Use BEFORE claiming done, before writing a PR body, when receiving review feedback (especially unclear or technically questionable), when requesting review, or after completing a feature/bugfix. Enforces technical rigor over performative agreement, evidence-based completion claims, and a prumo-specific review checklist tuned to recurring incident classes (BOLA, run-state TOCTOU, error swallowing, schema drift, ApiResponse envelope drift, stale TanStack cache). Triggers on phrases "looks good", "should be fine", "I'll fix it", "ready to merge", "tests pass", "done", "complete", "fixed", "ready for review", "let me draft the PR". Be pushy about running it.
Use BEFORE claiming done, before writing a PR body, when receiving review feedback (especially unclear or technically questionable), when requesting review, or after completing a feature/bugfix. Enforces technical rigor over performative agreement, evidence-based completion claims, and a prumo-specific review checklist tuned to recurring incident classes (BOLA, run-state TOCTOU, error swallowing, schema drift, ApiResponse envelope drift, stale TanStack cache). Triggers on phrases "looks good", "should be fine", "I'll fix it", "ready to merge", "tests pass", "done", "complete", "fixed", "ready for review", "let me draft the PR". Be pushy about running it.
Code Review (prumo)
Three practices, one principle: evidence before claims.
Verification gate — run the actual commands before saying "done".
Requesting review — small PR, clear risk + test plan, label needs-review.
Receiving feedback — technical rigor over performative agreement.
Plus: a prumo-specific review checklist grounded in the bug classes that have actually shipped here. Every checklist item exists because we paid for it once.
Core principle
Technical correctness over social comfort. Verify before implementing. Ask before assuming. Evidence before claims.
Every claim in a review (yours or theirs) must link to one of:
A file:line reference (e.g. backend/app/api/v1/endpoints/extraction_runs.py:88).
A command output (test run, lint run, grep, git diff).
A doc reference (docs/reference/extraction-hitl-architecture.md §4.1).
"It should work" / "it looks right" / "probably fine" are not evidence. If you find yourself typing one of these phrases, stop and run something.
When this skill fires
Before any of these statements leave your mouth, run this skill:
"Ready to merge", "ready for review", "let me draft the PR".
"I'll just push this and see".
Receiving a review comment, especially when your instinct is to immediately agree or immediately disagree.
If the user explicitly asks for a code review, security review, or PR draft — this skill is mandatory.
The verification gate (the Iron Law)
No completion claim ships without fresh evidence.
IDENTIFY command → RUN command in full → READ output → CONFIRM it matches the claim → THEN claim
Skipping any step is lying, not reviewing.
Commands that count as evidence on prumo
Claim
Evidence (run from repo root unless noted)
Backend tests pass
make test-backend — exit 0, no FAILED / ERROR
Backend lint clean
make lint-backend — exit 0, no diff after format
Frontend tests pass
npm run test:run — exit 0
Frontend lint + typecheck
npm run lint + npm run typecheck
Migration applies cleanly
cd backend && alembic upgrade head — exit 0
Migration is reversible
alembic downgrade -1 && alembic upgrade head
Bug fixed
New test reproducing the bug now passes
Endpoint authorized
grep for ensure_project_member in the endpoint
No N+1
Run the request with SQL echo, count queries
Red flags — stop and run something
Using "should" / "probably" / "seems to" about your own code.
"I refactored this, should be equivalent" without a test re-run.
Self-congratulation before verification ("nice, that should do it").
Committing because the change "looks small".
Trusting an agent's report instead of reading its tool outputs.
Full protocol: ../debugging/verification-before-completion/SKILL.md (the canonical gate; references/verification-before-completion.md now redirects there).
The prumo review checklist
Apply to every diff before requesting review and before approving someone else's PR. Each item exists because we shipped this bug class before — see the linked git history.
A. Authorization (OWASP API #1 — BOLA)
Every endpoint touching project data calls ensure_project_member (or a stricter role check) before the data access. Pattern: see backend/app/api/v1/endpoints/extraction_runs.py:88.
When the endpoint takes run_id / article_id / template_id but notproject_id, resolve the project and run the membership check. Don't trust the URL.
Manager-only operations call the manager check, not just membership. Re-grep both layers.
Frontend never controls authorization — server is authoritative. If the client decides who can do what, write it down as a defect.
Why: BOLA is the #1 OWASP API risk and the #1 historical bug class on prumo. Audit playbook: references/bola-audit.md.
B. RLS + multi-tenancy
New tables have RLS enabled and a policy. Check backend/alembic/versions/0018_* for the helper pattern (is_project_reviewer).
Policies are written in terms of auth.uid() and project_memberships, not raw user IDs.
If the migration relaxes RLS, the PR body explains who gains access and why.
Why: Supabase RLS is our second line of defense. Migration 0018 relaxed reviewer writes for a reason — every RLS relaxation needs the same scrutiny. Full checklist: references/rls-review.md.
C. Run-state and concurrency (TOCTOU)
Run-state transitions go through run_lifecycle_service, not ad-hoc run.status = ... assignments.
State checks (if run.status == X) and the subsequent write happen in the same DB transaction, or use a conditional UPDATE ... WHERE status = X RETURNING ... so two requests can't both win.
HITL session opens / closes are idempotent — re-running the same request doesn't double-create rows.
Celery tasks that mutate run state re-fetch under lock; don't trust the state captured at enqueue time.
Why: Run-state TOCTOU has bitten us multiple times — see commit 1994ceb fix(backend): resolve 31 auto-found bugs across HITL/extraction stack. Race-spotting guide: references/race-conditions.md.
D. Error swallowing
No .catch(() => ({ success: true })) or any catch that returns a success-shaped object without context.
Promise.all is only used when every child must succeed. Otherwise use Promise.allSettled and surface partial failures explicitly.
No except Exception: pass or bare except:. Either re-raise, log + raise, or convert to a specific exception.
Empty query results aren't treated as success. if not rows: return ok() is almost always wrong on prumo — usually a missing membership check is silently hiding the data.
Why: Commit 5493631 fix(frontend): resolve 17 auto-found extraction hook + service bugs was almost entirely error-swallow fixes. Deep dive: references/error-swallowing.md.
E. Schema drift (SQLAlchemy ↔ Pydantic ↔ TypeScript)
Every Optional[X] in Pydantic matches a nullable=True in SQLAlchemy and vice versa.
Defaults are defined in one place: prefer DB-side defaults via Alembic, mirror in SQLAlchemy server_default, do not also default in Pydantic unless you mean "API will fill this in".
Enum values are in sync between Python enum, DB type, and frontend type. New variants require a migration.
Frontend types regenerated / updated when the response shape changes.
Why: Three of the 31 backend bugs in the auto-fix wave were schema drift. Reference: references/schema-drift.md.
F. ApiResponse envelope
Every API response uses the ApiResponse envelope ({ data, ... }) consistently. No endpoint returns the bare payload.
Frontend unwraps once. If a hook returns data.data.foo, that's a double-unwrap bug — see 7100956 fix(qa): drop double-unwrap of ApiResponse envelope in useRunAIExtraction.
Mutation responses are unwrapped at the same layer as queries — pick a layer (fetcher vs hook) and stick to it. Mixed layers = bugs.
Why: Envelope inconsistency caused at least four shipped bugs. Rules: references/api-envelope.md.
G. TanStack Query cache
Cache keys include every variable that scopes the data: at minimum project_id, usually run_id, often article_id. Missing scope = leaks between projects.
Mutations call queryClient.invalidateQueries for every list/detail key whose data they changed — not just the obvious one.
Optimistic updates have a rollback path on error. If onError is empty, the cache will lie after a failed mutation.
When the backend autoadvances a Run stage (PROPOSAL → REVIEW), the frontend invalidates the run detail key.
Why: Stale-cache bugs are silent — users see old data and assume their click failed. Cache-key playbook: references/tanstack-cache.md.
H. Migrations
Migration revision IDs match the file name. Down-revisions point to the actual previous head.
Destructive migrations (drop column / drop table) have a documented rollback or are gated behind a feature flag.
No data migration in the same revision as a schema migration that locks the table — split them.
Read docs/reference/migrations.md before touching backend/alembic/versions/.
I. Tests
New behavior has a failing test that now passes (TDD or test-with-fix is fine; "I'll add tests later" is not).
Bug fixes include a regression test asserting the original symptom is gone.
No flaky sleep-based tests — use deterministic fixtures or freeze_time.
Backend test names describe the scenario, not the code path. test_ensure_project_member_blocks_outsider > test_ensure_project_member_3.
J. Style + meta
All code, comments, commit messages, and PR body are in English (CLAUDE.md §1).
Commit message follows Conventional Commits (feat, fix, refactor, docs, test, chore, ci, perf) with an optional scope, e.g. fix(hitl): close session on abort. See git log on dev for the in-repo style.
PR targets dev, not main.
PR title is the same imperative as the headline commit. PR body has the structure below.
K. Decisions & docs land with the code
If this PR makes an architectural decision (new dependency, storage strategy, cross-cutting pattern, core abstraction), the ADR is in this PR under docs/adr/ — superseding, never editing, an existing one. Not a "docs follow-up".
A new/changed endpoint under backend/app/api/v1/endpoints/ or a new extraction_* table updates the matching docs/reference/ doc (and its last_reviewed) in this PR, or the PR body says "no doc change needed".
The PR's Definition of Done block (.github/PULL_REQUEST_TEMPLATE.md) is filled — evidence ticked, not hoped.
Why: ADRs and reference docs that lag the code drift out of sync — the glossary-sync fitness check only catches concept-vocabulary drift, not endpoint/architecture docs. Bind the decision to the diff that makes it.
Requesting review
Make the PR easy to review
Small. One concern per PR. If you can't summarize the change in one sentence, it's two PRs.
Targeted at dev. Never open against main.
Labeled. Add needs-review (and security if it touches auth / RLS / endpoint exposure).
Self-reviewed first. Run through the prumo checklist above before pinging anyone.
PR body template
## Why
<1–3 sentences. What problem, why now. Link issue if any.>
## What changed-<bulletperlogicalchange, groupedbyarea>-<linktofile:linewherethechangeisnon-obvious>## Risk<Whatcouldbreak? Whichareasareexercised? Migrations / RLS / state-machinechangesgetaparagraph.>## Test plan- [ ] `make lint-backend` clean
- [ ] `make test-backend` passes (note any newly added tests)
- [ ] `npm run lint` clean
- [ ] `npm run test:run` passes
- [ ] Manual: <stepsareviewercanreproduce>## Out of scope<Whatyounoticedbutdeliberatelydidnotfixhere, withlinktofollow-upissueiffiled.>
Why this template: reviewers need risk + reproduction, not a code rehash. The diff is the "what" — you owe them the "why" and the "how to verify".
After pushing
Watch CI. Do not request review on a red branch.
Resolve your own review comments on the diff (notes for the reviewer) before pinging.
If the PR is non-trivial, write a "review entry point" comment: which file to read first, which is mechanical.
Never skip VERIFY. The reviewer might be wrong. You owe them the courtesy of checking, not the performance of agreeing.
Banned phrases (performative agreement)
"You're absolutely right!"
"Great catch!" / "Great point!"
"Thanks for the feedback!"
"Good idea, I'll do that."
These are social lubricant masquerading as engagement. They commit you to changes you haven't evaluated. Replace with one of the four real responses below.
The four real responses
Restate the requirement.
"You're asking me to change ensure_project_member to a manager check on this endpoint because writes here should be manager-only — confirming."
Ask a clarifying question (when truly unclear, not as a stall).
"When you say 'wrap this in a transaction', do you mean the whole handler or just the two-statement state transition?"
Push back with technical reasoning.
"I checked backend/app/api/v1/endpoints/extraction_runs.py:88 — ensure_project_member is already called. Did you have a different endpoint in mind, or am I missing a path?"
Just start working — when the request is clear, correct, and small.
When to push back
Push back when the suggestion is:
Factually wrong (grep proves it, link the file:line).
More expensive than the problem (YAGNI — "we might need this someday" is not a reason now).
In conflict with docs/reference/extraction-hitl-architecture.md or docs/reference/migrations.md.
A style preference dressed as a correctness claim.
Push back politely with evidence, not opinion: "Here's what the code does today, here's why I picked this, what am I missing?"
Fix in this PR but in a separate commit: Important, scope-adjacent improvements.
Open a follow-up issue: Out-of-scope cleanups, "while we're here" suggestions that double the PR size.
Don't silently drop a comment. Either resolve it with a reply ("won't fix because X") or link the follow-up issue.
Full protocol: references/receiving-feedback.md.
AI-assisted review etiquette
If an AI agent (you or another) is producing review comments:
Every comment cites file:line or a command. No vague pattern-matching.
No "consider doing X" without explaining the concrete risk if X isn't done.
No mass-flagging style nits that the linter would catch. Make the linter catch them instead.
The human reviewer's time is the scarcest resource — prioritize Critical → Important → Minor, in that order, and stop at the first 10 items.
If you (the agent) are receiving AI-generated review on a PR, apply the same standard: demand file:line, demand the concrete risk. AI review without evidence is noise.
Any automated surface reviewing a pull request uses this same skill as
its single source of truth: the checklist above is WHAT to review;
references/automated-pr-review.md is the orchestration contract —
how to identify the PR, the dedup rule, the ## Claude review comment
format, and the comment-only hard rules. Keep review knowledge HERE,
never inlined in routine prompts or workflow files (inline copies rot;
the clone always carries the current version of this skill).
Workflow integration
TDD / feature-dev: review-after-each-task. Run the verification gate before saying a task is done.
Multi-task subagent runs: each subagent's "done" report is unverified — re-run the gate yourself before believing it.
Pre-merge: full prumo checklist + verification gate. CI green is necessary, not sufficient.
Hotfix on main: still goes via dev unless explicitly approved otherwise. Hotfixes get more scrutiny, not less.
Bottom line
Evidence before claims.file:line or command output, every time.
Small PRs, full self-review, clear PR body.
Technical rigor over performative agreement. Verify what the reviewer said. Push back when wrong. Implement when right.
Run the verification gate. No exceptions.
Verify. Question. Then implement. Evidence. Then claim.
