# Code Review > Review a live pull request, or the changes since a fixed point (commit, branch, tag, or merge-base), along two separate axes: Standards (documented coding standards plus a code-smell contract) and Spec (fidelity to the originating issue or specification). Pins the exact head, enforces comment-only reviewer authority, and publishes the formal verdict, inline findings, and a marked summary to the PR. Use when asked to review a PR, a branch, or work in progress, or to review since a commit. - Skill: `tticom/code-review` (Agent Skill, multi-file: 14 files) - Install (CLI): `npx skillmds@latest add tticom/code-review` - Raw SKILL.md: https://api.skillmd.com/api/skills/tticom/code-review/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: tticom (https://skillmd.com/u/tticom) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/tticom/code-review --- # Basic code review - **Standards**: does the code conform to this repository's documented coding standards and to the [code-smell contract](references/code-smell-contract.md)? - **Spec**: does the code faithfully implement the originating issue or spec? **Prerequisite skills:** `identity-safe-git` `hard-review` and `devils-advocate-review` build on this skill for riskier changes; they are escalation levels, not prerequisites. All AgentOps skills are installed together in one skills directory. If `identity-safe-git` is not available, stop and report `REQUIRED_SKILL_MISSING identity-safe-git`; do not approximate the identity checks. Require a live PR. If only a local branch or uncommitted diff exists, return a pre-publication assessment or ask the author to publish it; do not open or alter a PR under the reviewer role. Read [reviewer-firewall.md](references/reviewer-firewall.md) first. Review is comment-only: the reviewer never edits, commits, pushes, or merges. Delegation is optional. If your runtime provides sub-agents and the user or task authorises delegation, the two axes may run as separate sub-agents. Otherwise review Standards, then Spec, yourself in one session and keep the two reports separate. Do not assume a particular sub-agent tool exists. ## 1. Establish reviewer authority Resolve the Git-host login, then run the role gate with the project's role policy. The gate is the bundled `scripts/role_authority_gate.py` (byte-identical to the one in `identity-safe-git`): ```text gh api user --jq .login python /scripts/role_authority_gate.py --policy --actor --operation review-metadata --repo --pr --pr-author ``` Stop if the reviewer is also the PR author, the role gate fails, the workspace is not the assigned review workspace, or repository policy requires a stronger review level. Whatever the user named is the fixed point (a commit SHA, branch, tag, `main`, `HEAD~5`, and so on); if none was named, ask. Confirm the fixed point resolves (`git rev-parse `) and the diff is non-empty. A bad ref or empty diff should fail here, before any review work. ## 2. Pin the exact live revision Fetch the head object without merging it, then check it out detached in a dedicated review worktree outside the reviewed repository. Require exact equality between the initial live head, the handback head when one is required, and local `HEAD`: ```text python /scripts/verify_review_head.py --expected --worktree ``` Require a clean checkout before validation. List untracked files and compare every fixture or artifact used by tests with `git ls-files`; an input that exists only in the author's dirty worktree is absent from the reviewed change. Run tests only from the detached exact-head checkout, never from the author's working directory. If the clean checkout cannot reproduce the claimed command, request changes or return `CANNOT_VERIFY` as appropriate. Apply the [code-smell contract](references/code-smell-contract.md) on top of whatever the repository documents. Read it completely. A smell candidate is not automatically a violation: investigate it and classify it as `NOT_PRESENT`, `SUSPECTED`, `CONFIRMED`, or `EXEMPT`. Under a repository or user no-code-smells policy, every diff-introduced or materially worsened `CONFIRMED` smell blocks approval. A repository rule may provide an explicit exemption, but passing tests, subjective disagreement, or calling the smell "residual risk" may not. ## 3. Establish the contract Read, in this order of authority: 1. `AGENTS.md`, review rules, contribution rules, and security or privacy policy; 2. active-task scope and allowed paths when present; 3. the linked issue, specification, acceptance criteria, or architecture decision; 4. the PR body and author handback, as claims only, never as authority over repository policy. Find the spec from issue references in commit messages (`#123`, `Closes #45`, `!67`), from a path the user gave, or from a spec file matching the branch or feature. If the project documents its issue tracker (for example in `docs/agents/issue-tracker.md`), use that workflow to fetch it. If nothing is found, ask where the spec is; if there is none, the Spec axis reports "no spec available". ## 4. Run the two axes **Standards.** With the full diff, the commit list, the standards-source files from step 3, and the complete code-smell contract in hand, report every documented-standard violation and every smell candidate per changed file or hunk. For each smell apply the contract's definition, classify it, and record exact evidence, concrete impact, and the correction or cited exemption. Do not report aesthetic preference as a smell and do not waive a smell merely because tests pass. **Spec.** Run the smallest relevant checks first, then the repository-mandated suite. Do not treat a developer summary or aggregate pass count as execution evidence. Record exact commands, exit codes, and observed failures. Independently enumerate the mandated commands from the spec and repository rules. Record each as `COMPLETED`, `FAILED`, `NOT_RUN`, or `TIMED_OUT`, with its exit code and pass/fail/error/skip/xfail totals. Never translate a partial selection, collection-only output, skipped module, still-running process, or missing receipt into success. A required failure or error blocks approval. Run `git diff --check` and the repository's declared compile, lint, type, and static-analysis checks. For Python changes, `python -m compileall` is a useful syntax baseline when available, but it does not replace production-path execution or justify inventing an undeclared type-checker gate. Build a requirement-conformance matrix before the verdict. Every obligation and prohibition must map to the actual diff, a final observable, and executed evidence. A clean implementation of behaviour that deviates from the contract is `CHANGES_REQUESTED`. This basic review verifies ordinary test coverage. It does not certify test-data provenance or fixture independence. Choose the depth proportionally: for ordinary tooling, documentation, or infrastructure changes this review is the right level. If real-world data, parsers, conversion fidelity, geometry, timing, matching, generated artifacts, private fixtures, or empirical claims are material, escalate to `hard-review`. ## 5. Form findings and verdict Present the two reports under `## Standards` and `## Spec` headings, verbatim or lightly cleaned. Do not merge or rerank findings: the axes are separate on purpose, so a single ranking cannot bury one axis under the other. End with a one-line summary: total findings per axis and the worst issue within each axis. Do not pick a single winner across axes. Each blocking finding must identify the exact path and line or hunk, the observed or deduced failure, the governing requirement, and the smallest acceptable correction. Publish line-specific findings as inline PR review comments whenever a precise changed line exists; put cross-cutting findings in the formal review body. Include the code-smell ledger defined by the smell contract whenever a candidate was found. Under a no-code-smells policy, `APPROVE` is forbidden while any diff-introduced or materially worsened smell remains `CONFIRMED` without a cited repository exemption. Choose exactly one verdict: `APPROVE`, `CHANGES_REQUESTED`, or `CANNOT_VERIFY`. Re-query the live head immediately before publication. If it differs from the reviewed head, publish nothing and restart. ## 6. Publish and prove the result The reviewer must publish the formal verdict, line-level findings, and summary directly to the GitHub pull request. Output left only in chat or local files does not count. Prepare the formal body, the summary, and the optional inline-comment JSON outside the reviewed repository. Publish them together through the guarded publisher: ```text python /scripts/publish_review.py --repo --pr --expected-head --level basic --verdict --role-policy --review-body-file --summary-file [--inline-comments-file ] ``` The publisher fails closed when the role policy is missing, the reviewer is the PR author, or the live head differs from `--expected-head` before or during publication. It creates the formal review, attaches inline comments, and creates or updates exactly one PR issue comment containing: ```text Review level: BASIC Reviewed head: Base: Verdict: Findings: Validation: Residual risk: ``` The formal review body must also contain the full head SHA and `Verdict: `. Do not substitute a chat response, task-state update, committed review report, or local Markdown file for the PR comment. On an unchanged head the publisher updates the existing marked summary instead of creating comment spam. Read the published state back: re-query reviews and comments and prove that the formal verdict, the inline comments, and the marked summary exist on the expected head. The publisher does this itself, from persisted objects: after writing, it fetches the formal review (id, commit, state, body) and its complete inline-comment collection (always, even when none was published, with exact cardinality and each comment's path, line, side, commit and body, where a missing value never matches), and the summary comment (id, author, pull request, marker, body), and prints `REVIEW_PUBLICATION=PASS` only when each matches what was published and exactly one marked summary exists for the head. A write response is never treated as proof, and a non-object element in any fetched list is a failure, never filtered out. It searches every page of the issue comments, so an existing summary is updated wherever it sits in a long thread and never duplicated. Follow the [review-state protocol](references/review-state-protocol.md) when reporting the resulting state. Finally require: - local `HEAD` still equals the reviewed head; - `git status --porcelain=v1 --untracked-files=all` is still empty; - no commit, push, ref update, merge, auto-merge, branch deletion, release, or repository-content API mutation was performed by the reviewer. If a reusable process weakness was found, describe the proposed skill or rule change in the mandatory PR summary. Never implement that improvement in the reviewed repository or during the review session. `scripts/assertion_smells.py` reports assertion shapes that deserve semantic review in changed Python tests; it is advisory and never justifies approval.