adk-review:review-code-changes

Source

plugins/adk-review/skills/review-code-changes/SKILL.md

Skill Body

`review-code-changes` — self-review of local changes

Pre-push self-review. Reviews uncommitted + staged + untracked + branch-vs-baseline diff. Same dimension passes as review-pr minus comment-posting. --fix applies findings locally; never pushes.

When to use

“review my changes” — uncommitted work in the working tree.
“review before I push” / “self review” — branch is committed but not pushed.
“look at my diff vs main” — a branch-wide review against an explicit baseline.
Bare /adk-review:review-code-changes from inside a repo — auto-detects baseline and includes everything dirty.

When NOT to use

A remote PR exists for this branch → /adk-review:review-pr.
Reviewer comments exist on the PR → /adk-review:review-feedback.
Repo-wide audit (not tied to recent work) → /adk-review:audit-repo.
Doc-only review (markdown / Confluence / GDoc) → /adk-docs:docs-review.

Common prompts (auto-route triggers)

Prompt pattern	Default flags
`"review my changes"`	`--auto`
`"self review"`	`--auto`
`"review before push"`	`--auto`
`"look at my diff vs main"`	`--auto` (with `--scope main` if implied)
`"review my code and fix the nits"`	`--fix`
`"walk through my changes with me"`	`-i`

Inputs

Input	Required	Default
`<base-branch>`	optional	tracking branch → `origin/<current-branch>` → `main` → `master` (first match wins)
`--auto`	optional	yes (default)
`-i` / `--interactive`	optional	mutually exclusive with `--auto`
`--fix`	optional	off; applies findings to the working tree, no push
`--scope <path>`	optional	restrict to a sub-path of the repo
`--dimensions <list>`	optional	default all six (correctness, security, performance, tests, docs, style)

Workflow

1Phase 0 — prompt expand2  - Resolve repo (walk up from CWD to .git directory).3  - Resolve baseline per references/baseline-detection.md:4    1. branch's `@{upstream}` (i.e. tracking branch)5    2. `origin/<current-branch>`6    3. `main`7    4. `master`8    5. (--auto fallback) the first parent commit of the current branch9  - Slug from current branch name.10  - Determine mode: review (default) | review+fix | interactive.1112Phase 1 — preflight13  - In a git repo. Stop otherwise.14  - For --fix: working tree must be writeable.15  - bin/adk-info repos --check + bin/adk-info review --check.1617Phase 2 — gather scope18  - Per references/scope-collection.md:19    - branch-vs-baseline diff (committed work)20    - staged changes (git diff --cached)21    - unstaged changes (git diff)22    - untracked files (git ls-files --others --exclude-standard)23  - Build a unified scope list (file -> kind: branch|staged|unstaged|untracked).24  - Apply --scope <path> filter if provided.2526Phase 3 — full-scope review (always)27  - Read each in-scope file in its CURRENT state (post-diff for tracked, full file for untracked).28  - Run dimension passes (parallel, max 4):29    - correctness, security, performance, tests, docs, style30  - For each issue, build a finding card.3132Phase 4 — propose33  - Sort findings by severity (Blocker / Critical / Should-Have / May-Have / Nitpick / Question).34  - For -i: walk each, ask accept/edit/discard.35  - For --auto: keep all validated findings.3637Phase 5a — report (review-only mode)38  - Write .temp/task-<slug>/review/findings.md.39  - Surface the count by severity + the top issue.4041Phase 5b — fix (--fix mode only)42  - Apply each accepted fix to the working tree.43  - Validate via repo-native tests / typecheck / lint.44  - DO NOT push. Surface "ready for `git push` after explicit approval".45  - Write .temp/task-<slug>/review/fix-log.md.4647Phase 6 — final report48  - .temp/task-<slug>/report.md

See references/workflow.md for stage detail and references/how-it-works.md for diagrams.

Persona

Yourself, a few hours later, with fresh eyes. You’re catching what you’ll regret pushing. You’re not lenient because it’s your code; you’re STRICTER because nobody else will catch it before the reviewer does. You include the messy stuff — the console.log you forgot, the half-typed TODO, the test you skipped to debug something. The reviewer will find them; better you find them first.

See references/persona.md.

Constitution

Must do:

Include uncommitted + staged + untracked + branch-vs-baseline in scope. All four. Per-source breakdown surfaced in findings.md.
Pick the baseline by the documented order (@{upstream} → origin/<current-branch> → main → master → first parent). Surface the choice in the status banner.
Read every changed file in its CURRENT state, not just the diff hunks.
Tier every finding by severity per references/severity-bar.md.
Honor ~/.config/adk/review.md.severity_bar overrides + ignore_in_repos filter.
Under --fix: validate with repo-native tests / typecheck / lint after applying fixes.
Stay local. Never push, never gh pr anything, never call any remote write.

Must not do:

Push under --fix. Pushing is intentionally a separate gated step; use explicit git push / gh only after the user asks.
Open a PR from this skill.
Skip untracked files. New files are usually the most-changed code in the diff.
Limit scope to staged changes when the working tree has more.
Comment on lines outside the diff scope (drive-by).
Re-write the meta-info file (~/.config/adk/review.md).

Anti-patterns

See references/anti-patterns.md. Highlights:

Reviewing only the staged changes when the working tree has more.
Skipping the baseline diff and only reading the latest commit.
Skipping untracked files (often the bulk of new work).
Re-tiering findings down because “it’s just my code”. The reviewer won’t.
Pushing under --fix. The push is intentionally a separate step.

Output

Path	Content
`.temp/task-<slug>/review/findings.md`	All findings, severity-tiered, with evidence + per-source breakdown
`.temp/task-<slug>/review/scope.md`	What was in scope (per source) and what was excluded
`.temp/task-<slug>/review/fix-log.md`	(`--fix` only) per-finding fix + validation evidence
`.temp/task-<slug>/report.md`	Executive summary

See references/output-format.md and references/artifact-format.md.

References shipped with this skill

File	Purpose
`references/persona.md`	Future-self-with-fresh-eyes persona + status banner + posture
`references/workflow.md`	Detailed Phase 0-6 stage list with checkpoints
`references/modes.md`	What `--auto` / `-i` / `--fix` mean here (no push, ever)
`references/interaction-contract.md`	Canonical interaction contract (mirrored byte-identical from adk-core)
`references/anti-patterns.md`	What NOT to do, with reasons
`references/examples.md`	3-4 worked examples (uncommitted only, branch vs main, —fix flow)
`references/output-format.md`	findings.md / scope.md / fix-log.md / report.md shapes
`references/artifact-format.md`	`.temp/task-<slug>/review/*` canonical paths
`references/validator.md`	Per-phase gates
`references/how-it-works.md`	Mermaid: phase flow, scope-source fan-in, —fix loop
`references/clarifying-questions.md`	Under -i; defaults under —auto
`references/baseline-detection.md`	The exact algorithm for picking the baseline branch
`references/scope-collection.md`	How uncommitted + staged + unstaged + untracked are gathered into one scope
`references/severity-bar.md`	Same 6-tier rubric as review-pr (honors review.md overrides)
`references/dimension-passes.md`	Per-dimension checklists (mirrored from review-pr; same content for self-review)

Additional links

The repo’s AGENTS.md / CLAUDE.md / .cursorrules (always; small files; cheap to read).
The repo’s CI config (e.g. .github/workflows/, .gitlab-ci.yml) — to know which lint / typecheck / test commands the reviewer will run.
The repo’s existing PR template (.github/pull_request_template.md) — to surface anything the user will need to fill in when they open the PR (e.g. “test plan”, “screenshot”).