🤖 fix: enable xhigh reasoning for gpt-5.2

[ThomasK33]

Enable xhigh thinking for openai:gpt-5.2 by updating the per-model thinking policy.

• Root cause: gpt-5.2 was falling back to the default policy (off/low/medium/high), so any xhigh selection got clamped before building OpenAI provider options.
• Fix: allow xhigh in getThinkingPolicyForModel() for gpt-5.2 (including version-suffixed and mux-gateway forms) and add tests.

Validation:

• bun test src/browser/utils/thinking/policy.test.ts
• make typecheck
• make static-check

---

📋 Implementation Plan

Enable xhigh reasoning for openai:gpt-5.2

Context / Problem

The newly released openai:gpt-5.2 model supports OpenAI’s reasoningEffort: "xhigh", but mux currently cannot actually request xhigh for this model.

Root cause (code-level)

Mux clamps the requested “thinking level” to a per-model capability subset via:

• src/browser/utils/thinking/policy.ts → getThinkingPolicyForModel() / enforceThinkingPolicy()
• This policy is used both:
  • in the UI (Thinking slider/options), and
  • in the backend request builder (via buildProviderOptions() which calls enforceThinkingPolicy()).

Right now gpt-5.2 is not special-cased, so it falls into the default policy:

• Default: ["off", "low", "medium", "high"]
• Result: any attempt to set xhigh gets clamped (typically to "medium"), so OpenAI never receives reasoningEffort: "xhigh".

OpenAI request construction is already correct once xhigh is allowed:

• src/common/utils/ai/providerOptions.ts maps ThinkingLevel → OpenAI reasoningEffort.
• src/common/types/thinking.ts includes xhigh: "xhigh" in OPENAI_REASONING_EFFORT.

Recommended approach (minimal change) — Update thinking policy for gpt-5.2

Net LoC estimate (product code only): ~+10–25 LoC (policy + comments; tests separate)

What to change

1. Allow xhigh for gpt-5.2 in getThinkingPolicyForModel():

   • File: src/browser/utils/thinking/policy.ts
   • Add a special case similar to gpt-5.2-pro and gpt-5.1-codex-max.

   Suggested policy:

   • openai:gpt-5.2 → ["off", "low", "medium", "high", "xhigh"]

   Notes:

   • Keep the gpt-5.2-pro branch above the new gpt-5.2 branch.
   • Use the same “version suffix tolerant” regex style already used elsewhere.
     • Example: ^gpt-5\.2(?!-[a-z]) so it matches gpt-5.2 and gpt-5.2-2025-12-11 but not gpt-5.2-pro.

2. Update comments to match reality

   • File: src/browser/utils/thinking/policy.ts
     • Update the “default policy” comment that currently implies xhigh is only for codex-max.
   • File: src/common/types/thinking.ts
     • Update the comment on OPENAI_REASONING_EFFORT.xhigh (currently says only gpt-5.1-codex-max).
   • Optional (nice-to-have): src/common/utils/tokens/models-extra.ts
     • The gpt-5.2 comment block doesn’t mention xhigh. Add a short note for consistency.

Why this works

Once the policy allows xhigh, the normal request path already does the right thing:

• UI can select/store xhigh.
• Backend uses buildProviderOptions(modelString, thinkingLevel, ...).
• buildProviderOptions() will:
  • preserve xhigh (no clamping),
  • set openai.reasoningEffort = "xhigh", and
  • include reasoning.encrypted_content so tool-use works correctly for reasoning models.

Tests / Validation

1. Update/add unit tests for the policy:

   • File: src/browser/utils/thinking/policy.test.ts
   • Add cases:
     • getThinkingPolicyForModel("openai:gpt-5.2") returns 5 levels including xhigh.
     • getThinkingPolicyForModel("mux-gateway:openai/gpt-5.2") returns same.
     • getThinkingPolicyForModel("openai:gpt-5.2-2025-12-11") returns same.
     • enforceThinkingPolicy("openai:gpt-5.2", "xhigh") === "xhigh".

2. Run targeted tests:

   • bun test src/browser/utils/thinking/policy.test.ts

3. Run repo-wide correctness gates (expected in CI):

   • make typecheck
   • make lint (or make lint-fix if needed)

Rollout notes / UX impact

• The Thinking slider will show an extra step for gpt-5.2 once selected.
• Command palette already offers xhigh; after this change, choosing xhigh on gpt-5.2 will no longer silently clamp back to medium.
• No changes required in provider config (knownModels.ts etc.).

Alternative approach (more scalable, higher scope)

Drive thinking policy from model metadata (e.g., a single authoritative model capabilities table that includes supported thinking levels).

Net LoC estimate (product code only): ~+80–200 LoC

This would reduce future “forgot to special-case model X” issues, but requires designing a shared model-capabilities schema and updating multiple call sites (policy derivation, UI, provider options clamping).

Execution checklist (when switching to Exec mode)

• Edit src/browser/utils/thinking/policy.ts to add gpt-5.2 xhigh support.
• Update relevant comments (policy + thinking mapping).
• Update src/browser/utils/thinking/policy.test.ts with new assertions.
• Run bun test ...policy.test.ts.
• Run make typecheck.

---

Generated with mux