Wrangle your threads.

Deterministic first, model last

Most of what you want to know is already in local state: whether a turn is running, whether it failed with a structured error, whether it is a registered automation, whether it was interrupted, and whether the agent left a status footer on its last message. ThreadBear reads that from a read-only local index and decides without a model call. Every classification records how it was reached — runtime, structured_error, automation, interruption, footer, luna, or unknown — so the provenance of every answer is auditable rather than implied.

An unchanged heartbeat costs nothing

The heartbeat runs from a macOS LaunchAgent rather than a Codex automation, precisely so an idle cycle can exit against local state without spending a token. The prototype that preceded ThreadBear did the opposite: it ran its classifier as a long-lived Codex task, accumulated more than five million input tokens, and replayed roughly 92,000–138,000 of them on turns that had nothing new to report. Deleting that cost model is the reason ThreadBear exists.

When judgment is required, it is one bounded call

Threads that survive the deterministic cascade are packed into the fewest calls that fit the model's advertised context and sent to gpt-5.6-luna at medium effort in a fresh, non-persisted session — no tools, no environment, no approvals, no history. Nothing accumulates, so nothing gets replayed on the next call. Medium is the default because the benchmark showed it carried a lower false-complete risk than low effort; both the model and the effort are configuration values, so a model deprecation is an edit rather than a release.

It only ever touches the title

A managed title is one canonical status emoji, a durable subject you can recognize, and the current next action when there is one. A subject you edited yourself survives. Prefixes never compound, and a stale next action is removed rather than stacked on top of the last one.

Nothing else moves. ThreadBear does not reorder your sidebar, invent projects, pin threads, edit Codex's global state, or drive the UI — there is no supported, reliable control surface for any of that, so it does not pretend otherwise. Your list stays in the order and the projects you left it in; only what each row says changes. Completed threads retire on their own after they have been complete and idle for 14 days, and nothing that is running, blocked, or waiting on you is ever archived.