review — audit + (on follow-up) walk paydown

Two modes. First turn: print the audit and stop. On a user follow-up ("fix the top one," "walk these," "do A," "pay some down"), apply the rubric below.

First turn: print the audit

Run the bundled review.py (it lives in this skill's scripts/ directory — reference it with the relative path; Codex resolves it against the skill root):

python3 scripts/review.py

Optional: --top N to surface more than the default 3 candidates.

Re-emit the helper's stdout verbatim in a fenced code block. Codex may collapse long bash outputs — if you don't print it yourself, the user might not see it. Copy exactly: no preamble, no summary, no "want me to fix the top one?" The fenced block preserves column alignment.

Then stop. The user picks the next move.

Paydown mode (only on user follow-up)

Work through requested entries one at a time. Confirm before each fix. Never auto-batch. Never auto-commit.

For each entry, read the registry file, the hotspot, and adjacent tests. Apply this rubric:

• Already fixed? If the marker/symptom the entry describes no longer appears in the hotspot file, say so and add the entry's letter to the drop list. Don't re-fix.
• Cold area? Churn=0 since created: and age >90d → propose deferring. ~20% of files generate ~80% of debt-related rework; don't pay down vanity refactors.
• Prudent-deliberate with payoff_trigger not met? Honor the trigger. Skip with a one-line "trigger not met: <quote>."
• Fix candidate? Propose the smallest change that resolves the entry. Improvement, not perfection — don't refactor surrounding code.

When you fix

• Read the repo first. Check the test framework, adjacent tests, the cached feedback commands. Adapt to what exists; don't impose a new style.
• TDD where tests exist. Write a failing test that pins the deferral, then make it pass. Don't weaken or delete existing tests to make a fix pass.
• No tests in this area? Surface that and ask: write one, or fix without?
• Explain why this resolves the entry. Cite the entry's payoff_trigger or body — don't commit code you can't explain.
• Risky fix? Auth, payments, migrations, public APIs, or ai_authored: true → run a fresh-context review of the diff before suggesting commit. Fresh-context review catches what the writer's motivated reasoning misses.
• Don't commit. Show the diff. The user runs the gates, drops the entry with drop A, and commits.

Pacing

Aim for 3–10 entries per session — continuous paydown outperforms stop-the-world batches. If the user says "do them all," push back once: unsupervised AI cleanup measurably increases duplicate blocks and short-term churn. If they insist, still one-at-a-time with diffs surfaced.

Speak plainly

The frontmatter uses a research taxonomy (quadrant, category) for ranking and grounding — it is not user-facing vocabulary. When you talk about an entry, describe it in plain words; never say "prudent-inadvertent", "reckless-deliberate", "code_rot", etc. to the user. Use the entry's body and a plain phrase (e.g. "a planned tradeoff", "a shortcut you knew about", "came up later") instead. The review.py output is already translated — match its tone.

Don't

• Don't ask the user to confirm before running review.py.
• Don't paraphrase the helper's stdout. Copy it verbatim into the fenced code block.
• Don't enter paydown mode on the first turn. Stop after the report. Wait for the user's intent.
• Don't auto-commit. Ever.