feat: per-model control of which sampling params are sent to the backend

[ikarys]

Summary

Some backends reject sampling parameters they don't support (e.g. temperature, top_p, top_k, reasoning_effort), returning HTTP 400 and blocking the conversation. OpenFox already stripped these via an omitParams field set by auto-config, but users had no manual control over which params are sent.

This PR adds a "Send" checkbox next to each sampling parameter in the provider model editor (Advanced section), so users can explicitly enable/disable individual params per model. Unchecking a param adds it to omitParams and strips it from outgoing requests; re-checking removes it.

What changed

• ProviderModal: New SamplingParamField component with a "Send" checkbox per param (Temperature, Top P, Top K, Max tokens). Unchecking disables + dims the input and adds the param key to omitParams. Re-checking re-enables it. Pre-existing omitParams entries (e.g. reasoning_effort from auto-config) are preserved on toggle.
• ProviderModal: Conditional "Re-enable reasoning_effort" checkbox in the Thinking section, shown only when auto-config omitted it. Lets users re-enable reasoning_effort without re-running auto-config.
• ProviderModal: Omitted fields now show a "Not sent" placeholder instead of a stale dimmed value, making it clear the value won't be sent.
• Refactor: Collapsed 4 near-identical sampling-param blocks (~80 lines) into one parameterized SamplingParamField component. Net LOC reduction.
• Comments: Genericized all vendor-specific comments and test fixtures to vendor-neutral wording.

How it works

omitParams: ['temperature', 'top_p']  →  temperature and top_p stripped from request body
omitParams: undefined                  →  all params sent (default)

The omitParams field was already wired through the full stack (ModelConfig → provider-manager → client-pure → request body). This PR exposes it in the UI so users can control it manually, without relying solely on auto-config detection.

Related issues

• bug: Mistral API returns HTTP 422 when thinking is enabled (reasoning field rejected in message history) #220 — A backend returns HTTP 422 when thinking is enabled (a field rejected in message history). Same class of problem: a backend rejecting a field OpenFox sends. This PR gives users manual control over which params are sent, addressing the broader pattern. (Note: bug: Mistral API returns HTTP 422 when thinking is enabled (reasoning field rejected in message history) #220 is specifically about reasoning in message history, handled by the existing sendReasoningInMessages flag — but the underlying need for per-field control is the same.)

AI-Enhanced Development

• AI Models: glm-5.2, claude-opus-5

Cache Impact

• No — omitParams strips generation-time parameters (temperature, top_p, top_k, max_tokens, reasoning_effort) from the request body. These are not part of the cached prompt prefix (system prompt + tools + messages) or the tool fingerprint. The dynamic context hash and KV cache are unaffected.

Test plan

• npm run typecheck — clean
• npm run lint — clean
• npx prettier --check (modified files) — clean
• npx vitest run web/src/components/shared/ProviderModal.test.tsx — 18 passed
• npx vitest run src/server/llm/client-pure.test.ts src/server/providers/auto-config.test.ts src/server/provider-manager.test.ts — 75 passed
• New tests: Send checkboxes rendered + checked by default, uncheck → omitParams + dim + clear, recheck → removed, preserve pre-existing entries, auto-config omitParams reflected, reasoning_effort re-enable

[co-l]

Merged, will be released in v2.0.119

AI-summary

Adds per-model control over which sampling parameters are sent to the LLM backend. Auto-config now probes each model to detect params the backend rejects (with a control-probe guard added during review to avoid false attribution), and the builder strips them from outgoing requests. The UI exposes per-param "Send" checkboxes plus a recovery control for reasoning_effort.

---

• omitParams on ModelConfig — strip rejected top-level sampling params (temperature, top_p, max_tokens, top_k, reasoning_effort) from outgoing requests; runs after all merges so it wins over queryParams additions
• Auto-detection — auto-config probes each model with an agentic-loop-shaped request (incl. a dummy tool) and reports params rejected via HTTP 400; a control probe skips detection when the backend rejects the probe structurally, and word-boundary matching avoids false positives (e.g. max_tokens_budget ≠ max_tokens)
• Accurate stats/retries — modelParams (turn stats + truncation-retry budget) now reflects the actual wire request, so omitted params are neither sent nor reported as sent
• UI — per-sampling-param "Send" checkboxes with disabled inputs, a "Re-enable reasoning_effort" recovery checkbox, and auto-config results pre-populating the checkboxes
• Plumbing — threaded through provider config, provider-manager, the providers API, and the streaming request path