feat: per-model inference parameters with extended thinking

[philmerrell]

Summary

• Replaces the global temperature/max_tokens knobs with a per-model supportedParams map keyed by canonical name (temperature, top_p, top_k, max_tokens, thinking, reasoning_effort, …). Admins author it on the model form, the runtime translates canonical names into provider-native shapes (Bedrock / OpenAI / Gemini), and users can override per-request from a new Settings → Advanced panel.
• Anthropic extended thinking on Bedrock works end-to-end: the canonical thinking value is an int budget that the translator wraps into {type: "enabled", budget_tokens: N} under the field Strands' BedrockConfig actually forwards (additional_request_fields, not the originally-attempted additional_model_request_fields). Sampling params that conflict with thinking are suppressed before dispatch.
• Validation at four layers so an admin or user can't ship a request Bedrock will reject: per-row form validators (min ≤ max, default ∈ [min, max]), thinking-specific FormArray validator (budget ≥ 1024, budget < max_tokens.default), Pydantic SupportedParams._check_thinking_invariants on save, and a final cross-param safety drop in _merge_inference_params at request time.

Backend

• backend/src/apis/shared/models/models.py — new ModelParamSpec + SupportedParams types added to ManagedModel/Create/Update. _check_thinking_invariants model_validator catches budget < 1024 / >= max_tokens at config time.
• backend/src/agents/main_agent/core/model_config.py — per-provider canonical → native translation tables, dot-path nested writes, thinking value shaping, and incompatible-sampling-param suppression. top_k and thinking route through additional_request_fields for Bedrock.
• backend/src/apis/inference_api/chat/routes.py — _merge_inference_params layers admin defaults + request overrides, respects locked: true (admin default wins), clamps numerics to bounds, and now drops thinking if >= max_tokens after the merge so direct API callers can't trigger a ValidationException from Bedrock.
• backend/src/apis/inference_api/chat/models.py — InvocationRequest.inference_params field.
• backend/scripts/seed_bootstrap_data.py — seeds default Claude chat models with a temperature/top_p/max_tokens supportedParams block.

Frontend

• frontend/ai.client/src/app/admin/manage-models/model-form.page.{ts,html} — full editor for supportedParams with per-row + cross-row validators (paramRowBoundsValidator, thinkingInvariantsValidator), inline error display, and a new thinkingBudget input kind. Save button is gated on modelForm.invalid so bad rows can't ship.
• frontend/ai.client/src/app/components/model-settings/model-settings.{ts,html} — collapsible Advanced section that renders inputs from selectedModel().supportedParams. Bounds are merged from catalog defaults + provider-specific bounds + admin spec; the thinking row's max is tightened to max_tokens − 1 and the toggle is disabled with an explanation when that window collapses below 1024. Auto-clears the thinking override if max_tokens drops below the budget. Transient amber clamp notices appear when typed values get coerced.
• frontend/ai.client/src/app/admin/manage-models/models/managed-model.model.ts — ModelParamSpec, SupportedParams, KNOWN_PARAMS catalog, and the new kind: 'thinkingBudget' + incompatibleWith fields.
• frontend/ai.client/src/app/session/services/model/model.service.ts — _inferenceOverrides signal keyed by modelId, persisted in sessionStorage.
• frontend/ai.client/src/app/session/services/chat/chat-request.service.ts — sends inference_params on the chat request body when overrides exist.

Test plan

• Backend: cd backend && uv run python -m pytest tests/agents/main_agent/core/test_model_config.py tests/agents/main_agent/property/ tests/agents/main_agent/test_fixtures_smoke.py (76 passing, includes new tests for the Bedrock thinking shape, top_k routing, sampling-param suppression, and disabled-thinking passthrough)
• Frontend: cd frontend/ai.client && npx ng build --configuration development (template type-check clean)
• Admin form: add a model, configure temperature/top_p/max_tokens/thinking with bad bounds (min > max, default out of range, thinking < 1024, thinking >= max_tokens); verify each shows an inline red bullet and Save is disabled.
• Admin form: save a Claude Sonnet 4.5 model with thinking default: 4096 and max_tokens: 8192. Confirm DDB round-trips supportedParams.
• Chat: with that model selected, open Settings → Advanced. Toggle Extended Thinking on; send a non-trivial prompt. Verify the response renders a reasoning content block before the text (proves additional_request_fields.thinking reaches Bedrock).
• Chat: lower Max Output Tokens to 800 → thinking toggle disables with caption "Set Max Output Tokens above 1024 to enable extended thinking (currently 800)."
• Chat: set Max Output Tokens to 5000, enable thinking with budget 4000, then drop Max Output Tokens to 3000 → thinking auto-clamps to 2999 with amber caption.
• Backend safety net: edit sessionStorage.inferenceParamOverrides directly to {"<modelId>": {"max_tokens": 1000, "thinking": 1024}}, send a prompt. Should succeed without reasoning, with Dropping thinking budget 1024 ... not less than max_tokens 1000 warning in inference-api logs.

Notes / follow-ups

• Tool-use round-tripping: this PR doesn't touch the agent loop, so if Strands' BedrockModel doesn't preserve Anthropic thinking content blocks (with signature) across tool-use turns, enabling thinking on a tool-using prompt will 400 from Bedrock on the second call. Worth a separate ticket — verify behavior before turning thinking on for any prod model.
• Phase 2 (per-user override surface) hooks already in place via ModelParamSpec.locked and the _inferenceOverrides signal — this PR is the admin-authored side; locked-respecting user UX is built but only exposed via the Advanced panel today.

🤖 Generated with Claude Code

[philmerrell]

Tracking issue created: #204 — captures the user-facing approach this PR ships and the follow-up work (per-user defaults, locked-indicator UX, tool-use × thinking round-tripping, convergence with #202). Marked as related to #201 and #202.