[insuffer]

What

Fixes #432: auto-generated session titles could leak the title model's own chain of thought. cleanGeneratedTitle only trimmed quotes/prefixes/punctuation, so when a reasoning model answers the title prompt with a <think>…</think> block, the raw reasoning text was stored as the session title (the sidebar showed titles like <think>The user is asking about…).

cleanGeneratedTitle now runs the existing stripThink helper first, so only user-visible text survives. Both title paths go through it (WebuiTitleService.generateTitle and maybeGenerateWebuiTitle).

Why

TITLE_GENERATION_REASONING_EFFORT is "none", but inherently-reasoning models (DeepSeek-R1, GPT-OSS with reasoning, …) ignore that and still emit thinking blocks, so the leak happens in practice — see the screenshots in #432. When the model produces only thinking, the title is now left empty instead of storing the reasoning text, so the session keeps its default title rather than exposing chain-of-thought.

How it was verified

Windows 11, workspace App/memmy-agent:

• 8 new unit tests for cleanGeneratedTitle: think/thought blocks (closed, multiline, unclosed), thinking-only → empty title, whitespace/punctuation still collapsed after stripping, backticked <think> mentions in prose preserved, empty/null input
• 2 new service-level tests in webui-title.test.ts: a provider answering <think>…</think>需求范围整理 stores 需求范围整理; a provider answering with thinking only stores no title
• Each new test was confirmed to fail on the unpatched tree (verified by reverting only the source change and re-running)
• tsc -p tsconfig.json --noEmit passes
• Full workspace suite on Windows: 5590 tests, 44 failures — the same pre-existing flaky/platform failures as an unpatched checkout of main (43 failures there, with membership shifting run to run among the same timing-sensitive files: exec-session-tools, channel-gui-sync, session-dag/builder-queue, plus platform-only document-parsing/build-assets/office-skills). No title-related test fails, and both new test files pass in the full run

Notes

• Reuses stripThink (src/utils/helpers.ts), the repo's established stripping mechanism, which already preserves backticked/prose mentions of <think> and Claude-family ant* artifact tags
• stripThink covers think/thought tags but not <reasoning> blocks; covering that inside stripThink itself would affect every call site, so it seemed better as its own change if maintainers want it
• Two suggestions from [Bug] 自动生成的会话标题会泄漏模型思考内容（包含 <think> 标签） #432 are intentionally left out to keep this PR focused, happy to follow up: falling back to the first user message when the model returns only thinking, and adding an explicit "do not include your reasoning" line to the title prompt

[insuffer]

Hi — first contribution to this repo. This fixes #432: when a reasoning model answers the title prompt with its own <think> block, the raw chain-of-thought was stored as the session title. The fix runs the repo's existing stripThink helper inside cleanGeneratedTitle and adds 10 tests (8 unit + 2 service-level), each verified to fail on the unpatched tree. The Linux CLI installer workflow run is waiting for fork approval — could a maintainer approve it so CI can validate the change? Happy to address any feedback.