merge

The user invoked /procoder:merge with arguments: $ARGUMENTS (The argument is the PR number or branch; with none, use the PR for the current branch: gh pr view --json number.)

First read .procoder/github/WORKFLOW.md and follow it — the repo's rules win over the defaults below. If it is missing, get the default via procoder templates, write it, then follow it.

A PR merges when EVERYTHING answers — never before.

1. Don't block on the waiting: spawn a BACKGROUND agent whose only job is to watch and report. The watcher never fixes, never replies, never merges — it gathers information; you act on it. Continue other work while it watches, and pick up the loop below each time it reports. Give the watcher the merge-watching protocol from .procoder/github/WORKFLOW.md verbatim: calibrate against previous runs of the same workflow, poll PER JOB in the foreground (gh run view <run-id> --json jobs — never a fire-and-forget monitor), report the FIRST failing job immediately with its --log-failed excerpt instead of waiting for the rest, poll dynamically (fast early and near the calibrated finish, never slower than 90s), and report on every state change.

Loop until done:

1. Checks: gh pr checks <pr> — every required check must pass. For a failing check, open its run (gh run view <id> --log-failed), fix the cause in the code, commit, push, and wait for the re-run.

2. Reviews: gh pr view <pr> --json reviews,reviewDecision and the review comments via gh api repos/{owner}/{repo}/pulls/<pr>/comments. Treat BOT reviewers (Copilot included) exactly like humans: read every comment and recommendation, and for each one either

   • fix it (commit, push, then reply to the thread saying what changed), or
   • reply with a concrete reason why not, or
   • when it is real, non-blocking, and too large for this PR, TRACK IT BEFORE MERGING: procoder todo add <title>, or a backlog story where it belongs. Write the printed template to its named path, fill its criteria, and name the originating PR. Then reply with that id. "Out of scope for this PR" is only true if something now holds it; said on its own it is how a real finding is answered and then lost. No comment is skipped silently. Resolve threads only after fixing or answering (gh api graphql resolveReviewThread), never to hide them.

   Tracking never waives a failing check, a correctness or security blocker, or a required review change. Fix those before merging. Step 2b still applies now to every real finding, including tracked follow-up work; if the detecting layer cannot be adapted safely in this PR, stop.

   How to receive review feedback — bot or human:

   • VERIFY before implementing: check each claim against the actual code. Reviewers (Copilot especially) lack full context; a finding can be factually wrong for this codebase. Wrong → push back in the thread with the technical reason, not a fix.
   • If ANY comment in a review is unclear, ask about it before implementing the others — comments are often related, and partial understanding produces the wrong fix.
   • Fix one finding at a time and re-run the covering tests after each.
   • Replies state facts: "Fixed in — ." No "great point", no thanks, no "you're absolutely right" — gratitude is not a verification, and performative agreement erodes the reviewer's trust in real agreements.
   • A reviewer asking for an unused "proper" feature gets the YAGNI answer: show it is uncalled and propose removal instead.

2b. The reflection step — MANDATORY whenever step 2 found anything real (bot or human). An escaped finding is a bug in our gates; fixing only the finding leaves its class open:

• For each real finding, name the layer that should have caught it BEFORE the PR existed: a linter (lint baseline/config), the pre-PR rubric (.procoder/github/REVIEW.md), a quality controller, a pinning test, or CI itself.
• ADAPT that layer now, in this same PR: enable the linter rule, add the rubric line, tighten the controller, write the pinning test.
• Record each as an entry in .procoder/github/LESSONS.md (shape: procoder templates), with the adaptation named. Then run procoder lessons — it flags any entry left unlearned, and an unlearned lesson is not done.
• False positives get no ledger entry — reflect only on real escapes.
• Then run procoder copilot-leak. Copilot's auto-review sometimes opens ISSUES rather than PR comments, and those escape the reading in step 2 entirely: the finding gets fixed and nothing records that the class happened. The command finds them, strips every trace of the code from them, and asks before it files or records anything. This is the only place a procoder command talks to GitHub about reviews — the gate reads the resulting ledger offline and never queries.