Vet

Skill brolag/neural-claude-code/skills/vet

Independently review a working tree, branch, pull request, or implemented plan before ship. Use for a material pre-merge gate that checks acceptance criteria, regressions, security boundaries, project consistency, cleanup quality, and active counterexamples in fresh context. Returns SHIP or HOLD with evidence. Don't use for trivial changes (<15 LOC) or docs-only edits.From its SKILL.md

Install

npx -y skills add brolag/neural-claude-code --skill vet

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

11 stars11 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

5.8 KB, as published. Nobody here has run it

Vet

Review the change as an evaluator with no stake in shipping it. Report material findings only and return exactly SHIP or HOLD.

This skill is the Claude Code port of the Neural Codex vet gate. Same contract, slash invocation. The default independent reviewer is a fresh Agent. Codex is optional.

1. Resolve the target

Accept:

--base <ref> for a branch diff
--scope working-tree|branch|pr
--pr <number> and optional repository
--spec <path> for acceptance criteria
trailing text as a review focus

Inspect status, branch, remotes, and the narrow project guidance relevant to touched files. If no diff or changed artifact exists, report that there is nothing to vet and stop.

BRANCH=$(git branch --show-current)
if [ "$BRANCH" != "main" ] && [ "$BRANCH" != "master" ]; then
  git diff --stat main...HEAD && git diff main...HEAD
else
  git diff --stat && git diff
fi

Honor --base / --scope if given.

2. Build a neutral review bundle

Create a temporary directory outside the repository and collect:

objective and acceptance criteria from the plan
base/head refs and target label
status, tracked diff, staged diff, and diff stat
an untracked-manifest.txt generated from git ls-files --others --exclude-standard -z
tests already reported by the implementer
relevant project constraints
no authoring narrative, praise, or proposed verdict

Input hygiene (non-negotiable). Pass the bundle and a neutral focus string. Do not prepend a narrative that explains, motivates, or justifies the change.

For every untracked status entry, resolve it beneath the repository and account for it in the manifest. Copy safe text files into an untracked/ subtree that preserves relative paths. For directories, inventory their files recursively without following symlinks. For large, binary, ignored, sensitive, or excluded files, record path, size, hash when safe, and the reason content was omitted. The reviewer must compare the manifest with status before issuing a verdict. An unexplained missing path prevents SHIP.

Treat refs, paths, and user text as data. Quote shell arguments, reject traversal outside the expected repository, and never include .env contents, credentials, or unrelated sensitive output.

3. Separate the evaluator

Launch a fresh Agent as the independent reviewer. It receives ONLY the neutral bundle. Never the author's session or reasoning.

Agent(subagent_type: "general-purpose", prompt: "<review contract> + <neutral bundle>")

If a second model is available (Codex CLI installed and logged in), you may also run the same bundle through codex exec for diversity. It is optional. The fresh Agent is the default gate.

codex exec -c model_reasoning_effort="high" "$(cat <<'PROMPT'
<the same review contract>
PROMPT
)" < /dev/null

The < /dev/null is required: codex exec reads stdin even when a prompt argument is given.

If genuine context separation is unavailable, perform the strictest possible local review, disclose that limitation, and return HOLD whenever independence is required by the plan or material uncertainty remains. Never present self-review as independent review.

4. Verify acceptance

Extract every safe machine-checkable command, required file, behavior, migration, policy, must_include, and must_not_include assertion from the plan.

Mark each criterion:

PASS: executed or inspected with concrete evidence
FAIL: contradicted by a command, diff, or counterexample
NOT RUN: unsafe, unavailable, or outside the current environment

A required FAIL or NOT RUN prevents SHIP.

5. Review adversarially

Prioritize failures with real impact:

authorization, trust boundaries, and secret exposure
data loss, corruption, duplication, rollback, and partial failure
races, stale state, ordering, retries, and idempotency
empty, null, timeout, degraded dependency, and version-skew behavior
compatibility, installer, migration, and documentation drift
dead code, generated slop, swallowed errors, unsupported claims, missing tests, and copy-paste that should reuse an existing symbol

Construct distinguishing inputs or safe probes for testable behavior. A green suite is evidence, not the whole specification. Where a runnable check exists, run it.

6. Report the verdict

Use this shape:

## Vet Review

Target: <target>
Reviewer: <fresh Agent | Codex (optional) | disclosed limitation>

### Verdict: SHIP | HOLD

### Acceptance
| Criterion | Result | Evidence |

### Findings
#### CRITICAL | HIGH | MEDIUM | SLOP: <title>
- File: path:line
- What can go wrong:
- Evidence:
- Impact:
- Fix:

### Probes
### Residual risk
### Summary

Return SHIP only when no material finding remains and every required criterion passes. Return HOLD for addressable findings, missing required evidence, failed criteria, or an invalid review boundary. LOW/SLOP findings are listed but do not block SHIP.

Usage Examples

/vet --base origin/main --spec plans/2026-07-12-hooks/plan.md
/vet --scope working-tree focus on installer compatibility
/vet --pr 42

Done when

The change has been challenged in fresh context, acceptance is evidence-backed, material findings are actionable, and the verdict is unambiguous.