/review

name: review
description: "Review Elixir/Phoenix changes before committing — bugs, security, Ecto, LiveView, and Oban anti-patterns, via parallel specialist agents. Use when asked to review changes, a component, a diff, or a PR."
effort: high
argument-hint: "[test|security|oban|deploy|iron-laws|all]"

Review Elixir/Phoenix Code

Review code by spawning parallel specialist agents. Find and explain issues — do NOT create tasks or fix anything.

Usage

/phx:review                          # Auto-detects task ID from branch/commits
/phx:review test                     # Review test files only
/phx:review security                 # Run security audit only
/phx:review oban                     # Review Oban workers only
/phx:review deploy                   # Validate deployment config
/phx:review iron-laws                # Check Iron Law violations only
/phx:review ENA-8931                 # Force Linear issue
/phx:review #42                      # Force GitHub issue
/phx:review .claude/plans/auth/plan.md    # Force plan / spec file
/phx:review --no-requirements        # Skip requirements coverage check
/phx:review --codex                  # Add Codex CLI as cross-model reviewer

Arguments

`$ARGUMENTS` = Focus area, task ID, or path to plan/spec file.

When no requirements argument is passed, the skill auto-detects a task ID from the branch and recent commits (see `${CLAUDE_SKILL_DIR}/references/requirements-detection.md`).

Workflow

Step 1: Identify Changed Files and Prepare Directories

Create output dirs before spawning agents — agents cannot create directories, so their writes would fail.

1. Determine SLUG via Glob on `.claude/plans/*/` (default: `"review"`) 2. Run `mkdir -p ".claude/plans/${SLUG}/reviews" ".claude/plans/${SLUG}/summaries" .claude/reviews` 3. Run `git diff --name-only HEAD~5` and `git diff --name-only main` 4. Save the diff base for pre-existing detection in Step 3b

Step 1b: Load Plan Context and Prior Reviews

• Read `.claude/plans/${SLUG}/scratchpad.md` for planning decisions and rationale
• Pass relevant decisions to agents as WHY-context (eliminates session archaeology)
• Check `.claude/plans/${SLUG}/reviews/` for prior output; if present, include a

consolidated summary as "PRIOR FINDINGS" with: "Focus on NEW issues. Mark still-present issues as PERSISTENT."

Step 1c: Detect Requirements Source (skip on `--no-requirements`)

Find a task/spec whose requirements should be cross-checked against the diff. Priority order (stop at first match): explicit arg → conversation context → branch regex → commit subjects → latest plan → none. Full table, regexes, and fetch mapping in `${CLAUDE_SKILL_DIR}/references/requirements-detection.md`.

Fetch the detected source into `.claude/plans/${SLUG}/reviews/.requirements-input.md` (Linear via `mcp__linear__get_issue`, GitHub via `gh issue view`, file via Read). Record `REQ_SOURCE` label (e.g. `"Linear ENA-8931"`) for the verifier heading. On fetch failure, set `SOURCE_STATUS=FETCH_FAILED` and continue — verifier will emit `NOT AVAILABLE` rather than block the review.

Step 2: Spawn Review Agents (MANDATORY)

Spawn each agent role once per review. Code analysis belongs to the spawned agents, not this context — a review that spawns zero agents has failed.

1. If `TaskCreate` is in your tool list (never ToolSearch for it), create a task per agent, set `in_progress` 2. For `/phx:review` or `/phx:review all`: select agents dynamically per the selection table in `${CLAUDE_SKILL_DIR}/references/agent-spawning.md` 3. For focused reviews (`test|security|oban|deploy|iron-laws`): spawn only the matching specialist from the focused mode table in the same reference 4. **If Step 1c succeeded** (REQ_SOURCE non-empty and `--no-requirements` not passed): add `phx:requirements-verifier` to the same parallel batch. Pass these prompt inputs: `REQUIREMENTS_TEXT` (content of `.requirements-input.md`), `REQUIREMENTS_SOURCE` (REQ_SOURCE label), `DIFF_FILES` (git diff --name-only output), `SOURCE_STATUS` (only if FETCH_FAILED), `output_file: .claude/plans/{slug}/reviews/requirements.md` 5. Spawn in ONE message with `run_in_background: true`. Do not pass the deprecated Agent `mode` parameter; Claude Code 2.1.212+ ignores it and subagents inherit the parent session's permission mode 6. Pass an explicit `output_file` per agent (mapping in the reference) 7. Include the CRITICAL prompt block: write by turn ~12, chat body ≤300 words 8. Scope every agent to the diff: pass `git diff --name-only` output with "Focus on NEW code. Pre-existing: one-line `{file}:{line} — {brief}`. Do NOT deep-analyze unchanged files." 9. **With `--codex`**: add `phx:codex-reviewer` to the same batch (prompt template in agent-spawning.md). Missing CLI degrades to SKIPPED.

Step 3: Collect and Compress Findings

Wait for ALL agents to complete before writing the verdict — a partial panel gives a misleading review. Count one completion notification per agent spawned; mark any task `completed`.

**Missing file fallback** — after each agent finishes, verify its expected `output_file` exists. If missing (turn exhaustion, error):

1. Append to `.claude/plans/{slug}/scratchpad.md`: `[HH:MM] WARN: {agent} did not write {expected_path} — extracting from message` 2. Parse findings from the agent's return message as fallback 3. Mark the section in the final review with `⚠️ EXTRACTED FROM AGENT MESSAGE (see scratchpad)` — never silent

**Verification-runner fallback** — if it times out, run directly: `mix compile --warnings-as-errors && mix format --check-formatted $(git diff --name-only HEAD~5 | grep '\.exs\?$' | tr '\n' ' ') && mix credo --strict && mix test`

**Context supervision** — for 4+ agents, spawn `phx:context-supervisor`:

Prompt: "Compress review agent output.
  input_dir: .claude/plans/{slug}/reviews
  output_dir: .claude/plans/{slug}/summaries
  output_file: review-consolidated.md
  priorit