Code Critic

/code-critic — Code Review

Apply this skill to review code changes against project standards.

---

Stance

Treat the implementation as a hypothesis under attack. Your default assumption is that something is wrong; your job is to find what.

Confirmation bias is the dominant failure mode of AI code review — guard against it by actively trying to disconfirm the implementation rather than verifying that it looks reasonable. If you cannot find a fault after genuine effort, that is itself a finding worth stating explicitly (see the Output Format section).

---

Before Reviewing

Selecting the diff

In the /feature workflow the implementation is committed before review, so a bare git diff (working tree) shows nothing — review the branch's committed changes against its base, the default branch named in AGENTS.md → Commands: git diff ...HEAD (or git log -p ..HEAD). Invoked ad-hoc on uncommitted work, review the working-tree diff instead. If unsure what changed, check git status and git log --oneline first.

Note for ad-hoc use: the push gate records a commit SHA (scripts/review-ok.sh), so a review meant to unlock a push must cover the committed state — commit first, then review.

Architecture Decision Records

Check if a docs/adr/ directory exists. If it does, list the ADRs you read by number in the review output. If none seemed relevant to the diff, say so explicitly — silence is not acceptable, since it is indistinguishable from skipping the step.

If the diff contradicts an ADR, quote the contradicting ADR clause and the offending diff line, and flag it as FAIL.

Implementation Plan

If plan text was provided (inline in the prompt, or via a file path), use it before reviewing. Use it as follows:

• Approval Summary / Acceptance Criteria: the human-approved contract.

Verify every AC-n has a committed test (via the [AC-n] tags in Test Strategy) that would fail if the criterion were broken — a criterion without one is FAIL.

• Contract section (if present): cross-check the diff's routes, form

fields/params, response shapes, error rendering, and schema changes against it. An undiscussed deviation from the approved Contract is FAIL.

• Requirements section: the source of truth for what should have been built —

used to verify Plan Compliance and that all specified edge cases are handled.

• Approach section: the agreed implementation strategy — used to verify the

code follows the intended design rather than an ad-hoc alternative.

• Edge Cases section: the enumerated scenarios that must be handled —

cross-reference against the code and tests.

• Test Strategy section: the agreed test coverage — cross-reference against

the actual tests written.

• Files section: the planned file manifest — cross-check against the diff. A

file in the diff that is not listed here (or listed but left untouched) is an undiscussed change; flag the mismatch and judge whether it is in scope.

• Out of Scope section: the explicit boundary — do NOT flag as missing

anything listed here, and DO flag as scope creep any code that strays into it.

If no plan was provided, skip the Plan Compliance checklist section entirely.

Test evidence. The review verifies coverage statically; whether the suite actually ran and passed on the reviewed state is separate evidence. In the /feature workflow that evidence is passed in (the summary of the latest full test-suite run). Treat that summary as a record of the run, not independent proof — it is produced by the implementing agent, and this review has no independent way to confirm it (whether the project's own CI provides that is outside this skill's scope). If no evidence was provided and you cannot (or may not) run the suite yourself, do not assume it is green — raise an Open Question: "no evidence the test suite ran on the reviewed state".

When flagging a plan compliance issue, quote the exact line from the plan that the diff violates, alongside the diff line that violates it. Paraphrasing the plan is not enough — the developer needs to see the literal mismatch.

Reading Beyond the Diff

You may read any file in the repository. The diff is the unit under review, but surrounding code, tests, configuration, and migrations are fair context — and often necessary to judge whether the change is correct. Some project-specific rules below may explicitly require it (cross-referencing a sibling file, checking that a registration or annotation is present).

---

Coding Standards

Single Responsibility

Functions: A function should do one thing. If a function contains a conditional branch that handles a fundamentally different concern (e.g. an admin path bolted onto a regular-user path, or a parsing path inside a persistence call), flag it. Severity:

• New function introduced with mixed concerns → FAIL
• Existing function extended with a clearly unrelated branch → FAIL
• Borderline case where extraction would hurt readability → NEEDS_DECISION

Classes: A class has too many responsibilities if you can identify more than one independent reason it would need to change.

When reviewing new code that touches an existing class:

• If the new code introduces a class with too many responsibilities → FAIL
• If the new code significantly worsens an existing violation (e.g., adding

several more methods to an already oversized class) → FAIL

• If the new code extends an already-oversized class without making it

meaningfully worse → NEEDS_DECISION: flag the pre-existing debt and let the developer decide whether to refactor now, file tech debt, or accept it

Error Handling

• In business logic, never use generic catch-all error handling. At process

boundaries (top-level controllers, scheduled jobs, async task entry points) a catch-all that logs and translates to a domain error is acceptable — judge by where the catch lives, not just what it catches.

• Every error must include context about what operation failed and why.
• Errors in critical paths must be logged with structured fields.

Naming

• Functions should describe what they do: calculateShippingCost, not calc or

process.

• Boolean variables/functions should read as questions: isValid, hasPermission.

Tests

You are responsible for both test quality (structure, naming, pattern) AND test completeness (coverage of the spec AND beyond). The /adversarial-qa skill is purely exploratory/adversarial in a running browser — it does not verify that the committed tests cover the plan, so that responsibility lives here.

Quality:

• Every public function with non-trivial behaviour must have tests covering

the happy path AND edge cases. Trivial delegators, generated code, plain data classes, and pure getters are exempt — but if you exempt a function, state which one and why in the review.

• Test names must describe the scenario: test_order_fails_when_inventory_insufficient,

not test_order_2.

• Use the Given-When-Then pattern.
• Never test implementation details — test behaviour.
• Tests that encode the plan's Test Strategy (the [AC-n]-tagged ones

especially) are the contract, not implementation detail: a diff that weakens, loosens, or deletes one so the suite passes is FAIL unless the review input documents an approved plan deviation covering it.

Completeness — against the plan:

• Every Requirement and every enumerated Edge Case in the plan must have at

least one test that would fail if the requirement were broken.

• Every branch introduced in the implementation (error paths, validation

failures, authorization denials, empty/null handling) must be exercised.

Completeness — beyond the plan (critical analysis):

Read the implementation diff and actively try to break it. You are doing this through static reading, not by running the code — for each branch in the diff, mentally construct an input or sequence of events that would exercise it, then ask whether a test covers that input. The plan is a starting point, not a ceiling. For each category below, ask "given the actual code in the diff, what would I do to make this fail?":

• Implicit branches: if, when/switch, ?:, early returns, exception

handlers introduced by the implementation but not called out in the plan. Each one is a behaviour worth testing.

• Boundary values: 0, 1, max, min, off-by-one, empty collections,

single-element collections, exactly-at-limit vs just-over-limit. Plans rarely enumerate these exhaustively.

• Input shapes the plan didn't mention: null, blank strings, whitespace,

unicode, very long strings, leading/trailing spaces, mixed case, duplicates, unsorted input.

• State and concurrency: stale reads, double submits, retries on the same

resource, partial writes, what happens if the operation is invoked twice.

• External dependency failure modes: timeouts, 4xx vs 5xx responses,

malformed payloads, slow responses — wherever the code calls out.

• Security-adjacent gaps: authz checks on every entry point, not just the

one the plan mentioned; injection-shaped inputs on any field that hits a query, template, or shell.

If you find a gap of this kind, flag it as FAIL (or NEEDS_DECISION if it is genuinely ambiguous whether the case is in scope) with a concrete description of the missing test, not just "more tests needed".

Privacy and Data Protection

Prefer build-enforced tests over prose here (same doctrine as the build-enforced guidance below). Any project holding personal data should implement these three fitness tests — until they exist, they are backlog items, not per-PR checklist prose:

1. Deletion by design — every table with a user FK either cascades on

account deletion or appears in an explicit, commented allowlist.

1. Public-surface whitelist — the model rendered on public/

unauthenticated surfaces is a distinct type whose fields are asserted against a whitelist, so a sensitive field cannot be added silently. (If public pages currently render from the full domain object, that restructuring is the prerequisite — worth its own task.)

1. No personal data in logs — log statements must not reference

sensitive/user-content field symbols or contact-detail fields.

For any of these that IS implemented, the reviewer's only duty is the standard build-enforced check: the diff must not WEAKEN the enforcement (deleting or disabling the test, adding an unexplained allowlist entry, restructuring code out of the test's scan scope). A weakened enforcement is FAIL. For any not yet implemented, check the corresponding invariant by hand only on diffs that touch it (new user-FK tables; public-view rendering; new/changed log statements).

Prose residue the tests cannot catch (diff-scoped, manual):

• Indirect serialization into logs — logging rich objects (toString()

of an entity, dumped request params) that embed personal data. FAIL.

• Personal data in URLs — emails, phone numbers, or values embedding

them in query params or path segments; URLs land in access logs and browser history. FAIL.

Consent, retention, and data-classification rules are deliberately NOT per-PR review items — they are product flows to be built as tasks and then protected by tests, not per-diff checklist prose.

---

Project-Specific Rules

This skill is project-agnostic; each project extends it without editing it. Check the repo-root AGENTS.md (not a module-level one) for a Review & Planning Guidance section. If it has a "Code review guidance" entry, read the file it names. If AGENTS.md has no such section, or the section exists but has no "Code review guidance" entry, fall back to checking docs/agent-rules/code-critic.md directly. If an entry names a file that doesn't exist, treat it the same as "no file found" below, but say so specifically (a named-but-missing file is a broken pointer worth surfacing, not just an absent extension).

If a file is found, apply every constraint in it alongside the base standards above: where it states a severity, use it exactly as stated (including the don't-weaken doctrine for anything it marks as build-enforced); where it states none — a pre-existing project doc without this skill's format, such as a style guide, CONTRIBUTING.md, or an engineering handbook — judge severity yourself using this skill's normal PASS/FAIL/NEEDS_DECISION framework. Either way, treat any PRIVACY anchors it contains as binding on the privacy rules above (sensitive categories, public surfaces, existing fitness tests), regardless of whether the rest of the file follows this skill's structure.

If no file is found either way, proceed with the base standards alone and say so in the review output (one line) — the gap should be visible, not silent.

---

Review Checklist

Review every change against this checklist. For each item, state one of:

• PASS — the rule applies to this diff and the diff complies. Cite specific

diff lines that demonstrate compliance for non-trivial items (test coverage, edge cases, error paths, security-adjacent code). Routine quality items (naming, readability) need only a brief explanation.

• PASS (N/A) — the rule does not apply to this diff (e.g., security checks

on a CSS-only change). Briefly say why.

• FAIL — rule violated, and you can cite the line and the rule that breaks.

Reserve FAIL for findings you are confident in. If unsure, prefer NEEDS_DECISION or Open Question — false FAILs burn developer trust faster than missed issues.

• NEEDS_DECISION — the correct approach is ambiguous and requires

developer input. State the options.

• Open Question — concerns where the diff might be wrong but you

cannot verify from source alone (depends on runtime config, prod data shape, deployment topology, external service behaviour). State the question and what evidence would resolve it. Use this freely — it is better to surface a hypothesis the developer can dismiss in 10 seconds than to stay silent on a real risk.

Architecture

If a layer/architecture fitness test enforces boundaries (see the layer rule in the AI Workflow plugin's own documentation), do NOT re-derive them by hand — verify only that the diff does not weaken it: deleting or disabling the test, widening a package glob, adding an unexplained exemption, or moving code out of the scanned layer. A weakened enforcement is FAIL.

• [ ] Changes respect service boundaries and existing ADRs
• [ ] No business logic in infrastructure or API layers
• [ ] No new dependencies introduced without justification

Plan Compliance (skip if no plan was provided)

• [ ] Implementation follows the Approach described in the plan — no undiscussed

design alternatives introduced

• [ ] All steps in the plan are accounted for in the changes
• [ ] Every acceptance criterion (AC-n) in the Approval Summary maps to a

committed test tagged [AC-n] that would fail if the criterion were broken

• [ ] Diff matches the plan's Contract section — routes, fields, response

shapes, error rendering, schema (skip if Contract is "None")

• [ ] All requirements from the Requirements section are addressed
• [ ] Files touched in the diff match the plan's Files manifest — no undiscussed

files (a file in the diff but not in the manifest is an undiscussed change)

• [ ] No code introduced inside the plan's Out of Scope boundary

Code Quality

• [ ] Every function has a single responsibility
• [ ] Classes have a single reason to change — new classes assessed per the Single

Responsibility standard above; existing classes flagged if the change worsens the violation

• [ ] Error handling is specific, not generic catch-all
• [ ] Code is readable — a new team member could follow the logic without extra context
• [ ] No implicit assumptions that should be explicit (add

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: cunhaax
• Source: cunhaax/ai-workflow
• License: MIT

Install and usage instructions live in the source repository linked above.