Forgejo: implementer failure and reset #157

Merged

lucas merged 1 commit from agent/issue-147-implementer-abandon into main

2026-10-30 13:06:12 +01:00

omp-deck commented

2026-10-29 15:21:34 +01:00

Collaborator

Closes: #147

Implementer sessions can now end as failed. Both the agent and a human can trigger it.

Agent: the new forgejo_abandon_implementation tool (implementer policy only, and only for the session's own issue) posts ## Implementation abandoned plus the summary on the issue. When the call succeeds, the bridge queues an implementation_abandoned row and wakes the dispatcher. The wake goes through a new bridge.onForgejoTrigger hook, which index.ts sets.
Human: closing the agent's PR without merging no longer archives the session. It queues an implementation_scrapped turn (continue-only, no permission check) whose seed tells the agent to call the tool. After that turn ends, whatever the outcome, the dispatcher queues the same abandon row with summaryPosted: false. The dedupe key leaves out that flag, so if the agent already called the tool, this second insert does nothing.
Reset (handleAbandon): waits for the session to go idle, then closes every open PR mapped to the session (new DispatchApi.closePull, PATCH on the pulls endpoint; branches are kept). It removes progress_label and adds the new optional implementation.planning_label. If the agent wrote no summary, it posts the fixed notice. Any problems are reported on the issue, or as a notification if posting fails. Finally it archives the session. The issue stays open.
If the agent closes the PR itself (the abandon handler does this), that close is dropped without archiving.
planning_label must differ from progress_label. It may equal intake_label.

Deviations from the plan:

enqueueAbandon is in a new module, forgejo/abandon.ts, not in dispatcher.ts, so the bridge doesn't import the dispatcher.
swapInProgressLabels is now a generic swapLabels(add, remove). The start path and the abandon path both use it.
If .omp/forgejo.yaml can't be read or parsed during the reset, the labels are left alone and the problem is reported on the issue.

Known behavior, unchanged: the issue stays mapped to the archived session until a new implementer starts. So a comment on the issue before that resumes the abandoned session, the same way comments on closed artifacts do today.

Verification: bun test apps/server/src/forgejo apps/server/src/bridge (858 pass). The new dispatcher tests cover the agent abandon, no planning label, a human scrap with the fixed notice, dedupe after the agent's summary, the agent's own close being dropped, label-failure reporting, and a restart on the issue afterwards. Server typecheck is clean apart from an existing system-prompt.test.ts error that this PR doesn't touch.

Closes: #147 Implementer sessions can now end as failed. Both the agent and a human can trigger it. - **Agent**: the new `forgejo_abandon_implementation` tool (implementer policy only, and only for the session's own issue) posts `## Implementation abandoned` plus the summary on the issue. When the call succeeds, the bridge queues an `implementation_abandoned` row and wakes the dispatcher. The wake goes through a new `bridge.onForgejoTrigger` hook, which `index.ts` sets. - **Human**: closing the agent's PR without merging no longer archives the session. It queues an `implementation_scrapped` turn (continue-only, no permission check) whose seed tells the agent to call the tool. After that turn ends, whatever the outcome, the dispatcher queues the same abandon row with `summaryPosted: false`. The dedupe key leaves out that flag, so if the agent already called the tool, this second insert does nothing. - **Reset** (`handleAbandon`): waits for the session to go idle, then closes every open PR mapped to the session (new `DispatchApi.closePull`, PATCH on the pulls endpoint; branches are kept). It removes `progress_label` and adds the new optional `implementation.planning_label`. If the agent wrote no summary, it posts the fixed notice. Any problems are reported on the issue, or as a notification if posting fails. Finally it archives the session. The issue stays open. - If the agent closes the PR itself (the abandon handler does this), that close is dropped without archiving. - `planning_label` must differ from `progress_label`. It may equal `intake_label`. Deviations from the plan: - `enqueueAbandon` is in a new module, `forgejo/abandon.ts`, not in `dispatcher.ts`, so the bridge doesn't import the dispatcher. - `swapInProgressLabels` is now a generic `swapLabels(add, remove)`. The start path and the abandon path both use it. - If `.omp/forgejo.yaml` can't be read or parsed during the reset, the labels are left alone and the problem is reported on the issue. Known behavior, unchanged: the issue stays mapped to the archived session until a new implementer starts. So a comment on the issue before that resumes the abandoned session, the same way comments on closed artifacts do today. Verification: `bun test apps/server/src/forgejo apps/server/src/bridge` (858 pass). The new dispatcher tests cover the agent abandon, no planning label, a human scrap with the fixed notice, dedupe after the agent's summary, the agent's own close being dropped, label-failure reporting, and a restart on the issue afterwards. Server typecheck is clean apart from an existing `system-prompt.test.ts` error that this PR doesn't touch.

omp-deck force-pushed agent/issue-147-implementer-abandon from efaa7dc1b1 to 7933859ec0

2026-10-30 13:05:18 +01:00

Compare