| name | review-changes |
| description | Review the working-tree changes (vs main) for correctness, concurrency, architecture, testing, and API/doc issues — following .github/CODE_REVIEW.md — and return a severity-graded report. Scales the machinery to the diff size: a single code-reviewer agent for a small change, or a fan-out-and-verify Workflow for a large/multi-unit one. Produces findings; it does not apply fixes (the caller does). |
Review Changes
Produce the pre-PR review of the current branch's changes against main,
following the canonical .github/CODE_REVIEW.md. This skill scales the review
to the size of the change — a single agent when the diff is small, a parallel
fan-out with adversarial verification when it's large — and returns a
severity-graded report. It does not edit code; the caller (e.g. /deliver's
code-review phase, or you) applies the fixes.
Principles
- One source of truth. Every reviewer — single or fanned-out — follows
.github/CODE_REVIEW.md: the severity rubric, scope, and the mandatory
adversarial re-evaluation. Don't invent criteria.
- Scale to the change. Don't pay for a fan-out Workflow on a 60-line diff;
don't review a 12-file multi-model change through one over-stretched context.
- Findings only. Return the report; the caller fixes (test-first) and may
re-invoke to converge.
- Adversarial verification cuts noise. On the large path, every Critical/High
finding is independently refuted before it's reported — the strongest lever
against false positives.
0. Gate — skip if there is no reviewable code
Code review exists to review code — Swift source, plus the committed
scripts under .claude/workflows/ (decision infrastructure) and Scripts/
(enforcement infrastructure: check-defaulted-witnesses.py runs from make lint and as its own CI Lint step, so a bug there silently weakens a
gate — it shipped as a false green once already, passing on an empty scan).
If the diff touches none of them, there is nothing to review — return
immediately, don't spawn any reviewer.
git diff --name-only origin/main...HEAD \
| grep -qE '\.swift$|^\.claude/workflows/|^Scripts/' \
|| echo "no-reviewable-code"
If that prints no-reviewable-code (e.g. a docs-only, config-only, or
skills-prose change), stop here and report "No reviewable code — code
review skipped." Do not run §1–§3. (JSON fixtures under
Tests/**/Resources/ accompany Swift changes; a fixture-only change with no
.swift is rare — note it and let the caller decide.)
Override — force-review in the arguments skips this gate entirely. The
caller has decided the change carries risk the file extensions don't show.
/deliver passes it for a reflexive delivery — one that rewrites the
skills, agents or review spec the pipeline itself runs, where the diff is
markdown but the blast radius is the pipeline (#407 shipped three defects
exactly there). This override lifts the gate; it does not choose a path —
§1 still decides, and a caller-stated full weight still takes §2b (Risk
overrides size), however prose-shaped the diff. Whichever path runs, aim its
brief at the change's own subject matter: which rule changed, whether
anything still depends on the old wording, and whether the new rule is
enforced or merely stated. Without this the gate and the caller contradict
each other and the skip silently wins.
If the only reviewable change is a script (.claude/workflows/ or
Scripts/), take the §2a single-reviewer path with a script-focused brief
in place of the Swift lens: input guards and executable throws, schema
shapes, tally / dead-agent handling, and prompt-injection surface (what
untrusted text reaches an agent prompt). For a Scripts/ checker add the
question that matters most: can it pass while measuring nothing? — an empty
scan, a typo'd path, a count compared instead of an explicit set.
.github/CODE_REVIEW.md's severity rubric and adversarial re-evaluation still
apply; its Swift-specific checks don't.
1. Scope the change
git diff --stat origin/main...HEAD
Judge size (heuristic, not a rigid count):
- Small / single-unit — one cohesive area (a method + its tests + a model +
docs), a handful of source files, roughly under ~200 changed lines → §2a.
- Large / multi-unit — several new models/methods/services, or a broad diff →
§2b.
When unsure, prefer small — escalate to the fan-out only when the diff is
genuinely broad.
Risk overrides size. Take §2b however small the diff is when either the
caller states the change is full weight, or the diff touches a risky
surface — concurrency (actors, Sendable, locks, Task), networking /
HTTPClient, Decodable / CodingKeys, or new public API. Size decides how
much there is to read; the risk surface decides how many lenses have to read
it, and a cohesive one-file concurrency change is the shape whose defects most
reliably survive a single pass (#401, #433, #461 each had one found late, by a
second lens or a second platform). A caller that passes full has already made
this judgement from the plan — do not re-derive it from --stat and quietly
overrule it.
2a. Small change → single agent
Spawn the code-reviewer agent on git diff origin/main...HEAD, telling it
not to build or run tests (see Reviewers don't build below). It follows
.github/CODE_REVIEW.md including its own adversarial self-pass, and returns the
full severity-graded report. Pass that report back to the caller. Done.
Reviewers don't build
No reviewer — single or fanned-out — may run make, swift build or
swift test. Reviewing is a reading task: the reviewer has the diff and the
source, and the caller has already run every gate before invoking this skill.
This is not a style preference. Every target shares one SwiftPM scratch
directory per worktree, so N parallel reviewers each starting a build is N
processes contending on one .build/.lock — and if any of them builds docs, the
SWIFTCI_DOCC manifest flip invalidates the others' build plans, so they redo
each other's work in a cycle rather than merely queueing. Observed once as ~10
zsh pipelines pinned at 100% long enough that the user killed them (see
knowledge/gotchas.md → Docs builds need their own scratch path).
If a finding genuinely cannot be settled without executing something, say so in
the finding and let the caller run it — serially, once.
2b. Large change → fan-out + verify Workflow
Invoke the Workflow tool with the script below and args: { base: "origin/main" }.
It fans out one reviewer per dimension (each following the spec, focused on a
single lens), dedups, adversarially verifies every Critical/High finding with
an independent skeptic (dropping any that don't survive), and returns the
reconciled findings by severity.
export const meta = {
name: 'review-changes-fanout',
description: 'Fan-out multi-dimension code review with adversarial verification',
phases: [
{ title: 'Find', detail: 'one reviewer per dimension, in parallel' },
{ title: 'Verify', detail: 'adversarial skeptic per Critical/High finding' },
],
model: 'opus',
}
const BASE = (args && args.base) || 'origin/main'
const SPEC = '.github/CODE_REVIEW.md'
const DIMENSIONS = [
{ key: 'correctness', focus: 'correctness & safety — logic bugs, behavioural regressions, force unwraps / try!, and input validation at system boundaries' },
{ key: 'concurrency', focus: 'Swift 6 concurrency — async/await, actor isolation, Sendable conformance, no blanket @MainActor, and justified @preconcurrency / @unchecked Sendable' },
{ key: 'architecture', focus: 'architecture — protocol + TMDb-prefixed implementation, service-layer boundaries, new API exposed on TMDbClient with the service constructed in its private init (TMDbFactory vends shared plumbing only — services are not registered there), required model conformances, AND sibling-convention conformance: a newly-added member of an existing family (service method, model, guarded method) must match its siblings — same input validation, same error case, same conformance set / decode strategy — or the divergence is flagged' },
{ : , : },
{ : , : },
]
= {
: ,
: ,
: {
: { : },
: {
: ,
: {
: ,
: ,
: {
: { : , : [, , , ] },
: { : },
: { : , : },
: { : , : },
: { : , : },
: { : , : },
},
: [, , , , , ],
},
},
},
: [, ],
}
= {
: ,
: ,
: {
: { : },
: { : },
},
: [, ],
}
= () =>
+
+
+
+
+
()
perDim = (.(
((d), {
: , : ,
: , : , : , : ,
}).( ({
: d.,
: (r && .(r.)),
: r && r. ? r..( ({ ...f, : d. })) : [],
}))
))
dimResults = .( perDim[i] || { : d., : , : [] })
covered = dimResults.( r.).( r.)
missing = dimResults.( !r.).( r.)
(missing.) {
()
}
all = dimResults.( r.)
seen = ()
deduped = all.( {
key =
(seen.(key))
seen.(key)
})
blocking = deduped.( f. === || f. === )
advisory = deduped.( f. === || f. === )
()
verified = (blocking.(
(
+
+
,
{
: , : ,
: , : , : , : ,
}
).( (v && v. ? f : ))
))
confirmed = verified.()
{
: confirmed.( f. === ),
: confirmed.( f. === ),
: advisory.( f. === ),
: advisory.( f. === ),
: blocking. - confirmed.,
: covered,
: missing,
: missing. > ,
}
To iterate on the script, edit the file path the Workflow tool returns and
re-invoke with { scriptPath } rather than resending it.
If the fan-out is interrupted or dies part-way, don't discard the run. Each
completed agent's return value is recorded as a {"type":"result",…} line in
journal.jsonl inside the run's transcript directory (the Workflow tool
reports the path), so the findings that landed are recoverable — reduce those
and report as partial, naming what is missing. Never re-present a salvaged
run as complete.
3. Output
Return the report in the .github/CODE_REVIEW.md shape — Strengths, Issues
grouped Critical / High / Medium / Low (each with file:line, what's wrong, why,
fix), and an Assessment (Ready to merge? + reason). On the large path, also
note how many Critical/High findings were dropped by adversarial verification
(droppedByVerification) — that number is a feature, not a gap. The caller
decides what blocks and applies the fixes.
Findings you are not going to fix get filed, not dropped. A Medium or Low
that is real but out of scope for this change dies in the report unless someone
writes it down — and this skill produces findings precisely because the caller
will not act on all of them. File those per
.github/ISSUE_FILING.md, and cite the issue
number beside the finding, so the report shows what was captured rather than what
was noticed.
Do not file Critical/High findings — those block, and filing one reads as
handling it.
On the large path, coverage is mandatory in the report: list
dimensionsCovered, and when partial is true name every dimension in
dimensionsMissing and say plainly that the review is incomplete. A
large-path report without a coverage line is malformed — treat it as
untrustworthy rather than clean. A missing dimension means nobody looked,
which reads identically to nothing was found unless you say so; that is the
False green family (knowledge/gotchas.md), the defect this repo hits most.
(§2a's single-reviewer path has no dimensions, so this does not apply there.)
Arguments: $ARGUMENTS