Review
Skill stevennitesh/programming-agent-skills/skills/custom/review
Practical workflow skills for coding agents: scope, debugging, cleanup, subagents, GitHub tracking, and verification.
npx -y skills add stevennitesh/programming-agent-skills --skill reviewAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 0 stars0 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.
What its author says it does
Copied from the file, not written here
Review one ordinary branch, WIP, staged, or "review since X" diff read-only from a fixed snapshot. Judge Standards ("built right?") and Spec ("right thing?") separately, then return one terminal report. Hand off local PRs or high-risk local diffs once to $convergent-pr-review.
SKILL.md
5.3 KB, ~1.1k tokens by cl100k_base, as published. Nobody here has run it
Review
Pin. Trace. Admit. Judge. Return.
Own one ordinary, read-only fixed-snapshot review.
Terminal: judge one immutable snapshot and return. Start no remediation, dispatch, tracker mutation, successor review, or snapshot recapture. The report grants no mutation or new-snapshot authority.
Read-only: Inspect only. Leave the worktree, index, commits, tracker, PR comments or reviews, and external messages unchanged.
- Standards: built right, against repository rules, meaningful nearby conventions, and the fallback smell baseline when standards are thin.
- Spec: right thing, against the authoritative request, bounded slice, acceptance criteria, and required proof.
Lens reset: keep the axes separate. Never merge, deduplicate, or rerank findings across them.
Review route: For a local PR or high-risk local diff, hand off once to $convergent-pr-review and stop. Carry the caller-supplied Charter, review mode, Spec requirement, Source Trace, fixed point, target inputs, and any carried finding IDs; $convergent-pr-review owns snapshot capture. Do not resume this skill when it returns.
Spec gate: The caller supplies Spec required: yes | no. Default to no for standalone review. When Spec required: yes, a missing, conflicting, or unresolved authoritative source makes the review incomplete; never skip or replace the Spec axis.
An incomplete review returns:
Review status: incomplete
Review target: <resolved target or unresolved>
Sources: Standards: <sources or not traced>. Spec: <source or unresolved>.
Blocker: <unresolved ref | empty or incomplete target | missing or conflicting Spec>
Skipped: judgment did not run.
Return boundary: caller
Mutation authority: none
Successor snapshot authority: none
1. Pin
Resolve the supplied fixed point to a commit SHA. Otherwise discover the repository default branch and merge base, record the resolved commit, and ask only when discovery is ambiguous.
Select one target:
- review tree:
git diff <fixed-point> <review-tree> - committed branch:
git diff <fixed-point> HEAD - working tree:
git diff <fixed-point> - explicitly staged only:
git diff --cached
A supplied review tree wins over every live target. Verify <review-tree>^{tree} and inspect its files through git show. For staged-only review, record the fixed point but review only the cached diff.
Capture the resolved fixed point, selected diff command, git status --short, and relevant commits. Read in-scope untracked files directly for working-tree review. Record Review mode: initial | remediation; default to initial. Remediation mode requires the original Charter, prior snapshot, carried finding IDs, and repair delta.
Return incomplete when a ref does not resolve, the complete target cannot be captured, or the target is empty.
2. Trace
Build one Source Trace before judgment.
Trace Spec in order:
- Current request or caller-supplied item, slice, criteria, path, issue, URL, spec, or legacy PRD.
- Issue or PR references in captured commits, following
docs/agents/issue-tracker.mdwhen available. - A matching source under repository conventions such as
docs/,specs/, or.scratch/.
When Spec is optional and absent, report Skipped: no spec available.
Trace Standards from AGENTS.md and its pointers, contributor or coding guidance, test and tooling configuration, and meaningful nearby conventions. When these are thin, read SMELL-BASELINE.md. Repository standards override it; label each smell a baseline judgement call; leave tooling-enforced style to tooling.
3. Admit
Read FINDING-CONTRACT.md. Apply its Anchor, Reach, Evidence, Impact, and Proportion gates before severity. Preserve the caller's Charter; when none is supplied for standalone review, use the requested slice and resolved Source Trace as the boundary.
4. Judge
Run Standards -> lens reset -> Spec.
Report only admitted findings through the shared finding interface. Standards findings also say documented-standard breach or baseline judgement call.
Judge Standards from its traced sources, then reset to the pinned target and Spec source. Judge Spec for omissions, incorrect behavior, scope creep, and proof gaps through useful interfaces or seams.
5. Return
Use exactly:
Review status: complete
Review target: <resolved fixed point> | <diff command> | <review tree or live target>
Sources: Standards: <sources>. Spec: <source or skipped>.
## Standards
<findings or "No findings.">
## Spec
<findings or "No findings." / "Skipped: no spec available.">
Summary: Standards: <count>, worst <severity or none>. Spec: <count/skipped>, worst <severity or none>.
Return boundary: caller
Mutation authority: none
Successor snapshot authority: none
Complete: Pin and Trace pass; every reported finding passes Admit; both applicable axes run with a lens reset; the report satisfies the finding contract; and control returns without mutation or successor review.
What ships with it: 4 files
6.3 KB alongside SKILL.md
agents/
- openai.yaml317 B
- ADVISORY-CONTRACT.md1.0 KB
- FINDING-CONTRACT.md3.0 KB
- SMELL-BASELINE.md2.0 KB