pr-loop

A self-assessed cousin of greploop. Where greploop drives an external bot's score by pushing every round, pr-loop computes the score locally with the make-pr rubric — so it can iterate fast and only touch the network to publish each improvement.

The score is self-assessed, so the anti-gaming guards are load-bearing:

• The score comes from make-pr's executed gates (real pytest/tsc/eslint runs) — the loop cannot raise the number by editing prose; it must make the checks actually pass.
• Stage 2 is weakest-link — one weak axis caps the score, so the loop fixes the real problem.
• Stopping short of 5/5 is honest by design (see exit conditions) — the loop reports "stuck at 4/5 because X" rather than inflating to 5/5 to terminate.

Inputs

• PR number (optional): default to the PR for the current branch.

Procedure

1. Identify the PR

gh pr view --json number,headRefName,baseRefName -q '{number,headRefName,baseRefName}'

Switch to the PR branch if not already on it. Abort with a clear message if there is no PR for the branch (suggest running make-pr to create one first).

2. Loop — max 5 iterations

A. Score

Invoke the make-pr skill to compute the current Confidence Score and the binding axis, applying make-pr/references/rubric.md. This runs the repo's gates (tests/build/typecheck/lint) and evaluates the five judgment axes. Record score and binding_axis for this iteration.

Do not re-derive the rubric here — make-pr owns it. If make-pr is not installed, stop and tell the user to install it.

B. Check exit conditions — stop if any hold

1. Success: score == 5. → Refresh the PR body (step E) and exit reporting success.
2. Cap: this was iteration 5. → Stop, report current state + remaining blockers.
3. Plateau: score did not improve for 2 consecutive iterations. → Stop, report what it's stuck on.
4. Needs human: the binding axis is a judgment fork the loop must not settle alone — typically Safety/blast-radius or Scope discipline that would require a design decision, a risky deletion, reverting someone's intentional change, or an irreducible risk that simply cannot be a 5. → Stop and surface the decision: "stuck at N/5 on <axis> because X; raising it needs you to decide Y." Do not guess, and do not inflate the score to escape.

C. Fix the binding axis

Make a real change that addresses the lowest axis — not the prose:

• Tests adequacy → add meaningful tests for the uncovered behaviour-critical path.
• Correctness → fix the bug / handle the edge case the axis flagged.
• PR hygiene → update the doc/decision-record; clean up commit hygiene.
• Build/typecheck/lint gate red → fix the failing check until it is genuinely green.

Keep each fix scoped to the binding issue. If fixing it would itself require a judgment fork, treat that as exit condition B.4 instead of guessing.

D. Commit and push

git add -A
git commit -m "<short imperative message describing the actual fix>"
git push

Write a descriptive, imperative commit message about the real change (e.g. add citation dedup test for multi-page sources), not a generic "loop iteration N". Honor the repo's commit conventions from its CLAUDE.md/AGENTS.md — including no AI-authorship trailer where the repo forbids it.

E. Refresh the PR body

Re-run make-pr in update mode so the PR body's Confidence Score, files table, and diagram reflect the new state. Then return to step A.

3. Report

Summarize like greploop:

Output format

pr-loop complete.
  PR:            #123
  Iterations:    2
  Final score:   5/5
  Fixes:         add citation-dedup test; handle empty-retrieval edge case
  Stop reason:   reached 5/5

If it stopped short:

pr-loop stopped (needs-human).
  PR:            #123
  Iterations:    3
  Final score:   4/5
  Binding axis:  Safety/blast-radius
  Why:           schema migration is not reversible without data loss
  To raise it:   decide whether a backfill + down-migration is in scope for this PR