# Review Anvil Readonly

> Multi-agent code review without any edits or commits — requested rounds of parallel reviewers, synthesis, default-on reproduction of uncertain material findings, and a final report. Use when the user wants a code review of a diff, branch, path, or PR but does NOT want any fixes applied automatically (no edits, no commits, no staging). Activates the `review-anvil` engine in read-only mode.

- Skill: `mrshu/review-anvil-readonly` (Agent Skill)
- Install (CLI): `npx skillmds@latest add mrshu/review-anvil-readonly`
- Raw SKILL.md: https://api.skillmd.com/api/skills/mrshu/review-anvil-readonly/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: mrshu (https://skillmd.com/u/mrshu)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/mrshu/review-anvil-readonly

---


# review-anvil-readonly

Thin preset that runs the multi-agent review loop without applying any fixes. Activates the [`review-anvil`](../review-anvil/SKILL.md) engine skill with `commit_mode=none` and a default of `rounds=1`.

Use the ASD-STE100-inspired language contract in
`../review-anvil/references/asd-ste100-inspired.md` for generated prose.

## How to invoke

The agent reads the user's request, then activates `review-anvil` with this argument string:

```
commit_mode: none, <pass-through any args the user specified>, rounds: 1
```

**Assembly order matters.** The engine parses left-to-right with first-occurrence-wins precedence. Pinned values (`commit_mode: none`) go *before* user args so the pin is authoritative. Defaults the user can override (`rounds: 1`) go *after* user args so any user-supplied `rounds: N` wins.

**Pin-rejection (defense in depth).** Pins for this preset: `commit_mode`. If the `review-anvil-pr` helper script is resolvable (this preset ships no script of its own), enforce mechanically: `bash <pr-helper-path> check-pins review-anvil-readonly "commit_mode" "$ARGUMENTS"` — non-zero exit means abort with the script's error verbatim. Otherwise apply the engine's prose pin-rejection algorithm ("Parsing" section); the abort-message preset name is `review-anvil-readonly`.

Pass-through args the user may specify (non-exhaustive — any engine param not pinned here works; each is documented in the engine's SKILL.md):

- `target: PR #42` / `target: src/auth/` / `target: branch` / `target: uncommitted` — what to review
- `focus: <topic>` — narrow the four-pillar focus, or `only: <topic>` to replace it
- `rounds: N` — the user's value wins over the default `rounds: 1`
- `agents: 2 codex + 1 claude` — custom reviewer mix
- `min_fix_severity: <sev>` — drives the would-apply/suggestions split in the read-only report
- `reproduction: auto|on|off` — default-on confidence gate for uncertain material findings before they become actionable report items
- `proof_runner: /absolute/trusted/path` — isolated executable proof runner; when unset, generated probes are retained but behavior findings that need execution remain deferred
- `adversarial: auto|challenge|targeted|full|strict` — optional adversarial review that attacks false positives and disproportionate/bloated fix plans before the final report
- `adversarial_rounds: 1|2`, `disagreement_policy: defer|comment` — tune the adversarial gate; it remains read-only and bounded
- `reviewer_timeout: <seconds>`, `report_path: <file>` — as in the engine

After the engine completes, surface the synthesized report inline. **Do not** follow with edits, commits, or any side effects — that's exactly what `commit_mode=none` rules out.

Read-only review still means no tracked edits, staging, commits, or pushes.
Temporary prompt/reviewer/report artifacts under the artifact-local,
self-ignored `.review-anvil/` directory are allowed. Retained proof bundles
live in a host-created private directory outside the reviewed worktree and Git
common directory; proof handling never edits tracked files or Git ignore
metadata. A configured trusted `proof_runner` executes generated probes only
against an exact disposable snapshot with network disabled, source and proof
inputs read-only, filesystem reads restricted to source/proof/runtime inputs,
writes limited to the proof runtime directory, runner-authored results, a
sanitized environment, and bounded resources. The engine never executes a
repository-provided runner or falls back to direct probe execution.
Adversarial review remains read-only and may drop false positives, defer
harmful or tech-debt-heavy fix plans, harden fix prose, or strip unsafe
suggestion blocks.

Default local policy: leave adversarial review off for ordinary fast/local
reviews. Append `adversarial: auto` when the user asks for a careful,
skeptical, high-confidence, low-noise, thorough, or production-ready read-only
review. Do not append it when the user asks for a quick/rough/sanity pass unless
they explicitly request adversarial review.

## When not to use this

- If the user wants fixes applied and committed, activate `review-anvil` directly with `commit_mode=per_fix` (or just leave it default).
- If the user wants the report posted back to a GitHub PR comment, activate `review-anvil-pr` instead.

## Engine reference

See [`../review-anvil/SKILL.md`](../review-anvil/SKILL.md) for: full parameter surface, mix policy, focus options, reviewer prompt template, and final report format. This preset is a one-line argument addition; all behavior comes from the engine.