agent-workflow
Implement and review repository changes within local policy.
SKILL.mdplugins/vouchington-workflow/skills/agent-workflow/SKILL.md
GitHubAgent workflow
Read applicable AGENTS.md and their declared adapter alias, otherwise
.agents/skills/agent-workflow/SKILL.md from the repository root if present.
Local policy overrides these defaults; do not reload an already-read adapter or canonical skill.
Use before implementation. Inspect checkout status without discarding work; follow local branch,
worktree, contribution, and CI policy. Read every applicable AGENTS.md from the repository root
through changed files, plus relevant documentation and tests. Local instructions govern this skill.
Use PR descriptions before creating or updating a pull request. Diagnose failed CI with CI-log review, then update confirmed harness gaps before handoff. Keep those records after checks pass; routine local successes stay in the evidence record.
- During implementation, use implementation.
- Before handoff, use evidence sweep.
- For review, use review.
- For surfaced feedback, use review response.
- Before any rebase or lease push, use four refs.
Load only the references required by the current phase. Keep scope changes explicit; a discovered blocker or follow-up does not authorize widening the task.
Close-out
This section is the single home for the hand-off retrospective and merge-note rules.
- Report actual validation, open pull requests, and remaining blockers.
- Hand-off retrospective. When you hand a pull request to the human as ready, write one retrospective from the session that did the work, which reads that session's transcripts and journal. A native stack gets one retrospective when the stack is handed off, covering every layer; never one per layer or for an intermediate commit.
- Give each hand-off its own
sourceEventId, such ashandoff-pr<number>(a stack uses its top pull request's number), on bothretrospective_checkandretrospective_save. The default ID fits only the first hand-off in a session; a second would be mistaken for it. - Merge note. If you are still running when the merge happens and it changes the outcome, such
as a merge-queue ejection and its fix, a follow-up push, or review comments addressed after
hand-off, append one short note with the blackboard skill. Name the
pull request and the retrospective's source event, and give the note its own
sourceEventId(for examplemerge-pr<number>): the writer rejects the retrospective's ID reused for different content. Never write a second retrospective.
Parallel by default
Split work into independent units and run them concurrently, one worker and worktree each, within any cap the human set. Dispatching to a peer session in another repository is a unit.
Serialize only on a named dependency:
- A consumer change needs an upstream API or fix that is not yet released.
- Units change the same files or contract. Stack them instead of waiting.
- Units would write in the same worktree.
For upstream work, open every independent pull request at once, shepherd them concurrently, merge the ready set together, and cut one release. Start each consumer adoption once an admitted release contains what it needs; start consumer work that does not need the unreleased change now. A ready, queued, or shepherding pull request does not pause other units.
Consumer wrapper or local instructions own: default branch, runner class, documentation root, review system, merge policy, and command catalog.