Start one durable Agent Skill execution Step for an existing RepoPilot run. Call this immediately before executing the named Skill, then pass the returned Step ID to repopilot_finish_step. Reusing the same idempotency key within a run returns the existing Step instead of creating a duplicate.

Finish a running RepoPilot Step with an explicit terminal outcome and evidence-backed summary. Use only after the associated Skill has stopped; do not use it to update a Step that is still running. Repeating the same status and summary is idempotent, while a conflicting terminal result is rejected.

Append one immutable, SHA-256 hash-chained evidence record to an existing RepoPilot run. Use after each material decision, tool result, Git reference, CI result, or error; never include credentials or raw secrets. This is append-only and repeated calls create distinct evidence records.

Create a human approval request for one high-risk action on a maintenance run. Use before merge, branch deletion, rollback, permission changes, secret changes, or another explicitly gated action. This records the request and may pause the run, but it never executes the requested action.

Search verified historical RepoPilot runbooks for a target repository using failure symptoms or repair terms. Call during triage before selecting a repair strategy; do not use it as proof that a current defect is identical. Returns ranked stored runbooks and does not modify repository or run state.

Archive a verified maintenance outcome as a reusable repository runbook. Call only after independent verification for the source run; do not archive an unverified hypothesis or a pull-request review run. This inserts a new runbook and appends runbook evidence.

Read the current title, body, labels, state, and URL of one GitHub issue in an allowed repository. Use this to establish source requirements before triage; do not use it for pull requests. This performs a read-only GitHub API request and requires GITHUB_TOKEN.

Read current metadata for the pull request bound to a RepoPilot review run, including its immutable head SHA and change totals. Use before reviewing files or publishing a review. The run, repository, and pull request must match, and the call fails if the GitHub head changed.

Read one 100-file page of changed-file metadata and GitHub-provided text patches for the pull request bound to a review run. Use consecutive page values until hasNextPage is false. The tool verifies the head SHA before and after reading and returns null patch values for binary or unavailable diffs.

Create a GitHub pull request from an existing repair branch for a maintenance run. Use only after the patch and focused verification are complete; this is the highest automatic repository write allowed by pull_request_only. It creates an external PR and appends a git_reference evidence record, but never merges it.

Read GitHub combined commit statuses and Actions check runs for a pull request, then append the result as ci_result evidence. Use during verification or read-only review after a PR exists. For review runs, the PR number and current head SHA must still match the run.

Build a redacted Proof Bundle from a terminal maintenance run, verify its evidence chain, and create or update the run-scoped managed GitHub PR comment. Use only for the PR recorded in that run's create_pull_request evidence. Repeating the call updates the same marked comment instead of posting duplicates.

Create or update one managed GitHub review summary comment for the exact pull request head SHA bound to an active read-only review run. Call after reading all relevant diff pages and checks, while the Reviewer Step is running. This publishes findings but never approves, requests changes, modifies code, or merges.

Squash-merge a GitHub pull request only after a human approved this exact merge for the same maintenance run. Do not call for pull-request review runs or without the matching approval ID and version. The approval is atomically consumed before the external merge, so it cannot authorize another action.