Agent Team terminal runs become unrecoverable: fallback rejects readiness failure, stop retains reservations, and headless PATH differs
The problem
Environment - 1DevTool: v1.50.0 - Platform: macOS 26.5.2 - Hermes Agent: 19.0 - Claude Code: 2.1.215 - Codex: 0.144.6 - Project: local Git repository - Orchestration mode: Agent Team - Substrates tested: terminal and headless Summary Agent Team runs can enter an unrecoverable lifecycle state. A terminal member may be marked as waiting for submission confirmation, but the documented headless fallback rejects that run because it is not considered a “proven pre-submission fallback.” Stopping the Team then fails to confirm process shutdown. Member reservations and worktrees remain retained, while the Team stays stuck in stopping and affected members become uncertain. A separate test also found environment inconsistency between terminal and headless substrates. The headless process could not find rtk, although the executable was available in the caller environment at /opt/homebrew/bin/rtk. This appears to involve three connected orchestration problems: 1. Readiness state and fallback eligibility are inconsistent. 2. Terminal completion/process shutdown is not reconciled reliably. 3. Headless processes do not preserve the required caller PATH. Steps to reproduce 1. Create a Git-backed project in 1DevTool. 2. Start an Agent Team with named members for Hermes, Claude, and Codex. 3. Assign: - Hermes: planning - Claude: implementation - Codex: testing/review 4. Use terminal substrate or auto, allowing 1DevTool to create visible agent terminals. 5. Wait for the Hermes member to initialize. 6. Inspect the Team: 1devtool-agent-v2 team status --team=<teamId> 7. In the affected run, the Hermes member enters: state=readiness-test activity="Submit confirmation needed" needsAttention=true attentionKind=submit 8. Collect the run: 1devtool-agent-v2 collect --run=<runId> --timeout=600 9. The collector eventually returns without a final result: state=pending stillRunning=true 10. Send a follow-up through the same Team member. 11. The follow-up is created but also remains pending because the member is still stuck before submission. 12. Attempt the documented headless fallback: 1devtool-agent-v2 fallback --run=<runId> --action=headless 13. The command fails with: Only a proven pre-submission fallback can use this action 14. Attempt to stop the Team: 1devtool-agent-v2 stop --team=<teamId> 15. The command fails with: One or more delegate processes did not confirm shutdown; their reservations were retained 16. Inspect the Team again. 17. The Team remains: state=stopping 18. Affected members become: state=uncertain activity="Process close not confirmed" error="Resources remain reserved until process death is confirmed" Observed behavior Readiness and fallback mismatch team status explicitly reports: - state=readiness-test - activity="Submit confirmation needed" - needsAttention=true - attentionKind=submit This indicates that the prompt has not been submitted to the agent successfully. However, fallback --action=headless rejects both the original run and its retry with: Only a proven pre-submission fallback can use this action The state exposed by team status and the state used by the fallback eligibility check appear inconsistent. Terminal result-correlation failure A separate Codex terminal Team was started to investigate the issue. The Codex OS process remained alive and appeared to be working, but: - collect reached its 600-second timeout - the run still reported running - the member remained engaged - no correlated final result was returned This suggests that the terminal
Desired outcome
Imported from UserJot — outcome was not captured separately.
Roadmap progress
PENDING
—
REVIEWING
—
PLANNED
—
IN PROGRESS
—
COMPLETED
—
Discussion0
Newest first- No comments yet. Add context or ask a question below.
Sign in to comment, vote and follow this request.
Sign in to reply