fix(backend): log the call warnings a Provider raises

[willdady]

Closes #411

The problem

The AI SDK reports, per call, that a setting it was handed was not honoured. Nothing in Platypus read the array, so an Operator who set Temperature on an Agent running Bedrock — or on a model that rejects sampling parameters via Anthropic direct — was adjusting a control disconnected at the far end. No error, no log line, nothing in the UI.

Concretely, against the pinned versions: Bedrock drops seed, presencePenalty and frequencyPenalty and clamps temperature into 0–1; Anthropic strips temperature, topP and topK on the models that reject sampling parameters; both warn on a model id they don't recognise.

The change

installProviderWarningLogger() points the SDK's warning hook at pino. Each warning becomes one warn line carrying structured fields — provider, model, warningType, feature/setting, details/message — plus a sentence that stands on its own:

WARN: amazon-bedrock (anthropic.claude-opus-4-5-v1:0) does not support seed, so the value set for it was ignored
WARN: amazon-bedrock (anthropic.claude-opus-4-5-v1:0) did not use temperature as given: 1.5 exceeds bedrock maximum of 1.0. clamped to 1.0

The hook is a process global, which is the reason to use it: one call at startup covers the streaming Chat turn, the unattended run, sub-agent runs, title generation and memory extraction, and a generation call site added later needs no wiring. Assigning it displaces the SDK's own process.emitWarning default, so nothing is logged twice — there's a test for that.

Attribution is provider and model only, which is all the hook carries. Correlating a warning back to a run would mean threading async context through every path, and that is the cost the global avoids.

Notes for review

• Fields, not a stringified object, so an Operator can filter on feature: "did anything drop my temperature?" is the question this exists to answer.
• The clamp wording is deliberate. Bedrock reports a clamped temperature under the same unsupported type it uses for a parameter it drops, so "does not support temperature: … clamped to 1.0" would contradict itself. A warning carrying details reads "did not use X as given", and a test pins that.
• The hook fires per model call, not per turn, so a tool-calling loop repeats its lines once a step. Bounded noise — short, fixed text, only written by Providers that are actually dropping something — but worth a second opinion given Nested ZodError serialization turns one failed turn into thousands of log lines #414 and fix(backend): stop one rejected prompt from filling the log #422.
• Docs ship with the code: the Agents page callout now names the trap and says what to search the log for.

Testing

14 new tests covering each warning variant, a clamped temperature carrying details, several warnings from one call, a call with no provider or model, and silence when there are no warnings — plus SDK-driven cases through generateText and through a streamed ToolLoopAgent (the sub-agent construction). pnpm typecheck, pnpm lint and pnpm test all pass.

🤖 Generated with Claude Code