…ter timeout/disconnect (#2356) (#2492) @castrojo reported fresh duplicate contributor PRs AFTER his hive was upgraded to the image containing #2404 and #2435/#2455 (bluefin-lts#494 + #495, "avoid rechunker ordering cycle" — the SAME fix 17 min apart; common #919 + #921, same fix a third time). #2404 only split gh READ from WRITE in bin/gh-wrapper.sh so an agent CAN run `gh pr list` before starting — it is advisory, with no enforced gate, and two near-simultaneous sessions both see "no open PR" before either has pushed. #2435's failure cooldown covers the 30-min max-duration timeout (it routes through task_failed → recordTaskFailure), but NOT a disconnect-release. Root cause: the contribute-ws disconnect defer (pkg/dashboard/contribute_ws.go) nil-ed currentTask and logged "task released on disconnect" but recorded NO cooldown. A WS drop therefore removes the issue from activeIssues — the only double-assign guard — while recording nothing, so selectTask can hand the SAME issue to another session during the reconnect window (BASE_RECONNECT_DELAY_MS..MAX_RECONNECT_DELAY_MS = 1s–60s). Meanwhile the original relay keeps currentTask locally and re-asserts it via task_progress on reconnect (bin/contributor-relay.sh), so both sessions keep working the issue and each opens its own PR. This matches the live [contribute-ws] log for bluefin-lts#466: 16:00 RELEASED on disconnect (castrojo) → 16:00 ASSIGNED (castrojo) immediately → 16:04 COMPLETE. Fix: mirror the task_failed path and book the SHORT non-permanent failure cooldown (failedTaskCooldownMinutes) for the abandoned issue on disconnect, so it is not instantly re-admissible. The short window (10m) comfortably outlasts the reconnect backoff, so the returning session re-asserts and repopulates activeIssues before the cooldown lapses; a completion later resets the failure ledger. Only real issue tasks are booked (Number > 0) so synthetic pr-review tasks never poison an issue key. Regression tests: a disconnect-release now parks the abandoned issue in a cooldown and the next selection skips it in favour of another issue; a Number==0 synthetic review task is never parked. Fixes #2356 🤖 Generated with [Claude Code](https://claude.com/claude-code) Signed-off-by: Andrew Anderson <andy@clubanderson.com>

Uh oh!

/contribute sending "dupe" PRs #2356

Description

Activity

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions