2 Commits

Include renames

Author	SHA1	Message	Date
jpmschweitzerandClaude Fable 5	1769ec2803	feat(backend): adapt every completion to the detected backend (workspace T-137) ... One choke point in the sanitized client: the flavor is probed once (the boilerroom wrapper names itself on /health, a bare llama-server serves /props, Ollama answers neither) and every completion adapts. The agents' tool_choice "required" survives only on Ollama — advisory there, enforced by llama-server, an unbreakable tool loop through the wrapper. Through the wrapper every completion carries webber's session identity: session webber, eviction_order 20 in the decided ranking, configurable via settings. The wrapper's balancing and compaction signals are read from the response body's extra fields — an httpx event-hook variant was tried and never fires under the openai SDK. The enabling fix: the sanitized client was never in the request path. The provider assigned self._openai_client, an attribute nobody reads — OllamaProvider.client serves self._client — so every completion has bypassed the null-content sanitizer since the class was introduced. Exposed when the wrapper 503'd a session-less request the choke point should have named; the client now goes through the constructor's official openai_client parameter, and a wiring test pins provider.client to the sanitized type. The same bug exists in tatlock (its T-6, filed). Verified against the live wrapper from the dev server: flavor boilerroom detected, a tool-using explore run answered in 4.8 s with no tool loop, webber resident at rank 20, and the wrapper parked librarian and tatlock-experts to seat it — the ranking doing exactly its job. Six new tests (227 green), five mutation-checked: the rank default, the strip condition, the session-add condition, the no-cache-on-failure rule, and the client wiring. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-09-13 10:02:14 +02:00
jpmschweitzerandClaude	3e495daa73	fix(webber-api): clear mypy, and the dead code it was covering for ... 55 errors to zero. Nearly all of them traced back to two causes rather than 55. THE DECORATOR. @logged wraps ~24 functions across this package and was declared `def decorator(func: Callable):` with no ParamSpec and no return annotation, so it erased the signature of everything it touched. ToolResult.execute() is annotated `-> ToolResult`; through the decorator it came back Any, and mypy reported 33 no-any-return errors spread across the tools and agents. Each looked like a local annotation slip. All of them were one decorator. Typed with ParamSpec/TypeVar; the async branch casts at the await rather than loosening R, because loosening R would put the Any straight back into every caller. THE MISSING TYPE PARAMETER. BaseAgent was not generic, so _create_agent returned a bare Agent — Agent[Any, Any] — and pydantic_ai then typed every run() result as Any. BaseAgent is now Generic[CtxT] bound to AgentContext, _agent is declared on the base instead of reached through hasattr, and the three tool-registration functions take their agent's real context type. tools_streaming.py already did this; the other three had not been updated. Eight `execute` overrides carry a targeted ignore rather than a package-wide disable_error_code. Every tool narrows the base's **kwargs to its own named parameters, which is a real LSP violation — but nothing anywhere is typed as BaseTool, and every call site constructs the concrete tool. The abstract method earns its place by making a tool without execute impossible to instantiate. The reasoning lives in BaseTool.execute's docstring; the per-site suppressions mean an override that IS unsound still gets caught. BaseAgent.run_stream widened to AsyncIterator[str | StreamEvent], which is what callers already receive: task streams structured events, explore and plan stream strings, and the router branches on isinstance with a comment calling the string path legacy. The annotation now says what the code does. AND THE PART THAT MATTERS MORE THAN THE TYPES. Chasing the last error found that the Ollama sanitiser has been broken. It fetched the parent's chat getter with `AsyncOpenAI.chat.fget`, and openai made `chat` a functools.cached_property, whose getter is `.func`. Touching `.chat` raised AttributeError — meaning the content: null workaround that CLAUDE.md documents as live would have failed on the first completion any agent attempted. Confirmed in the running container (openai 2.46.0) as well as locally (2.15.0). Two things hid it. The line carried a bare `# type: ignore`, which suppressed precisely the complaint that would have caught it. And /agents/run and /agents/stream have served zero requests in 30 days, so nothing exercised the path. A mitigation can rot completely while every check stays green, if no check actually runs it. The lookup now reads whichever getter the descriptor exposes and raises a legible TypeError if openai adopts a third shape. tests/test_ollama_provider.py walks the chain an agent request walks, short of the network call — mutation-checked: all four fail against the old lookup. 215 passed, 23 skipped, plus the four new. mypy clean over 90 files. Co-Authored-By: Claude <noreply@anthropic.com>	2026-08-11 16:27:51 +02:00