Repository completion proof
Requires the rjmurillo/ai-agents checkout and its contributor-only sibling skills.
Inputs: PR number, intended acceptance criteria, and authorized task scope.
Apply the common execution contract.
Reuse the review config, pr-comment-responder, and the existing completion dispatcher.
This shortcut joins execution trust, criterion results, and platform acceptance.
Triggers
prove an ai-agents PR is completeverify before dispatching review commandsreview is finished but completion is uncertain
Process
Phase 1: Prove execution trust before dispatching review commands
Read config resolution in pr-review.
Use that documentation without invoking its standalone orchestration workflow.
Use its command map for that host rather than inventing a launcher.
Resolve the selected config's absolute path through that maintained resolution procedure.
For each invocation, bind PR_NUMBER and PR_REVIEW_CONFIG from the current task.
Do not rely on variables surviving separate tool shells.
From the repository root, the POSIX checkout form is:
GITHUB_SCRIPTS="${COPILOT_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-.claude}}/skills/github/scripts"
: "${PR_NUMBER:?Set the task PR number in this invocation}"
: "${PR_REVIEW_CONFIG:?Set the selected absolute config path in this invocation}"
uv run python "$GITHUB_SCRIPTS/pr/run_completion_gate.py" --config "$PR_REVIEW_CONFIG" --pull-request "$PR_NUMBER" --verify-config-only --json
Keep the default trusted remote-tracking ref unless the selected base requires another.
Any alternate anchor must resolve under refs/remotes/.
Do not substitute HEAD or a movable local branch to silence divergence.
Check anchor freshness against the selected remote revision through supported reads.
Fetching changes local Git state and requires the task's authority.
| Observation | Branch |
|---|---|
Successful verification and verified: true |
Begin verified finding handling. No completion claim follows yet. |
| Untrusted config or command file | Preserve the surfaced difference for human inspection. |
| Verification impossible | Restore verifiable Git evidence through an authorized owner; do not dispatch. |
| Installed config accepted by origin | Retain origin and command-trust evidence; run criteria before claiming success. |
Path containment does not establish execution trust.
A plugin root containing the consumer checkout is not disjoint.
The dispatcher trust contract
checks the origin, remote anchor, and tracked command files separately.
Use --approve-untrusted-config only after explicit human inspection and approval.
Never use it for an unverifiable Git state.
Phase 2: Execute completion criteria after verified review work
First read the current merged state through the verified test_pr_merged.py producer.
Use the selected config's launcher with --output-format json.
Inspect its JSON merged field, rather than treating exit zero as unmerged.
When merged is true, inspect the merged artifact and the requested downstream result, then stop Phase 2.
Do not invoke pr-comment-responder or the open-PR completion gate on that branch.
Continue only with Phase 3.
When the merged-state read is unavailable, stop Phase 2 and report the decision unresolved.
Do not dispatch open-PR criteria while merged is unknown.
When merged is false, continue with the open-PR steps below.
Pass trust evidence and the selected config to pr-comment-responder.
For actual findings, let reviewer-findings verify claims before responder changes.
Reconcile outcomes per comment ID through pr-comment-responder.
An acknowledged thread or local reviewer verdict is an intermediate result.
After the final push or resolution, run the selected config's wait_for_settled_zero.
Require three consecutive complete-and-zero readings before trusting thread completion.
Use its configured interval; a one-shot completion gate does not settle bot scans.
Skip settling only when no live bot reviewer can reopen threads.
Record that condition explicitly, rather than assuming a current zero is stable.
See the settling contract.
Then execute the same dispatcher without verify-only:
Bind the task PR and the same selected absolute config in this invocation.
GITHUB_SCRIPTS="${COPILOT_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-.claude}}/skills/github/scripts"
: "${PR_NUMBER:?Set the task PR number in this invocation}"
: "${PR_REVIEW_CONFIG:?Set the selected absolute config path in this invocation}"
uv run python "$GITHUB_SCRIPTS/pr/run_completion_gate.py" --config "$PR_REVIEW_CONFIG" --pull-request "$PR_NUMBER" --json
This command dispatches configured verifiers; inspect their effects before execution. Do not run it merely to test this library under a read-only task. Use an existing authorized evidence destination if persistence is required.
Require successful execution, all_passed: true, and every configured criterion's result.
Inspect fail-open criteria and dispatch errors before claiming measured completion.
Check pagination completeness, suppressed-finding dispositions, and reachable resolution SHAs.
The executable criteria
currently require zero unresolved threads, including security threads.
Its nearby exemption comment does not override that predicate or platform policy.
Preserve unresolved security work for the responsible owner; do not force-resolve it.
For failures, route only the failing criterion back to its producer. Use repository-pr-recovery when CI or policy attribution is uncertain. Use existing dispositions procedures for deliberately unaddressed findings. Do not manufacture approval markers or weaken predicates to make the gate green. Run focused verification after edits, then the required full gate at the final head.
Phase 3: Distinguish readiness, acceptance, and downstream completion
Join the criterion result with the effective-policy proof from repository-pr-recovery.
Verify required checks on the revision and event that GitHub evaluates.
Re-read readiness after a new commit, base update, or policy change.
If the PR already merged, inspect the merged artifact instead of failing its old readiness goal.
The open-PR completion config deliberately requires merged == false.
For authorized shipping, hand off to ship. Observe the merged state separately when the user requested a merged change. For releases or consumer behavior, hand off to repository-consumer-proof. Keep operator-owned publication or adoption work visible until independently verified.
Report three separate results: execution trusted, criteria passed, requested outcome observed. Name any inaccessible policy read and its responsible operator. A successful verify-only invocation proves only the first result.
Verification
- Trust evidence identifies the config, command files, and remote anchor.
- Executed criteria and platform evidence prove the requested outcome separately.
Anti-Patterns
Follow the shared outcome boundary.
Evolution
Follow the shared maintenance policy.
Scripts
Each question below has a runnable instrument. The dispatcher lists them and runs one by id:
python3 "${COPILOT_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-.claude}}/skills/repository-completion-proof/scripts/instruments.py" --repo . list
python3 "${COPILOT_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-.claude}}/skills/repository-completion-proof/scripts/instruments.py" --repo . run <id> -- <args>
Ids: verify-config (Phase 1), merged-state, settled-zero, completion-gate, resolution-shas, suppressed-findings (Phase 2), merge-ready, live-state (Phase 3).
It answers which script computes each fact and passes the child exit code through.
It does not choose the PR, the config, or what to claim. Those stay in the phases above.
No script resolves the config path. The maintained procedure in pr-review owns it.
Whether live-state and merge-ready join the check event to the head revision is not proven. Confirm that in Phase 3.
Provenance and Maintenance
reverify.json holds the checks that tie this skill to the dispatcher and config it names. Run the drift check after any edit:
python3 "${COPILOT_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-.claude}}/skills/skill-bootstrap/scripts/bootstrap_drift.py" check --repo . --skills-root .claude/skills --name repository-completion-proof --require