git:20261010.474217f to git:20261012.364f696

---

name: continue-pr

argument-hint: "[PR number]"

metadata:

internal: true

---

# Continue a Bot PR

Finish the repair on its existing branch, using the live PR as the work record.

Read its diagnosis, attempted fixes, review findings, and current checks before

deciding what remains. A response or review can be complete while the repair is

still unfinished.

## Required skills

- Load `/tend-ci-runner:run-tend` and its repo-specific overlay first.

- Load `/tend-ci-runner:fix-a-bug` before fixing, `/tend-ci-runner:push-commits`

before pushing or withdrawing the PR, `/tend-ci-runner:monitor-ci` to verify

the head, and `/tend-ci-runner:merge-pr` for its landing decision.

- Load `/tend-ci-runner:resolve-conflicts` for conflicts and

`/tend-ci-runner:post-to-github` before updating the thread.

## Establish ownership

```bash

BOT_LOGIN=$(gh api user --jq .login)

gh pr view <number> --json author,state,isDraft,headRefOid,headRefName,headRepository,baseRefName,commits,body,comments,reviews,statusCheckRollup

uv run --script \

"${CLAUDE_PLUGIN_ROOT}/scripts/active_subject_runs.py" \

"https://api.github.com/repos/$GITHUB_REPOSITORY/pulls/<number>"

```

in `/tend-ci-runner:merge-pr`; defer to an independent dedicated owner. A human

takeover or maintainer instruction to stop leaves the work with them. Read any

related in-flight PR before treating it as the owner of this repair: an

investigation or partial fix can leave work here.

## Complete the remaining work

Read inline review comments too (`gh api

"repos/$GITHUB_REPOSITORY/pulls/<number>/comments" --paginate`). Reproduce the

remaining failure on the live head and continue diagnosis from the prior

evidence. Apply a safe fix, verify it, and push to this PR's branch. Use an

isolated `$TMPDIR` worktree when the caller is also handling other subjects;

retain the original checkout for its remaining work.

A previous failed attempt or uncertain diagnosis calls for further diagnostic

work; it does not by itself transfer the repair to a maintainer. Stop for a

current blocker you cannot clear, such as inaccessible evidence or tooling, an

external dependency, or a semantic decision with no defensible default. State

what is blocked and what would unblock it. Re-check a recorded blocker against

current state before relying on it.

When prior attempts failed to clear the same failure, another push needs new

evidence or a materially different approach. Continue diagnosis; if neither

emerges, record what evidence is missing instead of repeating the attempted

fix. Each push starts another review round.

Re-check head and state before expensive verification and before pushing per

**Re-check the head SHA before the expensive verify, not just before the push**

in `/tend-ci-runner:push-commits`. Preserve any sibling's commits. A review that

arrives while this run works remains with its reviewer per

**Reviews arriving during verification** in `/tend-ci-runner:push-commits`.

Once the remaining findings are addressed, poll the pinned head per

`/tend-ci-runner:monitor-ci`, including a head this run did not push. Handle

failures as remaining repair work, then complete the landing decision per

is disproved, withdraw only the bot's PR after the branch-state recheck in

`/tend-ci-runner:push-commits`.

Update the PR's diagnosis and verification when the work changes them. An

unchanged blocker needs no repeat comment. A deduplicated response or review

suppresses that post, not the remaining repair work.