MANDATORY: Follow the complete shared lifecycle contract in ${CLAUDE_PLUGIN_ROOT}/skills/shared/run-lifecycle.md with declared skill review.

Code Review Skill

MANDATORY: READ ENTIRE FILE before composing user-facing prose: ${CLAUDE_PLUGIN_ROOT}/skills/shared/readability-style.md.

Thin wrapper around "${CLAUDE_PLUGIN_ROOT}/scripts/larch.sh" review core. It owns flag parsing, session setup, the outer diff-mode round loop, fix application, final summaries/issues, run logging, and cleanup. review core owns one gather→dispatch→collect→aggregate→vote→emit round (review aggregate-findings may no-op when disabled, when fewer than two findings are present, or when merge/dispatch/validation fails and leaves findings.md unchanged).

Anti-halt continuation reminder. After every child Skill tool call (e.g., /design, /review, /release, /issue, /implement) returns AND after every Bash tool call that completes a numbered step or sub-step, including scripts/larch.sh checks run-relevant, IMMEDIATELY continue with this skill's NEXT numbered step — do NOT end the turn on the child's cleanup output, on a Bash result, or on a status message, and do NOT write a summary, handoff, status recap, or "returning to parent" message — those are halts in disguise. This applies to ALL step boundaries from Step 0 through Step 5, and to ALL sub-step transitions within Step 3's review loop (3a→3b→3c→3d→3e→3f→loop back to Step 1). Critical: in diff mode, the review loop (Steps 1→2→3) repeats until convergence (0 findings, or Step 3f classifies the just-fixed round as non-substantial — a main-agent classification of accepted-and-fixed work, not a reading of reviewer prose) or the fixed cap of 2 — completing one round's substantial fixes does NOT mean the review is done. → shared/subskill-invocation.md#anti-halt Continue after child returns. Treat every script and child-skill result as input to the next step, not as a stopping point.

Parse flags from $ARGUMENTS: a nested call may begin with exactly one internal --lifecycle-parent-context <absolute-context-path> pair; bind it to LIFECYCLE_PARENT_CONTEXT and remove it before parsing --diff, --dynamic-archetypes <N>, --session-env <path>, --step-prefix <prefix>, --subagent, --difficulty <TRIVIAL|MODERATE|HARD>, and --run-id <ID>. Read ${CLAUDE_PLUGIN_ROOT}/skills/shared/run-id-flag.md only for shared --run-id flag semantics. Flags may appear in any order before the positional description; --diff and positional description are the two mode activators and are mutually exclusive. --dynamic-archetypes must be 0..1; when absent, LARCH_DYNAMIC_ARCHETYPES_MAX may supply the same range, default 0.

Mode activation is fail-closed: if --diff and positional description are both present, print **⚠ --diff cannot be combined with a description. Use --diff alone for branch diff review, or provide a description without --diff. Aborting.** and exit. If neither is present, print **⚠ /review requires either --diff (branch diff review) or a description of what to review. Examples: /review --diff, /review implementation of auth module, /review error handling in scripts/. Aborting.** and exit.

Progress and prompt pins: read step-name-registry.tsv; reviewer prompts preserve code-quality / risk-integration / correctness / architecture / security; specialist prompts are rendered through ${CLAUDE_PLUGIN_ROOT}/scripts/larch.sh render specialist; description mode preserves the ### In-Scope Findings / ### Out-of-Scope Observations dual-list contract hints.

Commands: ${CLAUDE_PLUGIN_ROOT}/scripts/larch.sh review core and review compose-findings are Rust-owned through the verified bootstrap. Rust: scripts/larch.sh review tally-code-votes, scripts/larch.sh review emit-tally, scripts/larch.sh review log-phase, scripts/larch.sh review aggregate-findings, scripts/larch.sh review prune-nit-findings, scripts/larch.sh review reviewer-prune, scripts/larch.sh review gather-context, scripts/larch.sh review dispatch-panel, scripts/larch.sh review collect-findings, scripts/larch.sh review check-reviewer-failure-threshold, scripts/larch.sh agent dispatch-voters, scripts/larch.sh agent dispatch-waterfall, and scripts/larch.sh agent launch-claude-subprocess. Harnesses: crates/larch-cli/tests/review_commands.rs, crates/larch-cli/tests/review_tally_commands.rs, crates/larch-cli/tests/review_dispatch_panel.rs, crates/larch-cli/tests/voter_dispatch_commands.rs, and crates/larch-cli/tests/waterfall_commands.rs.

review reviewer-prune, reached through ${CLAUDE_PLUGIN_ROOT}/scripts/larch.sh, owns prune-decision status and env writing.