feat: allow classifierReasoningLevel "off" to disable classifier thinking

[bsamek]

classifierReasoningLevel: "off" is rejected by validation, but the runtime already supports it

Summary

AutoModeSettings.classifierReasoningLevel only accepts low | medium | high | xhigh | max, so there is no way to ask the classifier to disable thinking. Meanwhile, createClassifierCompletionPlan() in classifier.ts already has a complete effectiveLevel === "off" execution path (it selects the simple completion and omits the reasoning option) — it's just unreachable via config, because the config validator excludes "off".

Why "off" matters

Several classifier-worthy models support a true non-thinking mode via Pi AI's thinkingLevelMap:

Model	thinkingLevelMap.off	Effect when no reasoning option is sent
fireworks/.../deepseek-v4p1-flash	"none"	anthropic-messages API sends thinking: { type: "disabled" }
fireworks/.../glm-5p2	"none"	openai-completions sends reasoning_effort: "none"
fireworks/.../glm-5p3	null	nothing sent; server default (thinking on)
github-copilot/gpt-5.6-luna	null	clamps up to minimal (→ low effort)

With no reasoning preference (the current behavior when classifierReasoningLevel is absent), these "off-capable" models still run their server-default reasoning on every classification. Requesting "off" explicitly makes the fast stage genuinely non-thinking — a large latency win for a gate that runs on every non-read-only tool call.

Proposed change

Accept "off" in ClassifierReasoningLevel and in config validation. The existing clamping semantics already handle models that cannot disable thinking: clampThinkingLevel(model, "off") walks up to the nearest supported level (e.g. minimal for GLM-5.3 / GPT-5.6 Luna), so behavior degrades gracefully instead of failing.

Full diff attached as 0001-allow-classifierReasoningLevel-off.patch. It touches:

• extensions/auto-mode/types.ts — add "off" to ClassifierReasoningLevel; narrow EffectiveClassifierReasoningLevel so the union shape is unchanged (all Exclude<EffectiveClassifierReasoningLevel, "off"> sites are unaffected).
• extensions/auto-mode/config.ts — add "off" to CLASSIFIER_REASONING_LEVELS; update the diagnostic message.
• docs/configuration.md — document "off" and the clamping behavior.

Verification

Tested locally against v1.16.0 with Pi AI's real model metadata:

• isClassifierReasoningLevel("off") === true; garbage values still rejected.
• createClassifierCompletionPlan(deepseek-v4p1-flash, "off", …) → { effectiveLevel: "off" }, simple-completion path, no reasoning option sent → pi-ai emits thinking: { type: "disabled" } on the wire.
• createClassifierCompletionPlan(glm-5p3, "off", …) → clamps to minimal (same as clampThinkingLevel semantics used for other levels).
• End-to-end: ran pi with classifierModel: fireworks/accounts/fireworks/models/deepseek-v4p1-flash and classifierReasoningLevel: "off"; classifier decisions (fast + detailed stages) round-tripped successfully with no reasoning content in responses.

Suggested test

test("classifierReasoningLevel off resolves to the off path for off-capable models and clamps otherwise", async () => {
  const raw = makeStubCompletion("raw");
  const simple = makeStubCompletion("simple");
  const offCapable = { id: "m", provider: "fireworks", reasoning: true, thinkingLevelMap: { off: "none", low: "low", high: "high" } };
  const plan = createClassifierCompletionPlan(offCapable, "off", raw, simple);
  assert.equal(plan.completeFn, simple);
  assert.deepEqual(plan.reasoning, { mode: "explicit", requestedLevel: "off", effectiveLevel: "off" });
  assert.equal(plan.reasoningLevel, undefined);

  const thinkingLocked = { id: "m", provider: "fireworks", reasoning: true, thinkingLevelMap: { off: null, low: "low", high: "high" } };
  const clamped = createClassifierCompletionPlan(thinkingLocked, "off", raw, simple);
  assert.equal(clamped.reasoning.effectiveLevel, "minimal");
});