fix(session): notify parent when subagent sessions finish #35066

Activity

1. opencode-agent

   assigned

   kitlangton

   on Jul 3, 2026
2. opencode-agent commented on Jul 3, 2026

   

   opencode-agentBot

   on Jul 3, 2026 – with OpenCode Agent

   ContributorAuthor

   More actions

   @kitlangton tagging you here for the subagent parent-notification issue. Draft PR: #35041
3. dreampuf

   added a commit that references this issue on Jul 20, 2026

   test(opencode): add regression test for lost terminal-iteration wakeup

   ...

   1391879
4. EnigmaJazz commented on Aug 11, 2026

   

   EnigmaJazz

   on Aug 11, 2026

   More actions

   Occurrence: headless serve — child completion never notifies the parent

   Observed today on opencode 1.18.15/1.18.16 headless opencode serve (Linux), driving an autonomous SDD workflow through the serve HTTP API:

   • 5/5 orchestrator sessions had a foreground task tool part stuck running forever while the child sub-agent session had already finished its work (its parts show complete reads/writes, batch-flushed). The child-completion notification never reached the parent, so the parent turn never resumed.
   • The same binary/providers/network in interactive TUI sessions: 226 task parts across all sessions — 132 completed + 89 errored + 5 running — i.e. 96%+ resolve. The TUI receives the completion; the headless serve does not.
   • The stuck-parent signature appears to ride on a tool-completion envelope drop: in a controlled probe on a freshly-recycled serve, one session ran read/glob/bash — glob and bash flipped to completed with output present, the read stayed running with no output for 300s until the caller aborted. Work completes; the completion envelope selectively never lands; the turn stalls.
   • Recovery on our side works by artifact-checking the child work from disk (the work is done), but the parent session itself never completes — exactly the gap this issue addresses.

   This issue (and the draft PR #35041) would fix the class we are hitting at scale. Happy to add any of the probe details or raw part-state dumps if useful.
5. NamedIdentity commented on Aug 27, 2026

   

   NamedIdentity

   on Aug 27, 2026

   More actions

   #45482 addresses this — a caller's Task call is now held until the children it started have settled, and an idle parent is woken so a late result is actually observed.
6. neriousy commented on Oct 1, 2026

   

   neriousy

   on Oct 1, 2026

   Member

   More actions

   In V2, when a subagent session finishes, its result is posted to the parent session as a <subagent ...> message with the final state, and an idle parent is resumed by default so it can act on the result.

   Wired up in #34320; delivery is shared since #46054.

   Closing as completed.
7. neriousy

   closed this as completedon Oct 1, 2026
8. Premik

   added a commit that references this issue on Oct 4, 2026

   fix(runner): queue work when already running instead of discarding

   ...

   4e78ca3