fix: expose per-model thinking tiers and forward reasoning_effort

[m1yuan]

Problem

pi/OMP hide the xhigh and max thinking tiers for all Command Code models, and silently clamp them to high. Root cause: every model is registered with reasoning: true but no thinkingLevelMap, and pi-ai's getSupportedThinkingLevels gates xhigh/max behind an explicit non-null map entry.

Fix

Two changes:

1. Per-model thinkingLevelMap — each model registers a map built from its supported reasoning efforts, so pi exposes exactly the tiers the model supports (e.g. DeepSeek V4 Flash shows off/high/max; Claude Sonnet 5 shows low→max). Unsupported tiers map to null and are hidden from the UI.

2. reasoning_effort passthrough — the selected thinking level is forwarded to /alpha/generate as params.reasoning_effort, matching the official CLI. The off level and unsupported tiers (mapped to null) are omitted.

Where MODEL_EFFORTS comes from

The Provider API (/provider/v1/models) exposes no effort metadata — every model entry is just {id, name, context_length}. The MODEL_EFFORTS table in src/models.ts (27 models) was parsed from the official command-code CLI 1.7.0 bundle (dist/cli.mjs), which carries a reasoningEfforts array on each model definition.

This table will drift as upstream adds models or changes supported efforts. Re-extract by scanning the latest CLI bundle for id:"..." ... reasoningEfforts:[...] pairs. Models absent from the table get no thinkingLevelMap, preserving pi's default off→high tiers.

Evidence that /alpha/generate accepts reasoning_effort

Confirmed by intercepting the official CLI's real outbound request (NODE_OPTIONS fetch hook wrapping globalThis.fetch):

POST https://api.commandcode.ai/alpha/generate
params: { ..., reasoning_effort: "max" }   // with --effort max

And the CLI rejects unsupported efforts per-model:

$ command-code --model deepseek/deepseek-v4-flash --effort low
Unknown effort "low". Supported: high, max.

Additional change

COMMANDCODE_MODELS_URL is now resolved inside the entry function instead of at module load, so the env override is honored at call time (was fixed during import, breaking testability).

Testing

• 5 new tests in test-thinking-levels.ts: per-model tier visibility, reasoning_effort passthrough, off omission, null (unsupported) omission, unknown-model fallback.
• All existing tests pass (66 total, 0 failures).
• tsc --noEmit introduces 0 new errors (3 pre-existing from optional pi-coding-agent peerDep).
• Verified against live Provider API: 27/50 models mapped, deepseek/deepseek-v4-flash correctly yields ["off","high","max"] with max visible.

[patlux]

Thanks for tackling this — the underlying pi issue is real and forwarding supported reasoning_effort values is the right direction. I found three blockers that should be addressed before merge:

1. Keep MODEL_EFFORTS exactly aligned with the authoritative command-code@1.7.0 effort catalog.
2. Do not infer a per-model default that upstream does not publish.
3. Emit canonical OMP thinking metadata (mode: "effort" plus efforts) separately from pi's legacy thinkingLevelMap.

I verified that command-code@1.7.0 is currently npm's latest release. Its generated model catalog says models without an effort entry decide their own reasoning depth.

Non-blocking follow-ups:

• Use mkdtemp()/tmpdir() instead of the shared hard-coded /tmp/commandcode-models-thinking-levels-test.json path, and clean it up in finally.
• Please add @m1yuan to the Unreleased contributors section.
• A regression fixture/test containing the exact 1.7.0 effort catalog would help prevent unsupported entries from being introduced again.

[patlux]

The checked-in table does not match the authoritative command-code@1.7.0 catalog. These five entries have no adjustable effort surface in the published CLI: claude-haiku-4-5-20251001, moonshotai/Kimi-K2.5, moonshotai/Kimi-K2.6, meta/muse-spark-1.1, and tencent/hy3-paid. The CLI's generated reference/models.md marks their Efforts column as — ("models without an effort column entry decide their own reasoning depth"). Please remove these entries; otherwise the provider exposes and sends effort values that the official CLI does not support. The remaining 22 entries match the 1.7.0 effort map.

[patlux]

OMP 17.2.x expects canonical ThinkingConfig: { mode: "effort", efforts: [...], defaultLevel?, effortMap? }. mode is required, while this effortMap also contains legacy pi-only off/null values that do not belong in OMP's string-valued map. Please keep the legacy thinkingLevelMap for pi, but emit OMP metadata separately, e.g. thinking: { mode: "effort", efforts: [...efforts] }. An OMP identity map is unnecessary because omitted mappings already pass the supported effort through unchanged.

[patlux]

Please do not infer defaultLevel from array position. OMP actively applies this value on initial model selection and every model switch, so this makes Claude/GPT-5.6 default to xhigh, Gemini/Grok to medium, etc. The Command Code catalog publishes supported efforts but no per-model default, so this changes runtime behavior and potentially cost without an upstream source. Omit defaultLevel unless an authoritative default becomes available.