tester with a warning log
- initialize_application() emits one loud startup log line
(tenant_guard_active / tenant_guard_production) stating the
effective tenant
Unit tests cover the dev/test/prod x default/explicit-user matrix.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- delegate_to_* now receives a trimmed conversation history (last ~6
turns, 500 chars/turn) as context on both live direct-delegation
paths (streaming and steward non-streaming), via new
build_delegation_context helper
- _stream_direct_delegation restructured as an async generator: the
butler 'start' think message streams BEFORE the expert runs and the
success/error message right after it finishes, instead of all
messages arriving after the research completed
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Ollama's OpenAI-compatible API mishandles anyOf[X, null] parameter
schemas. update_wiki_page (content/title/tags/description) and
smart_create_wiki_page (path) now use empty-string/empty-list
sentinels translated to None inside the tool, following the
biographer pattern from 9d7ce39.
Adds a snapshot test that walks every registered librarian tool's
emitted JSON schema and fails on any anyOf[..., null].
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- 2-attempt short-backoff retry for GETs and the read-only
POST /query/* and /rag/search endpoints only; wiki writes are never
retried (duplicate-page risk)
- honor the defined-but-ignored LIBRARY_DESK_TIMEOUT config instead of
hardcoded 60s/30s per-call values
- hold ONE shared httpx.AsyncClient per librarian run via
library_client_session (contextvar), instead of constructing a
client per tool call; nested sessions are no-ops and custom targets
still get their own client
- read tools raise ModelRetry on transient HTTP errors (transport
errors, 5xx, 429) so Agent(retries=2) engages; write tools keep
returning safe failure messages
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- add LIBRARIAN_TIMEOUT config (default 180s) and enforce it with
asyncio.wait_for inside delegate_to_librarian, covering the live
paths (steward direct delegation and SSE streaming) that had no cap
- timeouts fail honestly: success=False with a curated butler sentence,
detail in logs
- set an explicit timeout on TatlockOllamaProvider's AsyncOpenAI client
from OLLAMA_TIMEOUT instead of the SDK default (~600s per LLM call)
- remove the contradictory unused 60s default from
AgentRequest.timeout_seconds; coordination falls back to the
configured budget
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- parse source_counts into HybridRAGResponse and additively parse the
shared-contract source_status/degraded fields when present (absence
tolerated, so deploy order between tatlock and library-desk never
matters)
- hybrid_search appends a one-line coverage note when a leg reported
'failed' (or degraded is set), falling back to inferring silent legs
from source_counts on older library-desk versions
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- run_librarian / run_librarian_stream raise AgentError instead of
returning/yielding error text as normal output; detail stays in logs
- delegate_to_* wrappers now put a curated butler-toned sentence in
DelegationResult.output on failure and never expose str(e), so
streaming's error branch is reachable and honest
- _execute_single_delegation propagates success; direct delegation only
records delegate_to_* as called when the expert actually succeeded
- librarian tools return user-safe messages instead of
'Error searching: {e}' strings that leaked internal URLs into
synthesis; coordination stream errors are curated as well
- ruff cleanups (TYPE_CHECKING forward refs, B904, unused locals) in
the touched files to keep them lint-clean
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The client parsed field names the live library-desk service never
returns, so every result rendered as "unknown (score: 0.00)":
- source_type/sources -> source + sources (icons key off sources values)
- rrf_score -> score
- context -> formatted_context
- related_dossiers are per-result; top level aggregates unique titles
- synonyms live inside the keywords dict as a {term: [synonyms]} map
Also stop sending zero limits (service 422s on limit < 1); disabled
legs now rely on the enable_* flags with limits clamped to >= 1.
Adds a recorded live response as a fixture plus contract tests that
pin the mapping (non-unknown sources, non-zero scores, icon coverage).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Mechanical Optional[X] -> X | None and f-string cleanups so subsequent
librarian changes lint clean against the dirty baseline.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
With the butler persona prompt attached, gemma4 reasons about calling the
calculator and then answers from memory with a different wrong product every
run; tool_choice=required via extra_body is advisory at best on Ollama's
OpenAI-compat layer. orchestrate_tool_calls() now uses a terse
TATLOCK_ORCHESTRATION_PROMPT; synthesize_from_results() keeps the persona,
so the user-visible voice is unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Rolls back the claudification backend preference: PREFER_CLOUD_BACKEND now
defaults to false, resolve_backend() picks Ollama first and uses Claude when
explicitly preferred or when the new Ollama startup health check fails. The
Steward retries mid-request failures on the other backend in both directions.
Also hardens the fallback itself: Anthropic SDK imports are lazy so a broken
anthropic package degrades to Ollama-only instead of crashing at import time
(root cause of the production outage since April), anthropic is pinned to a
pydantic-ai-1.27-compatible range, ANTHROPIC_MODEL defaults to claude-sonnet-5
(sonnet-4-20250514 retired 2026-06-15), sampling parameters are stripped from
Claude calls (Sonnet 5 rejects them), and the Steward timeout is configurable
(STEWARD_TIMEOUT, default 60s) since gemma4 needs ~35s warm for analysis.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
gemma4:e2b has native function calling with dedicated tool tokens,
achieving 100% tool selection accuracy in benchmarks vs 67% for
mistral-nemo-large, with 5-8x faster response times (2-4s vs 15-20s)
and lower VRAM usage (8GB vs 9.2GB).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
exclude_none was too aggressive — it stripped finish_reason: null from
intermediate chunks (which OpenAI includes). exclude_unset correctly
omits only fields never passed to the constructor (like reasoning_content
on content-only chunks) while preserving explicitly-set finish_reason: null.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Remove references to unimplemented get_benchmark_store from steward and
tool tracking tests
- Fix steward test fixture calling async initialize_application synchronously
by using sync register_household_members instead
- Rewrite tool tracking tests to assert actual logging behavior
- Change unit test fixture model from Tatlock to lorem-tester so unit tests
don't require external services
- Add session-scoped _initialize_app fixture to run Claude health check,
ensuring integration tests use Claude instead of falling back to Ollama
- Increase integration test timeouts from 30s to 120s to match OLLAMA_TIMEOUT
- Add Steward reasoning as ReasoningOutputItem in create_response_with_steward
so <think> tags appear in chat completion responses
- Add test_tatlock_ollama_fallback to verify Ollama fallback path works
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
OpenAI's API omits null fields in streaming chunks, but Tatlock was
including them (content: null, reasoning_content: null). This caused
parsing issues in Open WebUI's streaming handler.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
sse_starlette's EventSourceResponse added \r\n line endings that
Open WebUI couldn't parse. Switched to plain StreamingResponse with
manual SSE formatting matching OpenAI's exact format.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The Steward's internal routing analysis (DELEGATE, COMPLEXITY, etc.)
was being exposed in <think> blocks. This is implementation detail,
not useful reasoning for the user.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
PydanticAI handles tool_choice natively for Anthropic. The extra_body
hack caused an infinite tool call loop where Claude kept calling the
same tool because tool_choice was forced to "any".
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
AnthropicModel doesn't accept api_key directly; it must be passed
through an AnthropicProvider instance.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
All agents now prefer Claude API when ANTHROPIC_API_KEY is configured,
with automatic fallback to Ollama when offline or unconfigured. New
src/anthropic/ module provides model selection via get_model() factory.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add Paperless document search to HybridRAG pipeline
- Add volatile cache (weather, forecast, news, stocks) to HybridRAG
- Add include_documents and include_volatile params to hybrid_search
- Add 📑 and ⚡ icons for document/volatile sources
- Update Librarian prompt with new data source awareness
- Fix Biographer routing: personal memory queries now route correctly
- Add location keywords to Steward pre-fetch logic
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Strengthened personality prompt to prevent unnecessary apologies after
successful Librarian delegations. Added explicit "do NOT apologize"
instructions to both system prompt and synthesis prompt.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add explicit instructions to the Librarian system prompt to never
invent data when tools fail or data sources are unavailable.
- Report what failed specifically
- Never provide placeholder or made-up data
- Better to return no information than fabricated information
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Instrument the full request flow with trace spans for debugging:
- Wrap expert delegations (librarian/biographer/housekeeper) in spans
- Add orchestrate and synthesize spans to TatlockAgent
- Trace Steward analysis in preprocessing
- Start/end traces in response service with context management
- Simplify router by moving context handling to service layer
- Include tracing router in debug mode
- Remove benchmark recording from tool_tracking and steward service
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Remove the Redis-backed performance benchmarking in favor of the new
lightweight file-based tracing system which provides better debugging
capabilities for local development.
- Delete src/core/benchmarks.py
- Remove ENABLE_BENCHMARKS, REDIS_BENCHMARK_DB, redis_url from config
- Update memory_cache comment (now uses DB 1)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Adds JSON-based tracing system for local development that captures
the full request flow through Tatlock's multi-agent architecture.
- Trace/Span dataclasses with automatic timing and nesting
- Context-var based propagation for async-safe tracing
- trace_span async context manager for clean instrumentation
- Traces written to logs/traces/{trace_id}.json
- REST API for listing and retrieving traces (/traces)
- Standalone HTML viewer with timeline visualization
Enabled via DEBUG=true environment variable.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Rewrite system prompt with negative constraints and step-by-step process
- Set temperature to 0.1 for deterministic tool calling
- Sort room groups to top of device list (address positional bias)
- Add [ROOM GROUP] marker in list_devices output
- Update tool docstrings with explicit entity_id= parameter examples
- Add optimization findings doc (experiment log: 0% → 100% success)
- Add test script for room group detection regression testing
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Update all client endpoints to use /housekeeping/ prefix
- Add critical rule requiring list_devices() before control actions
- Add housekeeping API spec documentation
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Convert booleans to strings for Redis hset (Redis doesn't accept bool)
- Extract capability from delegate_to_X tool names for tracking
- Use loop_scope="module" for pytest-asyncio module-scoped fixtures
- Add note about using venv for tests in AGENTS.md
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Messages in reasoning_content should be plain text, not wrapped
in <think> tags. Removed wrappers from:
- delegation.py household think messages
- orchestration.py status messages
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
