Repository navigation
Completion callbacks intermittently fail to wake Mattermost callback-session #191
Description
Summary
agentctl launch claude-code completion callbacks intermittently fail to wake the target --callback-session when it's a Mattermost channel session. The coding agent finishes successfully, but the parent session is never notified, so results sit undelivered until a human pings.
Evidence (2026-06-25, agent personal)
Three sequential builds in repo c-h-/digest-audio, all launched with:
--callback-session "agent:personal:mattermost:channel:6hmcskm5jbyqpcu77qaa41nw8c" --callback-agent personal
| Session | Outcome | Callback woke main session? |
|---|---|---|
| f943d3f4 (Phase 1) | completed ~09:49 | ❌ no — 2h+ silent gap until user pinged |
| 608a6012 (Phase 2) | completed | ✅ (also had a safety-net cron) |
| 5d906345 (Phase 4) | completed | ❌ no — caught only via manual status check |
So ~2 of 3 completions did not deliver to the callback session.
Impact
Finished work is silently stranded; operator must poll agentctl peek/list or rely on out-of-band safety-net crons. Defeats the purpose of --callback-session.
Expected
On agent completion, a wake/notification reliably reaches the resolved --callback-session (Mattermost channel session) the same way it apparently did for 608a6012.
Notes / asks
- Is delivery best-effort and dropped if the callback session is mid-turn / yielded at completion time? All three parents were idle/yielded waiting.
- Any ret/queue for callbacks when the target session is busy?
- Repro: launch with a Mattermost channel callback-session, let parent yield, finish the child.
Workaround in use: per-launch isolated safety-net cron that checks the child and posts results.
Activity