Skip to content

Evidence for #766: duplicate code agent dispatch on conforma/user-guide#239 — one run cancelled after 30 seconds #3876

Description

@fullsend-ai-retro

What happened

Two code agent workflow runs were dispatched for issue #239 within 38 seconds: run 29010929584 (created 10:14:14, cancelled at 10:15:03 after 30 seconds of sandbox creation) and run 29010964798 (created 10:14:52, succeeded). The triage run 29010690405 completed at 10:13:55. A second triage run (29011167811, for a different issue) started at 10:18:32. The cancelled code run consumed compute resources for sandbox setup but no LLM token cost.

What could go better

A concurrency guard or dedup check at dispatch time would have prevented the second code run from starting. The compute waste was minor (30 seconds of sandbox creation), but in higher-volume repos this pattern could compound. Medium confidence this was a dispatch race rather than a deliberate retry, since both runs were for the same issue and the cancelled run had no meaningful progress.

Proposed change

Add this as evidence for issue #766 (dispatch-layer dedup and per-issue concurrency guard). The conforma org is newly onboarded and already exhibiting the duplicate dispatch pattern, confirming the issue is not org-specific. The fix in #766 should include a check at dispatch time that no other code agent run is in-progress for the same issue number.

Validation criteria

After #766 is implemented, verify that only one code agent run is dispatched per issue per triage event. The cancelled-run-followed-by-successful-run pattern should not recur.


Generated by retro agent from conforma/user-guide#241

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    Status
    Done

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions