Pre-PR gates (review kill-gates)

Run this after the render proof below on a frontend diff, and before a final code review and opening the PR. This skill is the open-PR gate.

The failure mode it kills: a PR that looks finished, then comes back from review with the same HIGH/blocking comments as the last five PRs.

When finished, write this line in visible assistant text (not thinking):

PRE-PR CHECK: ghost=ok; gesture-cancel=ok; except-rollback=ok; prod-path=ok; oracle=ok; siblings=ok; org-id=ok; description=ok; render=ok; console=ok; behaviour=ok; motion=n/a

Any gate that does not apply is n/a, never silently ok. A failing gate is a code fix, not a waiver. If a PR-gate hook reads the artefact from a state file, write it there in its OWN Bash call, before the call that runs gh pr create: a write inside the same command is not seen. poort's PR gate (tools/poort) ties the evidence to the exact commit.

Do not run gh pr create until every applicable gate is ok or n/a.

Gate 1 — Ghost version (feedback_no_ghost_version)

Error, empty, loading, and zero must not look the same.

For every async fetch this diff adds or touches:

• .error is read in the template (or the computed that feeds it).
• Failed fetch does not render as "0 items" / empty state / "—".
• Loading is distinct from empty.

Pattern: siblings already using EmptyState v-if="xAsync.error.value" are the template. A new ?? [] on a failed fetch is a ghost.

Gate 2 — Cancelled gesture must not commit

If the diff binds @pointercancel or a drag/resize pointerup:

• pointercancel is not the same handler as pointerup.
• Cancel reverts to the pre-gesture state and does not persist.
• Escape (if a keyboard cancel exists) takes the same revert path.
• A test exists that cancel does not call persist/onDrop.

Known incident: card-drag had a dedicated cancel; tile-resize rebound @pointercancel to onResizeUp which always persistDraft().

Gate 3 — except + live SQLAlchemy session

If the diff has except Exception (or a broad except) around DB work:

• db.rollback() runs before the loop/caller continues.
• The error is logged with context (logger.exception, not swallowed silent).
• A test proves the session is still usable after the failure (PendingRollbackError must not escape).

A logged-and-continued except without rollback poisons the rest of the request.

Gate 4 — Feature is wired into the production path

If the diff adds a column, flag, or helper:

• Search every writer of that model (ModelName( and db.add).
• The path users actually hit (board confirm, workflow node, public API, not only the new settings endpoint) writes the field.
• A test hits that production path, not only the helper in isolation.

Known incident: expected_start_date existed on create/update in revenue and never on confirm_placement / workflow deal nodes — the feature never fired for board placements.

Gate 5 — Test oracle = production predicate

If the diff adds a filter, realized-check, or date ladder:

• Python helper, SQL helper, and TS port use the same predicate.
• Frozen oracles (scripts/gen_*.py, *oracle*.json) call that helper, they do not reimplement a weaker filter.
• Regenerating the oracle is a conscious step, not a silent skip.

Gate 6 — All sibling consumers

If the diff changes a pattern used on more than one page (activity feed, empty state, contact-moment body, summary tile):

rg "<component-or-pattern>" frontend-v4/src/views/

• Every sibling page got the same change, or an explicit reason why one stayed different is in the PR body.
• "I migrated 3 of 4" is a blocking finding.

Gate 7 — Cross-org isolation test (RLS is dormant)

If the diff adds a table or an org-scoped endpoint:

• A test creates a row in org A and asserts org B reads the default / 404 / empty, not A's data.
• The query filters org_id (the only in-process isolation).

Gate 8 — PR description matches the code

• Routes claimed in the body exist in tests/e2e/00-smoke.spec.ts if they are new top-level routes.
• Invariants in the body match the code (e.g. "never below actual" vs "strictly greater").
• Behaviour delta section lists intended changes; anything else in the diff that changes behaviour is a bug.

Gate 9 — Sink coverage and failing-condition tests

If the diff fixes or adds a callback, handler, guard, redirect or early return:

• The fix sits on the sink (the function every path converges on), not on one of its callers. rg every caller/trigger and list them in the PR body or the VERIFIED: line.
• Every new guard or branch has a test that triggers its failure condition for real (a failing storage delete, a header-only file, a refresh that fails). A source-text test (?raw import plus regex or toContain on the source) pins text, not behaviour, and does not count.

Known incident: 2eb7c32c patched one of three onLogout() triggers; the demo-reset guard and the header-only rejection shipped with no test that hit their condition (#657 review).

If the diff adds CREATE INDEX in db/migrations/:

• The table is created in the same migration, or the index is declared on the ORM model instead (ensure_additive_indexes() builds it CONCURRENTLY). A plain CREATE INDEX on an existing table SHARE-locks it for the whole build during deploy. polpo-guard blocks this on write and on commit.

Gate 10 — Frontend render proof

Applies when the diff touches .vue, .tsx, .jsx, .css, .scss or .html under a frontend directory. Otherwise write render=n/a; console=n/a; behaviour=n/a; motion=n/a.

• Scope first. Routes/components from git diff --name-only. One line per route with the expected fact (a number, a label, a position) written BEFORE any render. "Looks right" is not a fact.
• Render at three widths. Per changed route: skill render-proof (<url> <out.png> "" 1440), then 768 and 390. Read each PNG and read the expected fact off it. A mismatch is a code fix, not a note.
• Console clean. Open each changed route in a browser (or an e2e run) and confirm zero console errors and zero failed requests.
• Behaviour. A changed flow on a demo or non-PII page: an e2e flow test passes (skill e2e, npx --no-install e2e run). Logged-in, hover or drag state: claude-in-chrome with the state visible in the screenshot, else UNVERIFIED.
• Motion. If the diff adds or changes transition, animation, @keyframes or motion props: run skill emil-design-eng.

Auto-run order (do not skip or reorder)

1. Render proof — only for a frontend diff: the render, console and behaviour gates above.
2. This skill (pre-pr-gates) — the ten gates above.
3. find-bugs — full-diff security/logic pass.
4. A code review against standards and spec (e.g. code-review from mattpocock/skills).
5. Open the PR — only after the PRE-PR CHECK: artefact is written.

The render proof is the "the screen is proven" gate. This skill is the "this may become a PR" gate. The code review is the last human-shaped read before gh pr create.