session-titler

Harness sessions are identified by timestamps by default, so a growing list quickly becomes a wall of numbers that is hard to scan at a glance. Existing auto-naming solutions either wait until the session goes idle before calling a model—leaving the list blank while the user waits—or fire a model call on every user message, which multiplies cost unnecessarily. session-titler splits the two ideas into complementary phases so you get a title immediately without paying for it upfront.

Phase 1 fires the instant a user message lands: keywords are extracted from the latest message (code blocks, URLs, and Markdown are stripped; stopwords are filtered) and a readable title appears with near-zero latency and no model call whatsoever. Phase 2 runs in the quiet window after the session goes idle, routing a single refinement call to the cheapest capable model in the registered directory (flash, haiku, mini, etc.). That one call upgrades the keyword string into a natural phrase and produces a one-line session summary in the same request, emitted as a session/caption-note event for list UIs and export tools. Titles follow the message language across Chinese, English, Japanese, and Korean; Latin titles are capped by word count, CJK titles by character count. Duplicate titles are deduplicated, cross-session repeats receive a numbered suffix, and any manual rename permanently halts automatic generation for that session.

If you regularly open many sessions in Harness, keep a long history list, or care about the marginal cost of every inference call, session-titler works out of the box with zero configuration after install and restart. Every behavioral knob—from model selection priority to phase toggles—remains tunable for teams that want a custom policy.