Expand description
One-shot run-title generation.
The dashboard displays, searches, and persists RunMetadata.title; this
module is what fills it in. At spawn, the daemon marks an eligible run
PendingTitle; dispatch_title makes one cheap LLM call over the
task prompt via the title_bridge worker, and collect_title
sanitizes the reply into the metadata. Everything downstream (persistence,
dashboard header, run search) already reads the field.
Best-effort by design: any failure - no usable provider or model, a full
pool that never frees, a provider error, an empty reply - leaves the title
None and the run displays its blueprint name exactly as before. The
title lands on disk with the next persistence write (the run-level
heartbeat guarantees one within a few seconds).
Structs§
- Awaiting
Title - A title call is in flight. Unlike
AwaitingCompaction, this does not hold the agent out of inference - titling runs alongside the first turn. - Pending
Title - This run wants a title;
dispatch_titlepicks it up on the next tick. Inserted at spawn for enabled, root, non-empty-task runs only. - Title
Results - The receiving end of the title-outcomes channel, as a world resource.
- Title
Settings - The
[title]config, as a world resource (inserted by the daemon at setup). Absent in worlds that never title (tests,lev runwithout it). - Title
Sink - The sending end, cloned into each spawned title job.
Functions§
- collect_
title - Collect system: store each finished title into its run’s metadata. A provider error or empty reply changes nothing; either way the in-flight marker comes off.
- dispatch_
title - Dispatch system: start the title call for each
PendingTitlerun.