Menu
Autonomous issue-to-PR workflow from next eligible issue to ready PR with consistent metadata handling.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Baseado na classificação ocupacional SOC
Open a conservative Robot SF PR with scope verification, freshness checks, and artifact discipline.
Use for an autonomous Robot SF issue-to-PR loop that selects eligible GitHub issues, implements one scoped issue at a time, validates, pushes, and opens PRs.
Guarded PR merger; merges merge-ready PRs after verifying label, CI status, branch protection, and preflight checks.
Use for an autonomous Robot SF PR review loop that fixes scoped review gaps, validates proof, resolves review threads, and applies merge-ready; not for merging.
Continuous goal autopilot; orchestrates implement, review, merge, and discover cycles with preflight validation and delegation failure recovery.
Clarify ambiguous GitHub issues by tightening scope and acceptance criteria, proposing solution options with pros/cons, and marking decision-required issues when maintainer input is needed.
| name | gh-issue-autopilot |
| description | Autonomous issue-to-PR workflow from next eligible issue to ready PR with consistent metadata handling. |
| category | github-issue |
| kind | orchestrator |
| phase | implementation |
| requires_write | true |
| requires_slurm | false |
| requires_benchmark_artifacts | false |
| delegates_to | ["gh-issue-sequencer","implementation-verification","pr-ready-check","gh-pr-opener","artifact-provenance"] |
| output_schema | issue_to_pr_summary.v1 |
| aliases | ["issue-to-pr","gh-issue-to-pr"] |
Use this when the user asks to "take the next issue" and execute through to a ready PR.
Primary integration is with repo project state and PR opening; details of branch validation and detailed issue creation are handled by child skills.
ll7/robot_sf_ll7ll7 Project #5mainBASE_REF=origin/main scripts/dev/pr_ready_check.shdocs/context/issue_713_batch_first_issue_workflow.mdChoose the next issue in order:
ReadyTodoTrackedWithin the same status, use gh-issue-sequencer as the source of truth for Project #5 ordering.
Tie-breakers after sequencing: no blocker labels, no linked PR, older open issue first, stronger evidence first.
Before accepting a ready candidate, inspect the issue body and recent comments for source_pr,
linked PRs, Closes/Refs, and explicit prerequisite PRs. Verify the referenced implementation
surface exists on current origin/main when the branch must start from clean main. If an open PR
already covers the issue, classify it as covered_by_pr, route it to state:running, and stop
instead of reimplementing. If the issue depends on an unmerged source PR, mark it blocked or
unavailable for clean-main work with the unblock condition "source PR merged to origin/main",
unless the user explicitly chooses a stacked-PR route.
Before auto-admitting a ready candidate, revalidate it against current origin/main history using
the issue's named symbol, failing test, and error signature — not issue-number references alone.
Search recently merged PR titles and bodies, and when a fix is named only by issue number, map its
merge commits back to the covering PR through the commit-to-pulls API so a fix filed under a
different issue number is still found. Then verify the candidate's failure signature, named symbol,
and failing file/line against current origin/main history and code. Classify the issue as
covered_by_pr and stop before claim or branch only when an exact merged fix implements the stated
boundary and its regression proof covers the reported failure. Record the covering PR rather than
treating a loose keyword, a change in the same file, or historical failure output alone as duplicate
evidence.
Two regression fixtures are required:
PosixPath serialization: stale because current main already replaced
json.dumps(asdict(arm_params)) with the tested subprocess-boundary serializer._run_batch_sequential 3-tuple return: #5480 was dispatched after PR #5486
merged under #5482 (not #5480), so an issue-number-only merged-PR search missed it; the covering
PR is found only by mapping the named failing test
test_run_batch_sequential_worker_failure_logs_warning and its
too many values to unpack (expected 2) signature to current origin/main, and the expected
outcome is covered/stale with PR #5486 linked and no implementation worker admitted.Delegated queue-scout output is only route evidence until verified by the main agent in the target repository. Before using a scout recommendation to select, claim, or branch for an issue, run:
uv run python scripts/dev/gh_issue_rest.py thread <number> --repo ll7/robot_sf_ll7
This tries the concise gh issue view --comments path and falls back to paginated REST only for the
known repository.issue.projectCards GraphQL failure. Confirm the URL belongs to
ll7/robot_sf_ll7, the issue is still open, ready labels/state are current, recent comments do not
block or supersede the work, and no linked/open PR already covers it. Treat stale state, wrong
repo-owner URLs, omitted recent comments, and duplicate PR coverage as expected scout failure modes.
Run preflight for required local tooling and the Scout publication linter:
uv run python scripts/dev/check_skills.py --preflight gh-issue-autopilot.
Then refresh credentials and branch baseline (gh auth status, git fetch origin).
Resolve the queue candidate and re-check the complete issue thread with
uv run python scripts/dev/gh_issue_rest.py thread <number> --repo ll7/robot_sf_ll7 before claim
or branch, especially when a delegated scout proposed the issue.
Check open PRs for duplicate coverage using the linked issue, head branch/scope, and title. Stop or update routing when an existing PR covers the work.
If issue statement is ambiguous:
decision-required (create if missing),Tracked,Acquire the cross-machine issue claim before any implementation branch or worktree setup:
uv run python scripts/dev/issue_claim.py acquire <issue-number>
The command atomically creates the remote ref agent-claims/issue-<issue-number> through
GitHub's create-ref API, which fails when the ref already exists. If it exits non-zero, another
PC or agent probably claimed the issue first; run
uv run python scripts/dev/issue_claim.py status <issue-number>, record the collision, and skip
to the next candidate. Do not reimplement the issue unless the remote claim is confirmed stale
and deliberately released.
After a successful claim, make the claim visible in GitHub issue/project state: move the issue to
In progress, add or preserve state:running when using state labels, assign the local actor
when practical, and add a short issue comment naming the claim ref, machine/thread, intended
implementation branch, and stale-claim cleanup condition.
Publication rule: use scripts/dev/gh_comment.sh, gh issue/pr comment --body-file, or REST
JSON input for Markdown-heavy comments. Do not put bodies containing backticks, YAML, commands,
or multiline Markdown into inline shell strings. Use REST endpoints for simple label writes when
gh would route through GraphQL, for example:
gh api repos/ll7/robot_sf_ll7/issues/<number>/labels -f labels[]=state:running
gh api -X DELETE repos/ll7/robot_sf_ll7/issues/<number>/labels/state:ready
Create the issue branch in a linked worktree for non-trivial implementation. Prefer the
AGENTS.md "Fresh Worktree Bootstrap" location, naming, machine-context symlink, and branch
freshness rules; use an in-place branch only for tiny or explicitly requested main-checkout work.
Implement inside accepted scope.
Run validation gate and rerun on failures only when fixable.
Commit with conventional message; if long-running benchmark evidence appears, classify artifacts.
Sync with latest origin/main, rerun readiness, and check artifact durability.
Re-check open PRs for the same linked issue, head scope, or title before opening a new PR.
Open a ready PR using gh-pr-opener; use draft only when explicitly requested or when the
PR body names a concrete reason review should be blocked.
After the PR exists and the issue is visibly covered by the PR, release the transient claim with
uv run python scripts/dev/issue_claim.py release <issue-number>. Do not release the claim
earlier unless the run is abandoning the issue and records the handoff.
For deferred important work, create follow-up issues and link them before final handoff.
agent-claims/issue-<number> as the cross-machine mutex for implementation selection. A
successful claim means this run may proceed; a failed claim means skip the issue, inspect status,
and choose another candidate. Project status, labels, assignments, and comments are visibility
surfaces, not the atomic claim.AGENTS.md "Worktree Teardown And Preservation" so tracked,
untracked, and ignored-but-important local changes are preserved or explicitly dismissed.Emit:
Use this skill for the scope named in its frontmatter description and registry metadata.
.agents/skills/skills.yaml.