review-change
Review a change with only applicable internal axes, classify every finding, persist fix-now work, and return one evidence-backed decision. Findings only; --adversarial N uses isolated reviewers; --synthesize fuses supplied reviewer tables. Triggers: "review-change", "review this change", "adversarial review".
How do I install this agent skill?
npx skills add https://github.com/gtrabanco/agentic-workflow --skill review-changeIs this agent skill safe to install?
- Gen Agent Trust Hubpass
The skill orchestrates automated code reviews by running local utility scripts and spawning isolated subagents. While the design includes isolation measures, it inherently processes untrusted code diffs, which creates a surface for indirect prompt injection.
- Socketpass
No alerts
- Snykpass
Risk: LOW · No issues
What does this agent skill do?
Review Change
Quality gate: run every applicable review and skip the rest, then synthesize and classify one report. Findings only; never edit or refactor.
Turn contract
Load and verify the canonical Turn contract (11 boxes) before ending every turn. Skill-specific additions (receipt closeout, isolation rule, applicability) live here and in REVIEW_PROCESS.md. Missing reference → STOP.
For a final PR review, the turn is incomplete until this additional box passes:
✓ Decision: REVIEW-PASS + PR exists →
`bun scripts/review-receipt.mjs emit --pr <N> --head "$(git rev-parse HEAD)" …`
RUN and exited 0 — the script posts the SHA-bound comment, re-reads the PR
comments, and exits non-zero unless the newest exact-HEAD
`review-change:pass` marker is confirmed, so its exit code IS this box
When the receipt emitter runs, it also records the affecting-path scope
manifest (fix/286) so a foreign-only commit after the review (a session-log or
toolstate commit on a shared checkout) does not void the receipt at the merge
gate — pass --scope-base main (the unit's base ref) and the runtime derives
the manifest itself; never hand-assemble a digest. A receipt without a scope
manifest stays head-bound (legacy semantics).
The receipt closeout is a precondition of the report, not a follow-up: do not
print the fixed report block on a non-zero exit. A clean report without a
current receipt must not recommend /audit-pr.
For REVIEW-FAIL or NEEDS-DECISION, list every open finding ID in the closing
recommendation, joined with +; the review must never hand off only the first
finding.
Review-end turn boundary
On REVIEW-FAIL or NEEDS-DECISION, this skill ends at the report: it
never invokes /fold-findings, /execute-phase, /unit-lane or
/triage-issue — those are separate user-initiated invocations,
run in a fresh turn (a programmatic outer driver may compose the same
sequence in-turn, per REVIEW_PROCESS.md). The review's only mutations are the
ledger commit (persist step 11) and, on REVIEW-PASS with a PR, the
receipt comment (persist step 12); it never folds a finding, executes a phase,
or edits code. A REVIEW-FAIL/NEEDS-DECISION turn closes when the report
block is printed (report step 13).
Consume the internal verification contract;
the reviewer checks the same frozen acceptance artifact (legacy (pre-feature 61) ACCEPTANCE.md
or post-feature-61 unit doc ## Acceptance criteria) as the executor before
mapping candidate evidence.
When to use
- Mandatory before every merge. Review in a context that did not implement
the change; if it did, stop and hand off to a fresh one.
execute-phasemay also recommend optional checkpoints at layer boundary, accumulation or sensitivity (#77). - Use when you need applicable reviews without irrelevant passes.
Scope
Default target is the current change (branch diff vs the default branch); accept a path/glob to widen or narrow. State the scope at the top of the report.
Review charter — dismantle the claim
The workflow's job is to demonstrate that what works, works — and works well: programmatically, engineering-wise, decision-wise, and against user expectation. This review exists to try to dismantle that claim; if it survives the attempt, that survival is the review outcome. So:
- In-scope defects are this review's job — never “out of scope”. A bug, a
security issue, or a guard skip inside the current feature's scope is a
review-changefinding: a feature is not complete while it has one. The reward is not delivery, it is meeting the acceptance criteria, adequate engineering without over-engineering, meeting other features' requirements (a page that needs auth verifies the session and the permission; a change to another table that a requirement depends on stays in scope even if it “looks” out of scope), and above all the user's expectation. - Ask, don't infer — the whole expectation surface. Where the plan, the diff or the requirement leaves the user's expectation unclear, surface it as a question in non-technical terms (a technical user asks for an app; a non-technical user asks for “an app to measure my heart rate”) — never as an assumed intent. This extends ask-don't-infer from acceptance criteria to every expectation the change is judged against.
- Repo-wide hunting is not this review. Open-ended bug hunting, security
research and broken-version investigation across the repository belong to
product-audit, where the sweep is scoped and budgeted; this review stays on the change at hand.
Step 0 — Discover the project & the change (always first)
Per Workflow conventions + documentation map, decide axes from:
- Project nature: UI (
docs/frontend/), web/mobile/CLI/library/backend, and optional recorded platform skills (extras only). - Footprint: what the diff touches (UI, API, infra, domain). An axis applies only when both project and footprint support it.
- Prior findings (optional retrieval aid, never authoritative): when the
index entry point exists,
agentic-workflow doc --query "<change topic>" --json-onlymay surface earlier findings, decisions and session-log entries on the same surface before the passes run. A missing index, an absent entry point, or an empty answer changes nothing — grep stays the exhaustive fallback, and the index is never a decision authority.
Applicability matrix (default; the project's docs refine it)
Every axis maps to a skill of the workflow's own internal review pack
(skills/review-* — installed with the workflow, so none can be missing):
| Axis — internal pack skill | Web | Mobile | Console/CLI | Lib/SDK | Backend/Infra |
|---|---|---|---|---|---|
review-code (correctness, simplification, dead code, duplication, arch) | ✓ | ✓ | ✓ | ✓ | ✓ |
review-security | ✓ | ✓ | ✓ | ✓ | ✓ |
review-verify (run it, confirm real behavior, tests) | ✓ | ✓ | ✓ | ✓ | ✓ |
review-design (UI/UX) | ✓ | ✓ | TUI only | ✗ | ✗ |
review-a11y | ✓ | ✓ | rare | ✗ | ✗ |
review-brand (voice/copy) | ✓ | ✓ | output text | ✗ | ✗ |
review-perf | ✓ | ✓ | ✓ | ✓ | ✓ |
review-seo | ✓ | ✗ | ✗ | ✗ | ✗ |
| API ergonomics / usage docs (inline pass) | if API | if API | flags/help | ✓✓ | ✓ |
review-implementation(the single classifier over the synthesized table — process step 8) andreview-debt(the debt transform over the classified table — process step 9) are not axis finders: they run once per review, not per axis.
Isolation rule (default — every pass, not only adversarial)
Each applicable pass, the verification pass (step 6), the classifier (step 8), and debt transform (step 9) runs
isolated/context-clean and returns only its fixed findings table plus
PASS | FAIL—never diff or prose. Spawn one fresh subagent/headless run per
pass; without those, use a fresh conversation. Pass only scope, its checklist,
and Step 0 docs; cap full non-diff reads at 10 (targeted ≤50-line reads/greps
excluded). The orchestrator retains tables, not sources, and fuses them in step
7. Inline fallback is sequential table reduction. A pass runs at its own tier or
stronger, never weaker.
Progressive loading — choose one review route
After applicability and isolation are established:
The reference allowlist is exactly the seven linked paths below. Never invent or
read another references/ path.
| Invocation route | LOAD in this order | SKIP |
|---|---|---|
| Default review | review process → adversarial recommendation → persist and decide → output and guardrails | synthesis, portability, adversarial setup |
--adversarial N | review process → adversarial setup before reviewers → adversarial synthesis before fusion → persist/decide → output/guardrails | portability |
--synthesize | review process → adversarial synthesis plus the supplied reviewer tables → persist/decide → output/guardrails | adversarial setup, portability |
legacy --merge | print the fixed migration refusal below and stop — zero git/forge mutation | everything |
Legacy --merge is removed — not an alias. Calling /review-change --merge prints this fixed migration refusal and stops before any git or forge mutation command runs:
migration: --merge is removed. Table fusion is --synthesize: pass the fixed
reviewer tables the same way and the synthesis contract fuses them. No
repository merge is performed by this skill.
Active review paths use only --synthesize/fusion language. Add
portability only when contexts, parallelism, slash
commands or tier controls are unavailable. docs/workflow/REPOSITORY_STATE.md
is evidence for output/guardrails, not a skill reference; that route owns NRS and
Architectural invariants rules.
Resources are one hop from this file. Fixed reviewer/synthesis/output contracts are literal. Missing required resource → stop; never approximate a review contract.
Portability
Keep reviewer contexts isolated. Use portability for sequential/headless fallbacks; never collapse independent adversarial passes.
Relationship to other skills
Orchestrates internal finders (review-code, review-security, review-verify,
review-design, review-a11y, review-brand, review-perf, review-seo), then
one review-implementation classifier and review-debt transform, isolated by
default; installed platform packs are optional. triage-issue is user-invoked,
never auto-run (D3), for independent proposals, audit findings, and
--prioritize-now runs. It is Stage 4: checkpoint reviews are
optional, the end review is mandatory and fresh. fix-now is routed to the
unit's fold cycle — a separate /fold-findings invocation, never run by this review;
replan-in-unit adds user-confirmed phases, and independent work becomes
proposals. audit-pr consumes
only the verified PR-comment receipt, never the
chat report; product-audit is the periodic sweep. On REVIEW-FAIL the manual
correction path is /fold-findings, then re-run /review-change on the changed
HEAD (bounded at two cycles — a third cycle never starts without an explicit
user instruction, per REVIEW_PROCESS.md); a programmatic outer driver may run
the same review→fold sequence.
Done when
- One synthesized/classified table covers every applicable axis, lists skipped
axes with reasons, and includes manual checks. Every finding has a destination:
fold, confirmed replan phase, surfaced decision, or user-routed proposal; none
is silently lost and review creates no backlog (D3). Print the closing
→ Next:block (clean →/audit-pr; recurring drift →/product-audit) and change no code.
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/gtrabanco/agentic-workflow/review-change">View review-change on skillZs</a>