fix(gc): honour Keep Alive, widen grace to 30d, pin Zed missing-thread auto-repair

[lukemarsden]

Fixes the incident where a spec-task session came back with a permanently blank agent panel, and the related "long running tasks stop being able to use their own shell".

Paired Zed PR: helixml/zed#79

What happened

spt_01kz6r8evrdtpepqd59sjm0eev, pinned with Keep Alive ON:

Aug 5 14:04:06  find-ai#23 merged
Aug 5 14:04:27  task correctly moves to done (terminal)
Aug 5 14:07     user still working in the session
~Aug 5 20:04    6h grace elapses -> os.RemoveAll on the task workspace
                WHILE the container is running
                -> shell breaks (cwd deleted), .claude-state destroyed
Aug 10 11:40    restart -> empty workspace; zed_thread_id points at a session
                claude-agent-acp no longer has -> load_session "Resource not
                found" -> silent fallback to an empty "New Zed Agent Thread"

The done-transition was correct — the task's only PR really had merged. The bugs are downstream of it.

1. Keep Alive was invisible to the reaper

liveSpecTaskIDsForReaper decided liveness purely from terminal status and mtime. Grepping KeepAlive across api/pkg/hydra and api/pkg/external-agent returned nothing — the flag protected the container from idle-shutdown but not the workspace, so an explicit "leave this alone" was silently overridden.

22 tasks are currently done with keep_alive = true, so this was not a one-off.

Keep Alive now outranks the status heuristic entirely, and is deliberately not time-bounded: expiring an explicit "keep this" after N days reintroduces the same surprise on a longer timer.

2. Grace period was 6 hours

Nominally a race guard, but reaping a task workspace destroys conversation history, so the window really means "how long after a task finishes might someone want to pick it back up". At 6h a PR merged at 14:04 lost its workspace the same evening.

Raised to 30 days. Bringing a finished task back weeks later is normal, and losing its history isn't an acceptable price for reclaiming disk early.

3. (Zed side) Auto-repair now actually fires

The recovery path here — recoverMissingThread + isAuthoritativeMissingThreadError — was already built and correct. Zed just never sent the thread_load_error that triggers it, because it suppressed failures with no request_id (which is exactly the reconnect case). Fixed in helixml/zed#79, so a dangling pointer now self-clears.

Validation

• go build + go test ./api/pkg/external-agent/ — pass
• New regression test asserts Keep Alive survives terminal + archived + stale, while an unpinned stale terminal task is still reapable so the GC keeps reclaiming. Verified it fails when the KeepAlive check is disabled.

Known follow-ups (not in this PR)

• The reaper is still blind to whether a container is running.
• A live work session still doesn't protect its own spec-task workspace: spec-tasks/ is reaped against liveSpecTaskIDs only, so liveSessionIDs offers no protection.

🤖 Generated with Claude Code

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.