fix(coding-agent): request low reasoning for session title generation

[ImStillBlue]

Summary

The background session-title request left reasoning unset in buildTitleOptions(). For reasoning-capable models, pi-ai's OpenAI-completions layer then fills the gap with the provider's "disabled" mapping — on OpenRouter-format models that is reasoning: { effort: "none" }. Endpoints that mandate reasoning (observed with Z.ai GLM 5.x models, e.g. z-ai/glm-5.3-flash) reject that with:

Runtime error (session_title_generation): Reasoning is mandatory for this endpoint and cannot be disabled. (HTTP 400)

Regular agent turns are unaffected because they always carry the session's thinking level — only the background title call (and therefore every fresh session on such a model) failed. The same behavior is publicly documented against OpenRouter in openclaw#24851.

What changed

• packages/coding-agent/src/core/session-title-generator.ts — buildTitleOptions() now sends reasoning: "low" explicitly, and raises maxTokens from 64 to 1024 so low-effort reasoning tokens don't consume the budget needed for the <title> output on token-based providers.
• packages/coding-agent/test/session-title-generator.test.ts — deterministic regression test asserting the title request carries reasoning: "low" and maxTokens: 1024 via a fake streamFn (no network, no tokens).
• packages/coding-agent/src/core/changes.md — fork-ledger entry per the changes.md contract.

Non-reasoning models are unaffected: pi-ai clamps the level per model and the OpenRouter reasoning branch only fires for model.reasoning === true.

Validation

• bunx vitest run test/session-title-generator.test.ts — 12/12 pass (including the new regression test).
• Root bun run check — pass (Biome, pinned-deps, ts-imports, shrinkwrap, install-lock, claude-sdk-platform-lock, tsc --noEmit, browser-smoke).
• Manually verified against a local install of omo 5.0.0-0.beta.31 / senpi 2026.8.31: patching buildTitleOptions the same way eliminated the recurring session_title_generation 400 on OpenRouter z-ai/glm-5.3-flash sessions.

Residual risk

• Titles on reasoning-capable models now include a small amount of reasoning output, marginally increasing cost per one-shot title call (bounded by maxTokens: 1024). An alternative would be retrying with a low effort only when the endpoint rejects disabled reasoning, but the simple explicit level avoids a second round trip for a known-broken class of endpoints.

---

Summary by cubic

Fixes session title generation on reasoning-mandatory endpoints by retrying once with low reasoning when the provider rejects the default reasoning-free title request, so new sessions on models like Z.ai GLM 5.x no longer fail with a 400 error.

• Default title requests stay reasoning-free with maxTokens: 64, so healthy endpoints pay no reasoning cost.
• Only a "Reasoning is mandatory" 400 triggers the retry with reasoning: "low" and maxTokens: 1024, leaving room for the <title> output once reasoning tokens count against the budget.
• Adds regression tests covering the default path, the fallback retry, and non-matching errors (no retry).

^{Written for commit 94960ba. Summary will update on new commits.}

[cubic-dev-ai]

All reported issues were addressed across 3 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[ImStillBlue]

Valid review — pushed 94960ba addressing it.

Confirmed both claims: DEEPSEEK_THINKING_LEVEL_MAP / OPENROUTER_DEEPSEEK_THINKING_LEVEL_MAP map low: "high", and the unconditional reasoning: "low" enabled thinking on every title call across all reasoning-capable providers (Anthropic included, which previously had thinking disabled for titles).

Redesign per your suggestion: the default title request stays reasoning-free with maxTokens: 64; generateSessionTitle() now retries once with reasoning: "low" + maxTokens: 1024 only when the provider error matches "Reasoning is mandatory". Healthy endpoints keep the zero-reasoning path; broken endpoints pay one extra round trip (bounded by the same narrow retry policy). Added three tests: default attempt carries no reasoning, the reasoning-mandatory 400 triggers the fallback with low/1024, and other provider errors do not.

Validation: vitest run test/session-title-generator.test.ts 14/14, root bun run check pass.