Repository navigation
Description
What happened
A delegated child's follow-up task can report completed and wake the parent while the child still has commands explicitly awaited with t3_background_wait. Different follow-ups can repeat this before the command finishes. The parent spends another model turn only to say it is still waiting or repeat known information.
The earlier report also records waiting children under Lineage's Previous agents: #21837 (comment)
Expected: task completion, task_status and Lineage agree with awaited work; distinguish interim progress from completion. This report does NOT claim an observed lost final result.
Diagnosis
Source: v0.1.10-nightly.20261111.4544, commit 1264c20.
Orchestrator.ts:742-753,8728-8809 creates qualifying notification follow-ups with completionWake=always. At 13177-13217,13469-13510 follow-ups finalize from their tagged runs without the original task's equivalent awaited-work gate. OrchestratorMcpService.ts:1491-1505 similarly derives follow-up readiness from task runs. threadRelationships.ts:98-117 prioritizes activityRunStatus; ThreadRelationshipsControl.tsx:548-575 groups completed/idle as Previous agents.
Steps to reproduce
Derived steps; NOT rerun as a clean repro:
- delegate_task to a Claude child; reply READY and finish the original task.
- t3_thread_send a follow-up: start a background command, mark it with t3_background_wait, then end the turn.
- Check whether the follow-up completes and wakes the parent while that command remains awaited.
- On its background-exit wake, start and await a second longer command, then end the turn. Check completion notifications and Lineage again.
Version
0.1.10-nightly.20261111.4544 (current server/client and inspected source). The 24-hour sample spans an update; pre-update server build is not fully established. The concrete Nov 12 example is after current server startup.
Environment
macOS Desktop -> Linux managed t3 serve service; Linux x64 7.0.0-43-generic; Node 26.12.1; claudeAgent, Claude Opus 5.5, Claude Code 2.1.322.
Evidence
Read-only sample: Nov 11 06:16 to Nov 12 06:27 UTC; 1,853 requests, 494,347,430 context tokens. Explicit intervals start at successful background_wait matched and end at the SAME command's terminal event.
Summarized log example (not verbatim), child TX command bripwel7g:
Nov 12 02:58:42.650 UTC: wait matched
03:05:32.476: parent completed follow-up
03:18:27.700: another completed follow-up
03:25:15.971: command terminal event
Both parent replies say the slice-3 gate is still waiting. Combined cache read: 283,912 tokens. Different follow-up IDs, not same-ID redelivery.
Overlapping subsets:
166 prompts / 206 requests: same notifying child still awaited; 52,937,454 context tokens (10.71%).
117 prompts / 117 requests: no tools, waiting/progress replies; 31,002,277 (6.27%).
34 manually reviewed waiting-only/known-information replies: 7,539,745 (1.53%): 7,505,950 cache read + 33,659 cache write + 136 uncached input; output 6,801 separately.
Not all early turns are unnecessary. These are attributed costs, NOT exact billing savings; cache reads are discounted.
Requests mapped by request reference/API start/prompt UUID, not ambiguous CSV wake markers; 14 rows without retained response blocks excluded from no-tool classification. Parent task_status calls: 0. No repeated task ID among 594 results in 476 completion messages. Lineage was not rechecked in this investigation; no result-loss/redo claim.Related issues
#21837 is the same underlying case; this adds awaited-interval and token-cost evidence and can be merged there. #32327 (open) covers exact repeats/redelivery, not general awaited-work holds. #33539/#34958 (open) concern nested delegates; #34958 excludes command roster attribution. #34568 merged and shipped in nightly 4564, not inspected 4544; no post-fix repro claimed.
Fix applied or workaround
None. No SQLite opened, settings/threads modified, source patched or service restarted. No update performed; recheck on a build containing #34568 is needed.
Filed by
Prepared by GPT-6.1 Sol through the Pi coding agent in T3 Code, on behalf of the reporter.
Activity
juliusmarminge commentedon Nov 12, 2026
This is the same problem as #21837, so I'm closing it in favour of that one. If it is different, say how and it will be reopened.