self-review

Read the whole diff as someone who didn't write it: gh pr diff NNN (or git diff main...HEAD), plus the PR body. Then check, in this order.

1. does it do what was asked — no more, no less

• the user's ask, verbatim, against what the diff delivers. scope quietly narrowed or widened is a finding.
• every behavior the PR body claims exists in the diff; every non-behavior it claims is true.

2. correctness

• trace the main path and the failure paths by hand. an error that becomes a 500 where a 4xx was meant, a refusal swallowed, a partial write with no rollback.
• concurrency and ordering: anything that runs after a redirect, a deploy, an await on a slow network — what happens if the world changed meanwhile?
• data: a write to a user's PDS or a row rewrite on their behalf needs explicit user consent, not a stored token. flag it.

3. the tests actually test it

• a regression test must exercise the real path (drive the endpoint / mount the real component, mock only the network boundary) and must fail on the pre-fix code — stash the fix and run it once to prove that.
• no tests of what the type system already enforces.
• frontend tests exist for pure logic and mounted components; e2e (frontend/e2e/*.mjs) for anything needing a real PDS.

4. the repo's rules (the ones that get missed)

• no paragraph comments; one short line only when genuinely non-obvious. rationale goes in the PR body / STATUS.md / docs.
• no deferred imports except for circular-import breaks.
• loq limits: just loq-relax <file>, never hand-edit loq.toml or golf lines.
• never revert formatter output.
• copy: lowercase voice, outcome not mechanism, no protocol vocabulary in user-facing strings, no promises the system doesn't keep (times, guarantees). toasts ≤10 words (toast-copy skill).
• a new surface for an existing action reuses that action's existing copy — grep for the sibling's toast/label before writing a fresh one (the queue's bare "liked" vs the menu's "liked ", #1917).
• a keyboard/shortcut addition checks the global shortcut map first; a collision is a question for the user, never a silent contextual override (#1915).
• frontend: no localStorage for auth; per-user prefs from the account-scoped store; redirectToLogin() helpers, never goto('/login'); $effect that reads and writes the same state needs untrack.
• new routes/routers are a product-surface decision — confirm they were asked for.

5. verification claims

• "verified on staging" means the user-facing surface was rendered and measured, not that an API returned 200. for UI diffs, the ui-check matrix (widths × themes × states) is the bar.
• anything slow or failing in CI was diagnosed (telemetry, logs) — not labeled flaky and retried.
• the PR body follows the repo shape: one-or-two-sentence punchline, minutiae under <details>, intentional non-behaviors listed.

output

Findings most severe first, each with file:line, what breaks, and the fix. Fix everything you found yourself in the same PR, then say what you changed. If nothing survives, say so plainly and what you checked.