pr-recheck

PR recheck

Follow-up pass on a PR that has already been reviewed once. Assumes a prior round of inline comments exists and the author has pushed fixes since.

Preconditions

Same as pr-classify: run gh auth status; require a PR number/URL or a checked-out branch with an open PR; verify open state via gh pr view <id> --json state,mergedAt. If merged or closed, surface that and stop.

Steps

1. Load prior threads. Fetch open review threads via GraphQL — REST does not expose thread resolution state:

   gh api graphql -f query='query($owner:String!,$repo:String!,$num:Int!){
     repository(owner:$owner,name:$repo){ pullRequest(number:$num){
       reviewThreads(first:100){ nodes{ id isResolved
         comments(first:20){ nodes{ path line body author{login} } } } } } } }' \
     -f owner=OWNER -f repo=REPO -F num=N
   

   Keep only threads with isResolved: false.
2. Load new work. Diff and commits since the previous review (gh pr diff <id> plus the commit log). Read the new code fully, cross-referencing touched files.
3. Build the verification matrix. For each open thread, classify one of: addressed / partial / not addressed / wont-fix (author replied), each with the file:line and the commit that addresses it. Output as a table.
4. Scan for new findings. Re-reading the changed code may surface fresh issues. Classify them Critical / Important / Optional per pr-classify rules — the anti-nitpick gate applies (drop Optional whose cost of ignoring is "none").

Delegation

Delegate the gather, keep the verification judgment. Hand steps 1-2 (load the open threads, load the new work and read the changed code) to cheaper-tier subagents, each returning compact evidence: the unresolved threads, and what changed since the last review with file:line. Keep on the main model the addressed / partial / not-addressed matrix (step 3), the new-findings classification (step 4), and the decision gate below. Resolving threads, posting comments, and approving are gated actions — never delegated to a subagent.