Continue PR Auto

Continue Work on a Pull Request (Unattended)

Pick up an existing pull request, resolve conflicts, fix CI failures, address reviewer feedback, and push updates.

This is the unattended variant used by automation (utils/continue-all-prs.sh). It must never wait for user input: resolve conflicts and address feedback with your best judgment and push, and when a branch is too stale for a mechanical merge, rework the feature to reconcile it with the latest master (step 3a). The only hard stop is a genuinely missing PR number. For interactive use where asking is acceptable, use continue-pr instead.

Arguments

• $0 (required): PR number in the current repository (e.g., 12345)

Process

1. Parse arguments and fetch PR metadata

Extract the PR number from $0. This skill runs unattended and must never stop to ask the user anything: if the PR number is not provided, stop with a clear error message instead of prompting.

Validate that the PR number contains only digits. Reject any non-numeric input immediately — do not pass unvalidated input to shell commands or GraphQL queries.

Determine which repository you are operating in — the public ClickHouse/ClickHouse or the private ClickHouse/clickhouse-private:

gh repo view --json nameWithOwner --jq .nameWithOwner   # or: git remote get-url origin

Remember it as $REPO and use it everywhere below instead of the literal ClickHouse/ClickHouse in gh commands, API URLs, and GraphQL repository(owner:, name:) arguments. The repository also decides whom to ask about unrelated CI failures (step 4): @groeneai in ClickHouse/ClickHouse, @oranjeai in ClickHouse/clickhouse-private.

Fetch PR metadata using gh if available, otherwise use WebFetch on the GitHub API:

GH_USER=$(gh api user --jq .login)
gh pr view "$PR_NUMBER" --json number,title,body,headRefName,baseRefName,state,mergeable,mergeStateStatus,author,url,headRepository,headRepositoryOwner,isCrossRepository,maintainerCanModify,statusCheckRollup,reviews,comments,reviewRequests

If gh is not available or not authenticated, use WebFetch to get the data. Append ?per_page=100 and follow the Link header for pagination (the rel="next" URL) to fetch all pages:

• https://api.github.com/repos/ClickHouse/ClickHouse/pulls/$PR_NUMBER
• https://api.github.com/repos/ClickHouse/ClickHouse/pulls/$PR_NUMBER/reviews?per_page=100
• https://api.github.com/repos/ClickHouse/ClickHouse/pulls/$PR_NUMBER/comments?per_page=100

Report the PR title, author, branch, and current state to the user.

1a. Verify "already integrated" claims before closing

Treat closing a PR as already integrated or obsolete as a high-confidence decision. Verify both of these independently against freshly fetched refs:

1. Historical ancestry: Resolve and print the exact base and PR-head OIDs, then check the candidate commit in this direction only:

   HEAD_REMOTE=origin
   if [ "$IS_CROSS_REPOSITORY" = "true" ]; then
       HEAD_REMOTE="pr-$AUTHOR_LOGIN"
       git remote add "$HEAD_REMOTE" "$FORK_URL" 2>/dev/null || git remote set-url "$HEAD_REMOTE" "$FORK_URL"
   fi
   git fetch origin "$BASE_BRANCH"
   git fetch "$HEAD_REMOTE" "$HEAD_BRANCH"
   git rev-parse "origin/$BASE_BRANCH" "$HEAD_REMOTE/$HEAD_BRANCH"
   git merge-base --is-ancestor "$CANDIDATE_COMMIT" "origin/$BASE_BRANCH"

   Derive $IS_CROSS_REPOSITORY, $AUTHOR_LOGIN, and $FORK_URL from the PR metadata as in step 2. Exit status 0 means the base contains the commit; any other status means it does not. Never check against HEAD, the PR branch, --all, or a worktree and describe that result as containment by the base branch. If this check is nonzero, do not claim the commit is in the base and do not close the PR as integrated.
2. Current effective state: An ancestor commit may have been reverted or backed out later. Ancestry, git branch --contains, git cherry, and patch-equivalence results are historical evidence only; none proves that the change remains effective. Inspect all later base-branch commits touching the changed paths, search for explicit and manual reverts/backouts, compare the current base tree with the PR's intended effect, and run the reproducer or regression test against the current base when feasible. If the change was fully or partially reverted, count it as not integrated. If the current effect cannot be established confidently, leave the PR open for human review.

Before posting a closure comment, state the fetched base OID and the evidence for both ancestry and the current effective state. Do not close from a remembered ref, a stale local branch, a matching subject, or a commit merely visible somewhere in the repository.

2. Check out the PR branch locally

Determine whether the PR branch is in the main repository or in the author's fork.

If the branch is in the main repository (ClickHouse/ClickHouse):

HEAD_REMOTE=origin
git fetch "$HEAD_REMOTE" "$HEAD_BRANCH"
git checkout --detach "$HEAD_REMOTE/$HEAD_BRANCH"

If the branch is in the author's fork:

Derive the fork clone URL from the PR metadata (headRepository.url or headRepository.owner.login + headRepository.name) rather than hardcoding the repository name (forks can be renamed). Use a pr- prefixed remote name to avoid colliding with existing remotes like origin:

REMOTE_NAME="pr-$AUTHOR_LOGIN"
FORK_URL="https://github.com/$FORK_OWNER/$FORK_REPO.git"  # from headRepository in PR metadata
git remote add "$REMOTE_NAME" "$FORK_URL" 2>/dev/null || git remote set-url "$REMOTE_NAME" "$FORK_URL"
HEAD_REMOTE="$REMOTE_NAME"
git fetch "$HEAD_REMOTE" "$HEAD_BRANCH"
git checkout --detach "$HEAD_REMOTE/$HEAD_BRANCH"

Immediately after either checkout completes, discard any state left by an earlier PR before recording the immutable baseline. The automation uses a dedicated worker worktree, so it must never preserve staged, unstaged, or untracked files from a previous session. This is the only authoritative record of the PR surface that existed when the worker began:

git fetch origin "$BASE_BRANCH"
# The automation may have already prepared a validated, conflict-free merge of
# the base branch into the pull-request head in this worktree (the triage phase
# of `utils/continue-all-prs.sh` records it). Keep that merge instead of
# throwing it away and paying for a second merge and rebuild. Accept the
# recorded commit only when it is exactly that merge: a two-parent commit whose
# first parent is the current remote pull-request head and whose second parent
# is the current base-branch head. Anything else - a stale marker from an
# earlier pull request, a commit with a different shape - is ignored.
VALIDATED_MERGE_FILE="$(pwd)/tmp/continue-all-prs/validated-base-merge"
CHECKOUT_TARGET="$HEAD_REMOTE/$HEAD_BRANCH"
if [ -s "$VALIDATED_MERGE_FILE" ]; then
    VALIDATED_MERGE=$(cat "$VALIDATED_MERGE_FILE")
    if git cat-file -e "${VALIDATED_MERGE}^{commit}" 2>/dev/null \
        && [ "$(git rev-parse "${VALIDATED_MERGE}^@")" = "$(git rev-parse "$HEAD_REMOTE/$HEAD_BRANCH" "origin/$BASE_BRANCH")" ]; then
        CHECKOUT_TARGET="$VALIDATED_MERGE"
    fi
fi
git reset --hard "$CHECKOUT_TARGET"
git clean -ffdx -e build/ -e tmp/continue-all-prs/
test -z "$(git status --porcelain)"
mkdir -p tmp
PR_BASELINE_DIR="$(pwd)/tmp/continue-pr-${PR_NUMBER}-baseline"
rm -rf "$PR_BASELINE_DIR"
mkdir -p "$PR_BASELINE_DIR"
INITIAL_PR_HEAD=$(git rev-parse "$HEAD_REMOTE/$HEAD_BRANCH")
# The pull-request head stays the baseline even when the worktree already
# carries the validated base-branch merge on top of it.
git merge-base --is-ancestor "$INITIAL_PR_HEAD" HEAD
INITIAL_BASE_HEAD=$(git rev-parse "origin/$BASE_BRANCH")
printf '%s\n%s\n' "$INITIAL_PR_HEAD" "$INITIAL_BASE_HEAD" > "$PR_BASELINE_DIR/state"
git diff --name-status "origin/$BASE_BRANCH"...HEAD > "$PR_BASELINE_DIR/name-status"
git diff --stat "origin/$BASE_BRANCH"...HEAD > "$PR_BASELINE_DIR/stat"
git diff --binary "origin/$BASE_BRANCH"...HEAD > "$PR_BASELINE_DIR/diff"

Do not modify, stage, or delete this artifact during the session. It must remain available through the push safety gate in step 7. The deterministic path and state file are required because each resumed worker turn starts a fresh shell process.

3. Resolve conflicts with the base branch (if any)

Use the PR's actual base branch from metadata (baseRefName) instead of hardcoding master — this ensures backport PRs targeting other branches are handled correctly.

Check if the PR has conflicts with the base branch:

git fetch origin "$BASE_BRANCH"
git merge-base --is-ancestor "origin/$BASE_BRANCH" HEAD || echo "needs merge"

If the branch is behind the base branch and is red (some checks didn't pass), or if it is behind the base branch for more than a week (regardless of checks success), or has conflicts (including when GitHub reports the PR as CONFLICTING or its mergeability as unknown), or if at least one CI failure is unrelated to this PR and its fix has already landed on the base branch (merging pulls the fix in and clears the red — see step 4), or if the merge queue removed the PR after its last commit (see step 4), merge:

git merge "origin/$BASE_BRANCH"

If there are merge conflicts, resolve them autonomously — this skill runs unattended, so never stop to ask:

1. List conflicted files: git diff --name-only --diff-filter=U
2. For each conflicted file, use a Task agent with subagent_type=general-purpose to resolve:
   • Read the conflicted file
   • Analyze conflict markers
   • Resolve intelligently: the PR's changes should generally take precedence for the code the PR modifies, while master's changes take precedence for unrelated areas
   • Stage the resolved file: git add <file>
   • For a genuinely ambiguous conflict, pick the resolution most consistent with the PR's intent and record that choice in the merge commit message. Do NOT use AskUserQuestion or otherwise wait for input.
3. Complete the merge: git commit --no-edit
4. Resolving conflicts is not finished until the merge is committed and pushed — push it (step 7).

3a. Reconcile a stale branch with the latest master

Resolving merge markers is not always enough. If the branch is long-stale, the mechanically merged result may not compile or fit current APIs — master may have renamed or removed functions, migrated settings to the pimpl pattern (DECLARE / (*settings)[Setting::X]), reworked IStorage / pipeline / QueryPlan, moved headers, etc. In that case, rework the PR's feature so it reconciles with the latest master — do not give up, defer, or leave it half-merged:

1. Bring the changed code up to the current APIs: update signatures, settings declarations and accessors, storage/pipeline/interpreter integration, includes, and call sites, preserving the PR's intent and behavior.
2. Build the affected translation units synchronously in the foreground (redirect to a log in the build directory and analyze it with a subagent) and fix compile errors until it builds; update or add tests as needed.
3. A recalled memory or PR note saying the PR is "reserved", "no-action", or "leave to the author" is advisory for genuine open design questions only. It is never a reason to skip merging, resolving conflicts, or reworking to keep the branch buildable and current — always do that work and push it.
4. Only when a full rework is genuinely infeasible in this environment — e.g. a submodule points at a fork that cannot be fetched or built here — do the most you can (merge, resolve conflicts, rework whatever you can build), push it, and state plainly what could not be verified and what remains blocked (e.g. the submodule needs a ClickHouse-org fork).

A CONFLICTING PR must not be left unresolved whenever you can push. Resolve the conflicts (steps above) and push:

• Determine pushability in this order:
  1. A branch in the main repository (isCrossRepository=false) is directly pushable through origin. Ignore maintainerCanModify for same-repository PRs; GitHub can report it as false because the field describes maintainer access to a fork, not access to a branch in the base repository.
  2. A fork owned by the authenticated $GH_USER (or a PR authored by $GH_USER) is directly pushable through that user's fork remote. Ignore maintainerCanModify for the authenticated user's own fork PRs; the owner can push regardless of whether base-repository maintainers are allowed to edit.
  3. Only for a cross-repository fork owned by someone else, use maintainerCanModify: true means push to the fork remote; false means the branch is not pushable by the authenticated user.
• For every pushable PR above, push the resolved branch — to origin for same-repository PRs and to the fork's remote for fork PRs (step 7). A contested, reserved, NA, dsgn, or "superseded" note does not block the mechanical conflict resolution and push; it only reserves the final design / merge decision. Resolving conflicts means keeping the author's intended change merge-clean against current master — it does not require the PR's design to be correct (that stays the human's call).
• If you cannot push — specifically, another author's cross-repository fork with maintainerCanModify=false, or a push actually fails for lack of permission — supersede the PR, provided the change is still wanted and is not obsolete, already fixed on master, already covered by another open PR, or design-rejected/contested. (In those excluded cases, report the state and leave the human decision; never open a duplicate of an existing superseding PR.) To supersede:
  1. Switch explicitly to supersede mode before creating the replacement target; do not reuse the original fork remote or branch name:

     PUSH_MODE=supersede
     PUSH_REMOTE=origin
     PUSH_BRANCH="continue-pr-${PR_NUMBER}-<short-desc>"
     SUPERSEDE_REASON="maintainer edits are disabled on the original fork"
     SUPERSEDE_REAUTHORED_UNSIGNED_CLA=false
     git ls-remote --exit-code "$PUSH_REMOTE" "refs/heads/$PUSH_BRANCH" && exit 1 || true

     The branch name must be new. Never update, delete, or force-push an existing replacement branch.
  2. Credit the original author. If they have signed the CLA, keep their original commits so their authorship is preserved in the history. If they have not signed it (a CLA note, or the CLA check is red on the original), create a new branch from the fetched base and re-create the intended change under your own authorship. Do not carry the original commits or a Co-authored-by: trailer, or the CLA check will block the new PR too:

     git checkout --detach "origin/$BASE_BRANCH"
     git switch -c "$PUSH_BRANCH"
     SUPERSEDE_REAUTHORED_UNSIGNED_CLA=true
     # Re-create and test the intended change, then commit it under the current user.

     Record the reason for using supersede mode and whether the replacement is a re-authored unsigned-CLA change before passing the step-7 gate.
  3. Push the replacement to the main repo (origin) under $PUSH_BRANCH, then open a new PR to the base branch with gh pr create, following .github/PULL_REQUEST_TEMPLATE.md. State that it supersedes the original and add Related: <original PR URL> (and Closes: <issue> if the original targeted one).
  4. Close the original with gh pr close <N> --repo "$REPO" --comment "...": say it is superseded by the new PR (link it), that you could not push the resolution here because maintainer edits are disabled on the fork, and thank the author. If the change is genuinely not worth superseding, resolve locally if useful and report the specific blocker (e.g. "resolved locally but the fork has maintainer edits disabled") — not a bare "needs attention".
• Check push access up front with gh api user --jq .login and gh pr view <n> --json author,isCrossRepository,maintainerCanModify,headRepositoryOwner,headRepository. Never infer that a same-repository PR or the authenticated user's own PR is unpushable from maintainerCanModify=false.

4. Analyze CI status and fix failures

First, check whether the merge queue rejected the PR. The PR's own CI runs on its head commit, but the merge queue re-tests the PR merged with the latest base branch (a gh-readonly-queue/<base>/pr-<N>-<base-sha> ref, workflow MergeQueueCI). A PR can therefore be all green and CLEAN yet keep being rejected - typically a semantic conflict: a test added on the base branch that pins behaviour this PR changes, or an API the PR uses that was changed on the base branch. Look for the latest merge-queue event:

gh api graphql -f query="
{
  repository(owner: \"${REPO%%/*}\", name: \"${REPO#*/}\") {
    pullRequest(number: $PR_NUMBER) {
      mergeQueueEntry { state }
      commits(last: 1) { nodes { commit { oid committedDate } } }
      timelineItems(last: 1, itemTypes: [ADDED_TO_MERGE_QUEUE_EVENT, REMOVED_FROM_MERGE_QUEUE_EVENT]) {
        nodes { __typename ... on RemovedFromMergeQueueEvent { createdAt reason beforeCommit { oid } } }
      }
    }
  }
}"

If the last event is a RemovedFromMergeQueueEvent (with a reason other than merged) newer than the last commit, and the PR is not back in the queue, the rejection is the first failure to fix. beforeCommit.oid is the merge-queue commit that was tested. fetch_ci_report.js with the PR URL detects this case and adds the failed jobs of that merge-queue run to its output (marked with a ⚠️ line and listed under MergeQueueCI). To look at the run directly, take its report links from the commit statuses (the check runs link to GitHub Actions jobs instead):

MQ_SHA=<beforeCommit.oid>
gh api "repos/$REPO/commits/$MQ_SHA/statuses" --paginate \
    --jq '.[] | select(.state != "success") | "\(.context)\t\(.state)\t\(.target_url)"'
node .claude/tools/fetch_ci_report.js "<a praktika.html?REF=gh-readonly-queue/...&sha=$MQ_SHA&name_0=MergeQueueCI&name_1=<job> URL from above>" --failed

The per-job JSON is also directly at https://s3.amazonaws.com/clickhouse-test-reports/REFs/gh-readonly-queue/<base>/pr-<N>-<base-sha>/$MQ_SHA/mergequeueci/result_<job>.json (gzip-compressed). Do not construct artifact URLs such as log paths by hand: Praktika stores attached files under normalized sub-result directories (e.g. .../mergequeueci/build_amd_binary/build_clickhouse/build_clickhouse.log), so guessed paths return AccessDenied. Follow the links of the failed results in that JSON instead, as printed by fetch_ci_report.js (also with --links), or use --download-logs. To fix the rejection, merge the latest base branch (step 3) even if the PR is otherwise current, reproduce the failure on the merged tree, and fix it in the PR - adapt the code to the changed API, or update a newly added test when the PR intentionally changes the behaviour it pins (say so in the commit message and the PR comment). Do not re-add the PR to the merge queue yourself, and do not dismiss a merge-queue failure as flaky without the same evidence as for any other failure.

Then fetch the PR's own CI reports:

node .claude/tools/fetch_ci_report.js "https://github.com/ClickHouse/ClickHouse/pull/$PR_NUMBER" --failed --cidb

For each CI failure:

1. Check if it is a known issue: Search for existing open issues or PRs that address this failure:

   gh issue list --repo ClickHouse/ClickHouse --state open --search "<failure_description>" --limit 5
   gh pr list --repo ClickHouse/ClickHouse --state open --search "<failure_description>" --limit 5

   Only dismiss a failure as unrelated if there is a concrete open issue or PR that matches. Do NOT dismiss failures without evidence.

   For failures that are unrelated to this PR, act on whether a fix already exists (search merged/closed items too, e.g. gh pr list --repo ClickHouse/ClickHouse --state merged --search "<failure_description>"):

   • If the fix has already merged into the base branch (the matching issue is closed by a merged PR, or the fixing commit is already on the base branch), update the branch: merge the base branch (step 3) so the fix is pulled in and the red clears on the next CI run, instead of merely dismissing it.
   • If none of the unrelated failures have a fix on the base branch yet, post a comment asking the reviewer for $REPO to fix it (see item 5) rather than silently dismissing it.

2. Investigate the failure: Download logs if needed:

   node .claude/tools/fetch_ci_report.js "<report_url>" --failed --download-logs

   Extract and analyze the relevant logs. Read the failing test files and the code they exercise.

3. Fix the failure: Make the necessary code or test changes. Each fix should be a separate commit with a clear message explaining what was wrong and why.

4. If the only failure is "CH Inc sync", fix it using the /fix-sync skill.

5. If you are confident that the failure is unrelated to the changes, post a comment, asking the reviewer for $REPO (step 1) — @groeneai in ClickHouse/ClickHouse, @oranjeai in ClickHouse/clickhouse-private — to investigate the failure: 🕵 @<reviewer>, investigate the failure: <link> and provide a fix in a separate PR. If the fix is already in progress, link it here.

6. Repeat until all failures are addressed or confirmed as known issues with links to open issues/PRs.