Arness Batch Merge

Smart PR management dashboard with guided per-PR review. Discover open batch PRs, delegate cross-PR analysis, then drive an interactive review/merge/refresh cycle.

This skill is a sequencer. It MUST NOT duplicate sub-skill logic. Arness-code-batch-merge handles: PR discovery, cross-PR analysis delegation, interactive review loop, merge execution, Feature Tracker updates, and chaining. Review logic lives in arn-code-review-pr; analysis logic lives in the arn-code-batch-pr-analyzer agent.

Pipeline position:

arn-code-batch-implement -> **arn-code-batch-merge** -> arn-code-batch-simplify
                               |
                               +-- per PR: review-pr -> approve/merge -> refresh

Step 0: Ensure Configuration

Read ${CLAUDE_PLUGIN_ROOT}/skills/arn-code-ensure-config/references/step-0-fast-path.md and follow its instructions. This guarantees a user profile exists and ## Arness is configured with Arness Code fields before proceeding.

After configuration is ensured, extract the following from ## Arness:

• Plans directory -- base path where project plans and CHANGE_RECORD.json files are stored
• Platform -- github, bitbucket, or none
• Vision directory (if exists) -- path containing the feature backlog
• Code patterns (if exists) -- path to code patterns directory

Step 1: Discover Batch PRs (Lightweight)

Platform Gate

If Platform is none: "Batch merge requires a platform (GitHub or Bitbucket) for PR management. Merge branches manually." STOP.

If Platform is github: verify gh CLI is available. If not: "The GitHub CLI (gh) is required for batch merge. Install it from https://cli.github.com/ and authenticate with gh auth login." STOP.

If Platform is bitbucket: verify bkt CLI is available. If not: "The Bitbucket CLI (bkt) is required for batch merge." STOP.

Scan for Batch PRs

Scan the Plans directory for CHANGE_RECORD.json files. Check all plan subdirectories: */CHANGE_RECORD.json, STANDARD_*/CHANGE_RECORD.json, SWIFT_*/CHANGE_RECORD.json.

For each CHANGE_RECORD.json found, check:

1. commitHash is populated (non-empty string) -- the feature was committed
2. nextSteps array contains at least one entry matching a PR URL pattern (contains github.com or pull or bitbucket or matches a PR reference like #123)

For each qualifying CHANGE_RECORD, extract:

• Feature name -- from projectName
• PR URL -- from the nextSteps entry containing the PR reference
• CHANGE_RECORD path -- full path to the CHANGE_RECORD.json file

Verify PR State

For each discovered PR, verify it is still open:

• Platform github: gh pr view <url> --json state

• Platform bitbucket: bkt pr view <id>

• If the PR is OPEN: include it in the batch

• If the PR is CLOSED or MERGED: skip it silently

If zero open batch PRs found: "No open batch PRs found. Run /arn-code-batch-implement first to create implementation PRs." STOP.

Collect the final list: PR number, URL, feature name, CHANGE_RECORD path, platform. This is a small data set -- it stays in context.

Step 2: Cross-PR Analysis (Delegated)

Spawn the arn-code-batch-pr-analyzer agent via the Task tool in foreground (the results are needed before proceeding), passing the model from .arness/agent-models/code.md as the model parameter (see plugins/arn-code/skills/arn-code-ensure-config/references/ensure-config.md "Dispatch convention" for fallback).

Pass to the agent:

• PR list (number, URL, feature name, CHANGE_RECORD path)
• Code patterns path (from ## Arness)
• Platform (github or bitbucket)
• Conflict-classification reference path: ${CLAUDE_PLUGIN_ROOT}/skills/arn-code-batch-merge/references/conflict-classification.md

The agent returns a concise summary containing:

• Status table -- each PR with CI status, review status, merge readiness
• Cross-PR findings -- shared files, overlapping concerns, integration risks
• Conflict map -- pairwise conflict classifications (clean / auto-mergeable / manual-resolution)
• Recommended merge order -- optimal sequence based on conflicts, dependencies, and readiness

Present the agent's summary to the user.

If the agent fails: Present the raw PR list without analysis. Offer to retry: "Cross-PR analysis failed. You can still review PRs individually, or I can retry the analysis."

Step 3: Guided Per-PR Review Loop

Ask (using AskUserQuestion):

"Ready to start reviewing PRs?"

Options:

1. Start with recommended PR -- begin reviewing the first in suggested order
2. Choose a PR to review -- pick a specific PR
3. Run batch-simplify first -- cross-feature quality pass before reviewing
4. Done for now -- exit

If Run batch-simplify first: Invoke Skill: arn-code:arn-code-batch-simplify. After it completes, return to this menu.

If Done for now: "Run /arn-code-batch-merge when ready to resume." STOP.

Selecting a PR

If Start with recommended PR: select the first PR from the recommended merge order.

If Choose a PR: present open PRs as numbered options. If more than 4 PRs, use layered navigation (first ask which group, then which PR within the group -- max 4 options per layer).

Review

Invoke Skill: arn-code:arn-code-review-pr with the selected PR number. The review-pr skill handles the full flow: fetch comments, validate, categorize, fix, local testing.

Post-Review Action

After review-pr completes:

Ask (using AskUserQuestion):

"PR #N reviewed. What next?"

Options:

1. Approve and merge -- merge this PR, update local, refresh analysis
2. Next PR -- move to next in recommended order
3. Re-check big picture -- re-run analyzer to refresh status
4. Done for now -- exit

Approve and Merge

Execute the merge:

• Platform github: gh pr merge <url> --merge
• Platform bitbucket: bkt pr merge <id>

Then update the local branch:

git checkout main && git pull

Clean up the feature's worktree (batch-implement worktrees only)

After the merge succeeds and main is up to date, remove the worktree and local branch created by arn-code-batch-implement. Only apply this cleanup when the merged PR's head branch starts with arn-batch/ — other branches (manual work, hotfixes) are left alone.

HEAD_BRANCH="$(gh pr view <url> --json headRefName -q .headRefName)"   # or equivalent for bkt
case "$HEAD_BRANCH" in
  arn-batch/*)
    SLUG="${HEAD_BRANCH#arn-batch/}"
    REPO="$(git rev-parse --show-toplevel)"
    WORKTREE_PATH="$REPO/.claude/worktrees/arn-batch-$SLUG"
    git -C "$REPO" worktree remove --force "$WORKTREE_PATH" 2>/dev/null || true
    git -C "$REPO" branch -D "$HEAD_BRANCH" 2>/dev/null || true
    ;;
esac

The --force flag is safe here: the branch's commits are already on main, so nothing is lost. Errors from both commands are ignored (the worktree or branch may have been cleaned up manually). Log the cleanup outcome alongside the PR URL in the merge summary: e.g. ✓ Merged #42 · worktree cleaned.

If the merge fails (CI checks, branch protection, conflicts): report the error: "Merge failed for [feature] (#PR): [error message]." Offer to skip and continue with another PR. Do NOT clean up the worktree — it still holds the only local copy of the work.

If remaining open PRs > 0:

• Re-run the arn-code-batch-pr-analyzer agent (foreground) with the updated PR list
• Present the updated summary
• If any remaining PR is now CONFLICTING that was not before: highlight the change
• Return to the post-review action menu for the next PR

Next PR

Select the next PR in the recommended merge order (skip any already merged). Invoke review-pr for that PR. After review, return to the post-review action menu.

Re-check Big Picture

Re-run the arn-code-batch-pr-analyzer agent (foreground) with the current open PR list. Present the updated summary. Return to th