Finished agents re-emit next_step nine times after subagent-stop #529

New issue

Open

opened 2026-09-03 20:22:43 +00:00 by jared · 0 comments

jared commented

2026-09-03 20:22:43 +00:00

Owner

Context

On 2026-09-03, in the same run as implementation 4 of ticket 524, two agent dispatches kept emitting next_step after they had already stopped.

Observed

test-writer (dispatch 27) and programmer (dispatch 29) transcripts each show nine identical assistant turns emitting next_step after the handoff was already recorded.

test-writer dispatch 27: nine next_step turns, 13:12:50 to 13:13:16 UTC
programmer dispatch 29: nine next_step turns, 13:18:34 to 13:18:56 UTC

Each loop costs about 25 seconds and a full dispatch's worth of tokens, repeated nine times. Suspect: the settlement handshake (ADR-0144, os-sdlc-runner subagent-stop) re-prompts an agent that has already stopped.

Reproduce

Inspect transcripts under ~/.claude/projects/-home-jared-dev-cc-os/98124b14-084a-4ade-a034-daad3945a395/subagents/
for dispatch 27 (test-writer) and dispatch 29 (programmer); find the re-prompt source between the recorded handoff and the ninth next_step turn.

Expected

Origin

Trigger: test-writer dispatch 27 and programmer dispatch 29 during implementation 4 of ticket 524
Improvised this session: none
Chain: nine next_step emissions per dispatch ← settlement handshake re-prompts a stopped agent ← DESIGN (ADR-0144, os-sdlc-runner subagent-stop)
Root candidate: this ticket
Where: os-sdlc-runner subagent-stop handling (ADR-0144)
Session: ae931982-f255-4800-b06a-0d845d6b1df7
Transcript: ~/.claude/projects/-home-jared-dev-cc-os/98124b14-084a-4ade-a034-daad3945a395/subagents/

## Context On 2026-09-03, in the same run as implementation 4 of ticket 524, two agent dispatches kept emitting next_step after they had already stopped. ## Observed test-writer (dispatch 27) and programmer (dispatch 29) transcripts each show nine identical assistant turns emitting next_step after the handoff was already recorded. ``` test-writer dispatch 27: nine next_step turns, 13:12:50 to 13:13:16 UTC programmer dispatch 29: nine next_step turns, 13:18:34 to 13:18:56 UTC ``` Each loop costs about 25 seconds and a full dispatch's worth of tokens, repeated nine times. Suspect: the settlement handshake (ADR-0144, os-sdlc-runner subagent-stop) re-prompts an agent that has already stopped. ## Reproduce ``` Inspect transcripts under ~/.claude/projects/-home-jared-dev-cc-os/98124b14-084a-4ade-a034-daad3945a395/subagents/ for dispatch 27 (test-writer) and dispatch 29 (programmer); find the re-prompt source between the recorded handoff and the ninth next_step turn. ``` ## Expected <one prose sentence: after subagent-stop records a dispatch's handoff, the agent must not be re-prompted, so a dispatch emits next_step exactly once.> ## Origin - Trigger: test-writer dispatch 27 and programmer dispatch 29 during implementation 4 of ticket 524 - Improvised this session: none - Chain: nine next_step emissions per dispatch ← settlement handshake re-prompts a stopped agent ← DESIGN (ADR-0144, os-sdlc-runner subagent-stop) - Root candidate: this ticket - Where: os-sdlc-runner subagent-stop handling (ADR-0144) - Session: ae931982-f255-4800-b06a-0d845d6b1df7 - Transcript: ~/.claude/projects/-home-jared-dev-cc-os/98124b14-084a-4ade-a034-daad3945a395/subagents/

jared added the

plugin/os-sdlc

filed-by/agent

labels

2026-09-03 20:22:43 +00:00

jared referenced this issue from a commit

2026-09-03 20:24:17 +00:00

feat(os-sdlc): implement-speed fixes — richer briefs, rubocop server, stall detection