Review changed code for naming, stale references, unnecessary complexity, and comment quality. Use after completing implementation work, before committing, or when the user asks to review or audit code.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Review changed code for naming, stale references, unnecessary complexity, and comment quality. Use after completing implementation work, before committing, or when the user asks to review or audit code.
allowed-tools
Read Grep Glob Bash
Code Review
Review the diff or specified files against these principles.
1. Names must reflect current reality
Variable and function names should describe what they ARE, not what they used to be.
If the underlying mechanism changed (e.g. FUSE → NFS), all related names must update.
Ask: would a new reader be confused by this name?
2. No stale references
After refactoring, grep for references to the old approach — dead detection logic, abandoned feature flags, comments mentioning removed code.
If something was tried and reverted, remove ALL traces. The codebase should look like the current approach was always the plan.
3. Simplify detection and guard logic
A gate like "can this feature run" should check the ONE thing that actually matters.
Don't chain fallback detections (binary exists OR source exists OR toolchain exists) when one check covers it.
4. Comments document WHY, not WHAT HAPPENED
Keep: non-obvious technical discoveries, platform quirks, "if you remove this, X breaks because Y."
Good: // com.apple.provenance causes SIGKILL when spawned as child process
Good: // umount while server is alive panics the macOS NFS client
Good: // NFS client uses cookie verifier to decide if cached readdir is valid
Remove: narrative of what was tried, what was abandoned, what was renamed.
Bad: // We changed this from cp to cat to work around the provenance issue
Bad: // Wrapper added because FUSE-T was zero-padding (see commit abc123)
Bad: // Previously this was called fuseAvailable but we renamed it
Remove: comments that just restate what the code does without adding reasoning.
Often no comment is needed at all.
5. Separate install-time from runtime
Install scripts set up prerequisites (toolchains, system deps).
Don't mix these. If the install script is compiling binaries that the runtime also compiles, one of them is wrong.
6. No intermediary artifacts
After iterating through multiple approaches, audit for logic that only existed as a stepping stone.
If a workaround was added for approach A and you switched to approach B, remove the workaround even if it's harmless.
The code should read as if written from scratch by someone who already knew the right answer.
7. Tighten guarantees after changes
When a change makes something guaranteed (e.g. a variable is now always set, a file is always created, a function always returns), audit downstream code that still guards against the old "maybe" state.
Redundant null checks, fallback defaults, and if (x) guards on values that can no longer be null are misleading — they imply a possibility that doesn't exist.
Prefer const over let when reassignment is no longer needed.
Never suppress a signal — fix the root cause._ prefix on unused params, // @ts-ignore, eslint-disable, as any — these all hide real issues. If a parameter is unused, DELETE it and cascade the removal to every caller. If a type doesn't match, fix the type. Run the checker, fix every error, repeat. Mechanical changes across many files is exactly what an agent excels at — there is no "too many callers."
8. When changes are tangled, start clean
If a file has been through 3+ rounds of conflicting edits, git restore it and re-apply only what's needed.
Don't try to surgically fix a mess — starting clean is faster and less error-prone.
9. Tests must verify actual values, not just collection sizes
When testing deduplication, normalization, or idempotent operations, assert on the stored value, not just toHaveLength(1).
Length alone doesn't prove the logic worked — a broken normalizer could silently drop one input, or store two different normalized forms that happen to match.
Pattern: add value in format A, re-add in format B, then assert both that the count is 1 AND that the stored value equals the expected normalized form.
Bad: expect(data.phones).toHaveLength(1) — passes even if normalization is broken
Good: expect(data.phones).toHaveLength(1); expect(data.phones[0]).toBe('+12125551234') — proves normalization recognized both formats
10. No thin wrappers or re-export-only modules
If a module only forwards another package's functions or types, delete it and import the original package directly.
Do not preserve local compatibility shims for scaffold code or unshipped branch work. Replace the scaffold outright.
Re-export barrels are not useful unless they define a real public boundary with project-specific semantics. Avoid creating files whose only job is export * from ....
Prefer using the upstream API name directly over inventing local aliases like discoverAll() for loadSkills().
11. Extract only around real ownership boundaries
Do not split files just to reduce line count. A new file should own a coherent responsibility that a reader can name.
Extract modules when they concentrate related policy, state transitions, resource handling, or domain-specific behavior.
Avoid moving one-off helper functions into a new file if the caller still needs to understand all the details to use them.
A good split usually reduces import pressure in the original file because dependencies move with the responsibility they serve.
If extraction increases total indirection without clarifying ownership, keep the code local.
12. Decouple tests from implementation — drive the system end-to-end
The most valuable test suite is the one most decoupled from the implementation it covers. Decoupled to the point where you exercise the backend by driving the frontend, and exercise a module by going through the same entry point a user goes through. A test that pokes at internals freezes the internals; a test that drives the public surface frees you to refactor everything underneath.
Prefer the outermost entry point that still gives a fast, deterministic signal. CLI binary > top-level exported function > internal helper > private method. If the CLI is what users invoke, write the test against the CLI. If a TUI is what users see, drive it through the same input pipeline real keystrokes hit, and assert on the rendered frame, not on intermediate state.
Test behavior, not structure. Assert on observable outcomes — exit codes, stdout, rendered output, files on disk, HTTP responses, persisted rows. Do not assert on which functions were called, in what order, with which intermediate shapes. Those are implementation details that should be free to change.
A passing test should mean a real user gets the right result. If a test can pass while the production path is broken, the test is wired wrong. Common smells: stubbing the thing under test, bypassing the router/dispatcher/parser, hand-constructing internal events that production would have built from input.
Harnesses must wire the system the way production wires it. Optional inputs that are always set in practice are part of the contract — pass equivalents in the harness (e.g. a settled status stream, the modal config, the same env vars). If you found a bug only by running against a real environment, the harness skipped something production sets; fix the harness so the next regression in the same shape is caught by the test suite, not by a user.
Coupled tests are a refactor tax. When renaming an internal function or moving a module breaks dozens of tests without changing any user-visible behavior, the suite is testing the wrong layer. Rewrite those tests against the outer surface and delete the brittle ones.
Reserve unit tests for genuinely tricky pure logic — parsers, normalizers, schedulers, state machines with subtle invariants. Everything else earns more value as an integration or end-to-end test.
13. A failing check can be a toolchain defect, not a code defect
Before hand-writing a type, adding a cast, pinning a value, or restructuring code to make a checker (type-checker, compiler, linter, test runner, build) pass, confirm the failure is a CODE defect and not a toolchain/environment artifact. Patching code to satisfy a broken or mismatched tool is a workaround that masks the real problem — the never-suppress-a-signal rule, one level up.
Reproduce under the exact toolchain that reports the failure — the CI/deploy version and config, not just your local one. A green local check proves nothing if it ran a different version: a pinned prerelease, a preview build, or version drift between your machine and CI can pass locally and fail in CI on identical source (or the reverse).
If the same source passes under the real/pinned tool, the code is correct — the fix belongs in the toolchain (pin the version, fix the config), not the code. When the output is ambiguous, probe the tool directly — force it to print the value, type, or error it actually computed — instead of guessing at the cause.
Keep local == CI. Confirm the checks that gate merge/deploy run the same toolchain the deploy runs; version drift makes every green check suspect, and "it passed locally" stops being evidence the deploy will.
14. Identity comes from auth, never from the caller
Resolve the acting identity — org/tenant/workspace id, user id, actor — from the authenticated request (API key, session, or a verified internal key), never from a caller-supplied argument. A public mutation that reads actorId/orgId from its args is spoofing surface, not a feature.
The only exception is a trusted internal call: accept a caller-supplied actor ONLY when a valid internal key is present; otherwise derive it from auth and reject the supplied id.
To scope to a child inside the authed scope (a specific channel, share, app), take the child id and verify it belongs to the auth-derived parent — don't trust a parallel parent id from the body.
15. Public API actions are thin auth + dispatch, not inlined business logic
A public route authenticates, validates, derives identity, and dispatches to a model function or an internal action that owns the work. Keep that wrapper thin — it is the auth/runtime boundary.
Don't inline third-party SDK / analytics / email logic into the public boundary because "it's only 3 lines." Push it behind the dispatch.
In Convex specifically: do NOT add 'use node' to a file holding many actions just to satisfy one — it forces every action in the file into the Node runtime. Split: auth+dispatch (default runtime) → services/<thing>.ts ('use node', owns the SDK call).
16. Public endpoints ship at minimum scope
One route, one item per request. Don't preemptively add /batch variants, paginated listings, or filter params before a concrete second caller needs them.
Inline the schema in the route; don't hoist a 4-line schema into a shared schemas.ts "for reuse" when nothing reuses it. Skip body-size caps unless the body is genuinely unbounded.
If a type already exists upstream (a schema validator, protocol package, generated client, SDK), use it directly — don't redeclare its shape inline, even partially. Redeclared shapes drift.
For a variant, derive it: Pick/Omit/Partial/Parameters/ReturnType/indexed access/Extract/Exclude; Convex/Zod equivalents Infer<typeof V>, FunctionArgs<typeof api.x.y>, z.input/z.output.
Bad: type UserSummary = { id: string; name: string; email: string } next to an existing User. Good: Pick<User, 'id' | 'name' | 'email'>.
18. Filter at the index, never take() + post-filter
When a list query returns rows from a sub-bucket of a table (active vs. archived, status === X), the index must do the filtering. Don't .take(N) the unfiltered query and .filter() in JS for the bucket you want.
Why it breaks: .take(N) returns the N rows at the head of the index's order. If the head is full of the other bucket, the post-filter returns zero even when the bucket has hundreds of older rows. take(N*2) only delays the failure.
Pattern: put the discriminator in an index and use .withIndex(..., q => q.eq(...)), or q.gt(field, 0) for the "present" bucket (Convex sorts undefined before defined values). Same for .first()/.unique() and "is there any X" probes.
19. Never .collect() — bound the read with .take(N)
Convex .collect() reads every matching row with no upper bound. For anything that accumulates per-tenant over time (sessions, events, ledger rows, audit rows), a query fine on day one eventually pulls thousands of rows on one reactive tick and silently degrades reactivity for the whole client.
Replace .collect() with .take(N) where N is a safety ceiling clearly above today's working set (100, 1000) — a cap against pathological data, not a UX paginator. Same for .withIndex(...).filter(...).collect().
If you genuinely need every row (a migration, admin dump, one-shot maintenance), say so with a comment and run it from a cancellable mutation/action, never a reactive query.
20. Don't remap rows just to rename or default fields
Object-literal .map()s that copy every field through to rename two or coalesce undefined → false are noise, and they drop new columns silently until someone updates the map. If the consumer needs every field, return the row; if a subset, Pick/Omit or destructure-and-rest. Only rename when the new name materially clarifies; only default when downstream truly can't handle absence.
Strip storage-only fields (_id, _creationTime, internal ids) once with a destructure-rest. Good: return rows.map(({ _id, _creationTime, ...row }) => row).
21. Internal APIs carry no version or compat machinery
When we own both producer and consumer (our own apps, services, and functions), do not add protocolVersion fields, version negotiation, capability flags, or "in case the other side is older" branches. Deploying is the version.
The one real compat axis is fields, not versions — and it's an asymmetry, not a knob: parse requests strictly (reject unknown fields), parse responses/events tolerantly (ignore unknown fields). That lets the producer add fields without a lockstep consumer release, which covers the only skew we actually have (components that update on their own schedule).
Smells: a version literal in a wire schema that nothing reads; an if (payload.v >= 2) branch whose only caller is code we deploy ourselves; a strict parser on a response, which turns every additive server change into a breaking one.
Exception: a genuinely external API (consumers we cannot redeploy) versions at the route (/v1/), never per-field.
22. Things that must agree need one owner
When two declarations must stay consistent but each can change alone, they drift. Give the shared decision one home and have each site compose from it — a parallel enum, picker list, or switch that restates a set living elsewhere is the common case (deriving the shape is rule 17; this is the wider rule for any values that must agree).
The harder half: where near-twins legitimately differ, justify the difference in the code. An undocumented divergence between otherwise-identical things is indistinguishable from a drift bug — a reader can't tell intent from oversight.
23. Don't hoist a single-use value into a named const unless it earns the name
A const extracted to module scope but referenced exactly once adds indirection without payoff: the reader has to jump to the declaration to learn the value, and the name restates what an inline value + short comment would say anyway. Inline it at the one call site and let a comment carry the WHY.
A single-use named const IS justified when at least one of these holds:
It's configuration that changes often or that an operator/reader is expected to tune (timeouts/limits grouped as knobs, feature thresholds, retry budgets that get adjusted).
It sits next to related consts and gains meaning from the cluster (a block of *_TIMEOUT_MS, a table of limits, sibling enum members) — the grouping is the documentation.
Declaring it independently is structurally meaningful: it's exported as part of a module's public surface, referenced by a type, or co-located with the data/file it parameterizes so a future second caller finds it.
Otherwise inline. The reviewer test: if the name only exists to label a literal used once, and it neither changes often nor lives beside kin, it's noise — fold it into the call site with a comment explaining the value.
Bad: const STOP_SESSION_CONNECT_TIMEOUT_MS = 5_000 declared on its own, used in exactly one runAction({ connectTimeoutMs: STOP_SESSION_CONNECT_TIMEOUT_MS }).
Good: connectTimeoutMs: 5_000, // 5s: a cold sandbox must not hang on the 60s default connect at the call site.
24. Batch loops isolate per-item failures — one bad item never starves the rest
Any loop over independent work items (sweep, cron batch, queue drain, fleet pass, fanout) must catch per item and continue — collect failures into the result (a failed list) or log them. An uncaught per-item throw aborts the pass, and because the runner (cron, scheduler, sync rail) re-selects the same ordered set next tick, a deterministically-failing item becomes a permanent head-of-line blocker: everything behind it re-queues forever while metrics show the job "retrying".
In transactional contexts (e.g. a database mutation that processes a page and advances a cursor) the failure is worse: the uncaught throw also rolls back the cursor/progress writes, so the same page re-runs forever. A caught error keeps the transaction alive — but the failed item's own pre-throw writes persist, so catch at item boundaries whose writes are idempotent or safe to re-run.
Early-exit on failure is only correct when the failure is provably batch-wide (the shared downstream is unreachable), never for item-specific errors. If the pass's caller needs failure visibility, catch-record-continue and rethrow the first error after the pass completes.
Watch the ordering trap even with isolation: if successful items stay in the candidate set (e.g. re-snapshotting everything before the blocker each retry), the sweep also needs its selection to exclude already-done work.
Accepted shapes: per-item try/catch + failed in the result; Promise.allSettled keyed by item; catch-record-continue then rethrow-after-pass.
Bad: for (const ws of candidates) { await archive(ws); await mark(ws) } inside an hourly cron — one unreachable VM freezes every candidate behind it, forever.
Good: the same loop with the whole per-item body in try/catch pushing { workspaceId, error } onto a failed array returned to the caller.
25. Renaming what users see means sweeping what tests see
User-facing copy, accessible names, routes, and data-* attributes are load-bearing for test layers that do not run in your gates — live-staging harnesses, journey suites, external monitors. A rename that keeps every local gate green can silently break them hours later in someone else's session.
When a change renames visible copy or restructures a surface, repo-wide grep the OLD strings/selectors — explicitly including test-harness and e2e packages — and update every anchor in the same pass. The diff that renames is the diff that sweeps.
The reviewer test: pick a renamed string from the diff and grep the repo for it. Any surviving hit outside the diff is a break waiting for the next live run.
Structural beats copy: where a journey needs an anchor, prefer an owned data-* attribute on the surface over its display text — then product copy can change freely without touching tests. Flag journeys that anchor on copy when a structural attribute already exists.
26. Leave the code better than you found it — a tidy diff is not worth an untidy repo
Never revert an incidental improvement to keep your diff focused. If the formatter, the linter, or a codemod also cleans files your change didn't touch, commit that too (as its own commit if it's noisy). Reverting it optimizes for how the diff reads at review time and taxes every future pass, which pays the same cost again and re-reverts it again. The repo's health outranks the diff's tidiness.
Dead things go in the pass that finds them. A dead column, argument, function, export, or a test whose only callers are tests — delete it, plus any comment defending it, right there. Don't file it; a note is a deferral, and the next reader has to re-derive that it's dead.
A comment you discover is false is a defect. Fix it in place. A wrong comment outlives the diff that would have corrected it and actively misleads — worse than no comment at all.
Fix latent bugs your change surfaces. Work that makes a rarely-run path run every time will expose ordering bugs, unreachable cleanup, and assertions that never fired. That is your bug now: it became reachable because of you.
Verify before you "improve". Confirm the thing is actually dead, wrong, or broken before deleting or rewriting it — grep for consumers, run the test. A confident deletion of something load-bearing is far worse than the untidiness you were fixing.
The limits: don't smuggle unrelated behavior changes into a feature commit, and don't rewrite a neighbouring module because you dislike its shape. This rule covers mechanical tidiness, dead code, false comments, and bugs you made reachable — not opportunistic redesign.
The reviewer test: after this change lands, is any file in the repo worse off, or carrying a known-wrong statement, than before it started? If yes, the pass isn't done.
Your task
Review: $ARGUMENTS
If no arguments given, review git diff --staged or git diff (unstaged changes).
For each issue found, cite the file and line number. Group by category. End with a clean/not-clean verdict.
