vs-roast-code
Use when the user says roast, roast-code, or tear apart code, or asks for a pre-flight, a ready-to-merge check, reasons not to ship, or names a healthy point to diff against. Prefer over simplify for roast requests. Classifies the change first and scales review depth to it, adding a cross-model second opinion for risky or substantial diffs. Readiness-only asks stay here; they are not a publish.
- License 许可
- MIT
- 所属仓库
- vltansky/vs
- 最近更新
- 2026-11-06
- Stars
- ★ 0
- Forks
- 1
01 前言 Front Matter
- name
- vs-roast-code
- description
- Use when the user says roast, roast-code, or tear apart code, or asks for a pre-flight, a ready-to-merge check, reasons not to ship, or names a healthy point to diff against. Prefer over simplify for roast requests. Classifies the change first and scales review depth to it, adding a cross-model second opinion for risky or substantial diffs. Readiness-only asks stay here; they are not a publish.
02 资源文件 54
- 参考文档agents/openai.yaml160 B
- 参考文档references/comedy-techniques.md3.6 KB
- 参考文档references/sin-categories.md3.5 KB
- 参考文档references/structural-review.md3.4 KB
- 脚本scripts/reject-roast-buckets.mjs5.4 KB
- 脚本scripts/reject-roast-preflight.mjs9.1 KB
- 脚本test/buckets.static.eval.ts3.7 KB
- 脚本test/classification.static.eval.ts6.7 KB
- 参考文档test/fixtures/buckets/bad-all-act-nits/roast.md221 B
- 参考文档test/fixtures/buckets/bad-dismissed-as-act/roast.md271 B
- 参考文档test/fixtures/buckets/bad-missing-act/roast.md274 B
- 参考文档test/fixtures/buckets/bad-slogan-only-roast/roast.md174 B
- 参考文档test/fixtures/buckets/bad-unbucketed/roast.md197 B
- 参考文档test/fixtures/buckets/clean-act-only/roast.md178 B
- 参考文档test/fixtures/buckets/clean-mixed-buckets/roast.md369 B
- 参考文档test/fixtures/buckets/copied-slogans-skill.md163 B
- 参考文档test/fixtures/buckets/phrase-copy-skill.md751 B
- 参考文档test/fixtures/buckets/slogan-only-skill.md133 B
- 参考文档test/fixtures/buckets/structure-paste-skill.md248 B
- 参考文档test/fixtures/ponytail-bloat/package.json131 B
- 脚本test/fixtures/ponytail-bloat/src/bloated-name.js317 B
- 脚本test/fixtures/ponytail-bloat/src/direct-name.js99 B
- 参考文档test/fixtures/preflight/apply-fix-no-ask/preflight.md319 B
- 参考文档test/fixtures/preflight/auto-fix/preflight.md184 B
- 参考文档test/fixtures/preflight/bad-ref/preflight.md142 B
- 参考文档test/fixtures/preflight/clean-ship/preflight.md302 B
- 参考文档test/fixtures/preflight/copied-slogans-skill.md136 B
- 参考文档test/fixtures/preflight/do-not-ship-act/preflight.md244 B
- 参考文档test/fixtures/preflight/empty-diff-waiver/preflight.md199 B
- 参考文档test/fixtures/preflight/empty-diff/preflight.md191 B
- 参考文档test/fixtures/preflight/fix-it-no-ask/preflight.md321 B
- 参考文档test/fixtures/preflight/forged-pin/preflight.md165 B
- 参考文档test/fixtures/preflight/mutate-no-ask/preflight.md327 B
- 参考文档test/fixtures/preflight/phrase-copy-skill.md272 B
- 参考文档test/fixtures/preflight/publish-invoke-skill.md169 B
- 参考文档test/fixtures/preflight/publish-run/preflight.md209 B
- 参考文档test/fixtures/preflight/ship-open-interview/preflight.md305 B
- 参考文档test/fixtures/preflight/ship-unresolved-pin/preflight.md92 B
- 参考文档test/fixtures/preflight/ship-with-act/preflight.md255 B
- 参考文档test/fixtures/preflight/slogan-only-skill.md140 B
- 参考文档test/fixtures/preflight/structure-paste-skill.md252 B
- 参考文档test/fixtures/preflight/vs-pre-flight-skill.md124 B
- 脚本test/fixtures/prioritized-review/auth.ts494 B
- 脚本test/fixtures/prioritized-review/profile-cache.ts530 B
- 脚本test/fixtures/prioritized-review/report.ts119 B
- 脚本test/fixtures/prioritized-review/thumbnail.ts240 B
- 脚本test/fixtures/secret-code/config.ts817 B
- 脚本test/fixtures/sinful-code/db.ts70 B
- 脚本test/fixtures/sinful-code/user-service.ts1.8 KB
- 脚本test/fixtures/small-change/src/duration.js226 B
- 脚本test/fixtures/small-change/src/duration.test.js205 B
- 脚本test/ponytail-pass.eval.ts2.6 KB
- 脚本test/preflight.static.eval.ts6.1 KB
- 脚本test/roast-code.eval.ts19.8 KB
03 SKILL.md 正文
Roast Code
Classify the change, then review it at the depth it deserves. First pass cleans. Second pass roasts what's left.
Building Block Composition
Roast Code is a building-block review tool. It consumes:
vs-deslopsemantics during Pass 1: simplify working code while preserving behavior. Follow../vs-deslop/SKILL.mdand runnode skills/vs-deslop/scripts/run-anti-slop.mjson in-scope named TS/JS files; do not skip that file pass. Do not reportCLEANwithout that named-file runner.vs-ponytailin composed mode when judging whether new machinery is necessary.- independent-advisors during Pass 2: Codex or another model is an independent signal, not the final judge.
vs-verifyafter fixes when the cleanup or review change could affect behavior.
Before delegating, load and follow
../vs-internal-shared/references/subagents.md.
Critical Rules
- Respect chat context — review ONLY files in scope. Never expand uninvited.
- Scale to the change — classify in Phase 1 and run only that class's program. A two-line fix does not get an evidence file, five tier headers, and a cross-model review.
- Keep SMALL output flat — after classifying a change as SMALL, announce
SMALL review, skip Pass 2 and the Sin Inventory, and use no taxonomy tier names anywhere in the response. Finding a normal bug does not unlock tiered ceremony; only a Phase 1 escalation changes the class. - Never manufacture findings to fill ceremony — an empty tier, an unused worst-offender spotlight, or an approval bar with no blocker is the correct output when the change is clean. Inventing a structural finding, a severity tier, or a blocker to justify the program you ran is a review defect.
- Own roast requests — if the user asks to "roast" code or changes, use this skill instead of
simplify. - Verify before roasting — only flag what you've confirmed. Being wrong kills comedy.
- Security first — redact, never quote — secrets, keys, credentials get escalated to the top, before anything else. But when reporting them: cite
file:lineand the variable name, and NEVER output the actual secret value in your response. Quote the shape only, with the value masked (e.g.,API_KEY = "sk-live-****"). The goal is to flag the sin, not to leak the secret into the transcript, the roast, or any follow-up fix plan. This overrides any roast convention about quoting actual code for specificity — for secret values, redact. The ban covers the whole response, not just the finding that flags the secret: masking it once does not license quoting it in a zinger, an aside, or a finding about some other sin. A secret is funniest masked. When committed credentials are present, explicitly tell the user to rotate/revoke them and move them to env vars or a secret manager. - Punch up not down — mock patterns, not people.
- Be specific — cite
file:line, quote actual code. Generic roasts are lazy. - Treat review output as advisory — verify every Act finding against the real code path and adjacent files before fixing or reporting it as true.
- Reject speculative review noise — skip unrealistic edge cases, vague rewrites, and fixes that over-complicate the codebase.
- Bucket every finding — Act, Consider, Noted, or Dismissed. An unbucketed list fails. A nit in Act fails. Advisors stay the shared Codex / compose / independent-advisors lane; they do not invent a new channel.
- Pre-flight is read-only go/no-go — when the user asks for a pre-flight, a ready-to-merge check, reasons not to ship, or names a healthy point, pin that point and review. Do not publish. Do not auto-fix until they say fix.
Tone: Senior dev who's seen too much + Gordon Ramsay energy. Not mean, not personal. "I'm roasting because I care."
Pre-flight mode
Enter this mode when the user asks for a pre-flight, a ready-to-merge check,
reasons not to ship, or names a healthy point (a commit, branch, tag, HEAD~n,
or any moment the app worked). Those phrases stay in this skill. They are not a
publish and they do not open a new slash skill.
Pin, then reuse one command
Default the healthy point to main when the user does not name one. Pin it
before any review:
git rev-parse --verify <healthy-point>
git diff <healthy-point>...HEAD
Reuse that same three-dot command as the Phase 0 scope. Capture the resolved
SHA. Fail-closed in the pin and stop before review when git rev-parse cannot
resolve the healthy point, or when the three-dot diff is empty.
Do not use the ordinary empty-scope prompt ("forgot to stage") in this mode. A bad or empty pin is not a roast; it is a stop.
Review as roast, verdict as go/no-go
Keep the existing lenses, Act / Consider / Noted / Dismissed buckets, and the HIGH-RISK advisor lane. Do not import a second Breaking / Incomplete / Conflict / Safe taxonomy. A change that would have been Breaking, Incomplete, or Conflict is Act or Consider; Safe is Noted.
This mode is read-only until the user explicitly says fix. Skip Pass 1 auto-apply, SMALL inline fixes, and Act auto-fix of CAPITAL / FELONY. Sweep the lenses and report; leave the working tree untouched.
Bounded interview (optional)
After the review, ask only about unresolved implicit decisions the code cannot
answer. Compose the questioning rules from ../vs-pushback/SKILL.md
— one question at a time, recommended answer first, Finding / Recommendation /
Impact. Do not reintroduce /grill-me. Skip the interview when nothing implicit
remains.
Closeout
Close with exactly one verdict:
- SHIP — zero Act findings, the healthy point stayed pinned, and the interview is done or was skipped.
- DO NOT SHIP — list every Act with file path, line number, and snippet.
SHIP is illegal while any Act remains, the healthy point is unresolved (missing pin, bad ref, or empty three-dot diff), or an interview question is still open.
On SHIP, the standalone handoff is /vs-ship-it. Do not publish from this
skill. Do not compose this review into vs-ship-it.
Score a pre-flight closeout with
skills/vs-roast-code/scripts/reject-roast-preflight.mjs (exit 1 is a fail).
Exclusive pre-flight cases live under test/fixtures/preflight.
Phase 0: Scope
Priority:
- Chat context (specific files under discussion)
- User-specified files/dirs
- Staged:
git diff --cached --name-only - Branch diff:
git diff main...HEAD --name-only - If none: ask
In pre-flight mode the scope is the pinned three-dot command
git diff <healthy-point>...HEAD. That replaces the staged / branch / ask
fallbacks. The pin still runs first and still fail-closes on a bad ref or an
empty diff.
Resolve the scope once and retain its exact diff arguments, base branch, paths, and kind for every later phase. Use these scope kinds:
explicit-uncommittedwhen a chat- or user-selected path is staged, unstaged, or untrackedexplicit-branchwhen selected paths belong to the committed branch diffexplicit-filefor a current-file review with no diffstagedfor the staged fallbackbranchfor the branch-diff fallback
Unrelated working-tree changes never replace the selected scope. If the user
selected an untracked file, retain it as explicit-uncommitted; do not infer
scope from a tracked-files-only status query later.
If empty: "Nothing to roast. Either your code is perfect (unlikely) or you forgot to stage."
Pre-flight empty or bad pin: stop before review. Do not ask whether they forgot to stage. The pin already failed.
Phase 1: Classify the Change
Measure before reviewing. One stat read plus a look at the changed lines decides which program runs, and the whole review is scoped to that program.
git diff --stat <resolved-scope-arguments>
For explicitly named files with no diff to measure (a file review rather than a change review), use the files themselves as the size signal.
Signal 1 — size: changed files, changed lines, and whether the change is docs, comments, tests, or formatting only.
Signal 2 — risk surface: does the change touch auth or authorization, secrets and credentials, crypto, query construction, shell or dynamic execution, untrusted input parsing, persistence or migrations, concurrency, payments, a public API or wire contract, or CI, release, permission, and ownership config?
Sniff it deterministically on the changed lines, then confirm by reading the hunks it hits — a keyword is a pointer, not a verdict:
git diff -U0 <resolved-scope-arguments> \
| rg -n '^[+-]' \
| rg -v '^[0-9]+:(\+\+\+|---) ' \
| rg -in 'password|secret|token|api[_-]?key|credential|authn|authz|authoriz|permission|jwt|crypto|exec|spawn|eval\(|innerHTML|SELECT |INSERT |UPDATE |DELETE |migrat|lock|mutex|Promise\.all|charge|refund|price'
Keep the pattern tight. A sniff that matches every TypeScript diff escalates every review and buys nothing; public-API and wire-contract changes are a judgment call on the hunks, not a keyword.
Paths carry the same signal: .github/workflows/, migrations/, *.sql,
Dockerfile, lockfiles, CODEOWNERS, and dependency manifests are risk surface
regardless of how few lines changed.
| Class | When | Program |
|---|---|---|
| SMALL | Docs, comments, or formatting only, or at most 2 files and 30 changed lines — and no risk surface | One integrated pass (Pass 1 lenses plus correctness, run inline). No disk-backed capture, no Codex, no tiers. |
| STANDARD | Everything else up to 5 files and 300 changed lines, no risk surface | Pass 1 inline, parent roast, tiered Sin Inventory, approval bar. Codex only if the parent pass ends with nothing confirmed. |
| HIGH-RISK | Any risk surface touched, or more than 5 files or 300 changed lines, or the user asked for a deep review, an exhaustive roast, or a second opinion | The full program: disk-backed evidence plus the Codex cross-model review. |
Set the class from this table before reviewing and announce it before any other review commentary. The class is a program choice, not a severity label. Keep it fixed unless the evidence meets the explicit escalation rule below.
Risk surface outranks size. A five-line diff that hardcodes a credential, loosens an auth check, or builds a query by string concatenation is HIGH-RISK. Size can only downgrade a change that touches no risk surface.
Announce the class in one line before reviewing, so the user can see which program ran and override it:
Scope: 2 files, 18 lines, no risk surface — SMALL review (single pass, no Codex).
Escalate on evidence, never on suspicion. Upgrade mid-review only when the
pass confirms a security, data-loss, or crash-level problem, the diff turns out
materially larger than the stat implied, or the user asks for more depth. Say so
in one line — Upgrading to HIGH-RISK: unsanitized input reaches exec at src/thumb.ts:44 — and continue with the larger program.
An ordinary bug that the cheap pass found is the cheap pass working. Fix it, report it, and stay in the class you started in — do not inflate a finding's severity to justify a bigger program.
Never downgrade a depth the user explicitly asked for, and never upgrade because the code looks interesting.
Pass 1: Simplify (auto-fix)
Pass 1 runs the vs-deslop on-demand anti-slop file pass on in-scope
named TS/JS files. Follow ../vs-deslop/SKILL.md and run
node skills/vs-deslop/scripts/run-anti-slop.mjs on those named files;
do not skip that file pass. Do not report CLEAN without that
named-file runner.
Clean the code first. The parent performs one integrated reuse, quality, and efficiency pass. Delegate separate review domains only when deep effort was selected and the shared budget permits it; verify findings before applying fixes.
SMALL and STANDARD: read the diff directly (git diff <resolved-scope-arguments>) plus the enclosing functions and their callers. Skip
the evidence capture below — under 300 changed lines, writing a patch file, an
index, and a manifest costs more than the change is worth. Sweep the three lenses
plus correctness and security inline and fix what is confirmed. SMALL then goes
straight to the SMALL Verdict section; STANDARD continues to Pass 2.
HIGH-RISK: persist the authoritative diff instead of printing it
into model context. Resolve the shared evidence-manifest.mjs, write git diff
(or git diff HEAD for staged changes) to an evidence file under
~/.vs/$PROJECT_ID/reviews/evidence/, then emit its manifest and a bounded
diff --git / @@ hunk index. Inspect only the relevant diff hunks, owning
source functions, callers, and tests. This keeps the full review surface
available without paying for it on every pass. Follow the
disk-backed evidence contract.
PROJECT_ID=$(git config --get remote.origin.url 2>/dev/null \
| sed -E 's#\.git$##; s#.*[:/]([^/]+/[^/]+)$#\1#; s#/#-#g')
[ -z "$PROJECT_ID" ] && PROJECT_ID=$(basename "$PWD")
REVIEW_EVIDENCE_DIR="$HOME/.vs/$PROJECT_ID/reviews/evidence"
EVIDENCE_TOOL="<resolved-vs-internal-shared-skill-directory>/scripts/evidence-manifest.mjs"
mkdir -p "$REVIEW_EVIDENCE_DIR"
DIFF_PATH="$REVIEW_EVIDENCE_DIR/diff-$(date +%Y%m%d-%H%M%S).patch"
HUNK_INDEX_PATH="$DIFF_PATH.index.txt"
git diff --no-color <resolved-scope-arguments> > "$DIFF_PATH"
rg -n '^(diff --git|@@)' "$DIFF_PATH" > "$HUNK_INDEX_PATH"
node "$EVIDENCE_TOOL" manifest "$DIFF_PATH" "$HUNK_INDEX_PATH"
head -200 "$HUNK_INDEX_PATH"
Replace <resolved-scope-arguments> from Phase 0 (--cached, main...HEAD, or
the explicit user scope). Search the complete disk-backed hunk index by filename
or symbol when a relevant hunk falls beyond the initial orientation sample.
Lens 1: Code Reuse
- Search for existing utilities that could replace new code
- Flag functions duplicating existing functionality — suggest the existing one
- Flag inline logic where an existing utility applies
Lens 2: Code Quality
- Redundant state / derived values that should be computed
- Parameter sprawl instead of restructuring
- Copy-paste with variation that should be unified
- Leaky abstractions breaking encapsulation
- Stringly-typed code where constants/enums exist
- Unnecessary JSX nesting adding no layout value
- Comments explaining WHAT (delete; keep only non-obvious WHY)
- AI slop: hallucinated imports, verbose boilerplate, defensive nulls on non-null types, wrappers adding zero logic
Lens 3: Efficiency
- Redundant computations, repeated reads, duplicate API calls, N+1
- Independent operations that could run in parallel
- Blocking work on hot paths (startup, per-request, per-render)
- No-op state updates in loops/intervals — add change-detection guards
- TOCTOU existence checks — operate directly, handle error
- Unbounded data structures, missing cleanup, listener leaks
- Reading entire files when only a portion is needed
Auto-apply
Aggregate findings from all three lenses. Fix each issue directly. Skip false positives — don't argue, just move on. Briefly summarize what was fixed.
Pre-flight: do not auto-apply. Sweep the lenses and report. Leave the working tree untouched until the user says fix.
Sweep for the other copies. Before calling a fix done, grep the repository for further copies of what you just changed — the same CSS selector, constant, threshold, validation rule, or copy-pasted block. A rule that exists twice and is fixed once ships the bug in the copy you did not read, and the diff will not show it to you. Fix every copy in scope, or name the unfixed ones as a finding.
Pass 2: Parent Roast + Gated Codex Review
STANDARD and HIGH-RISK only. A SMALL change already got correctness and security coverage in its single pass; a second sweep over 18 lines finds nothing the first one missed.
Run on the cleaned code. The parent owns the roast while Codex provides the cross-model second opinion. Do not spawn another local child merely to restate the parent's review.
Parent Roast
Gather intel first: imports/exports, callers, tests. Then sweep all scan lenses — find what Pass 1 missed:
- Correctness — runtime breakage, logic errors, null access, off-by-one
- Security — injection, unsafe input, missing auth
- Architecture — structural regressions, god objects, circular deps, mixed concerns
- Error handling — swallowed exceptions, silent failures, empty catches
For the Architecture lens, when changed code suggests a structural
regression, load and follow
../vs-architect/SKILL.md in diff-scoped composed
mode. Use it to deepen confirmed structural findings; do not open its standalone
candidate-selection gate, broaden beyond the review scope, or design an
interface. Roast Code still owns severity, fixes, and the verdict.
For every meaningful change, ask whether there is a structural simplification that would make the implementation smaller, more direct, and easier to explain. Look for ways to delete whole branches, helpers, modes, wrappers, layers, or incidental concepts instead of merely moving them around.
Ponytail pass
Run this bounded pass in every review class:
- Check once whether equivalent behavior or a canonical repository helper already exists.
- Check whether standard-library or native-platform behavior can replace a new dependency or custom implementation.
- Look for an entire dependency, file, wrapper, mode, branch, abstraction, or configuration entry that can disappear while the requested behavior stays complete.
- Reject speculative extensibility that serves no explicit requirement.
A Ponytail finding must name the removable machinery, the smaller verified alternative, and the behavior-preservation check. Do not report a vague "simplify this" preference or invent a finding when the chosen machinery is already the first complete rung.
Architecture findings must be framed as deepening opportunities, not vague "extract a service" advice. Include: Files/modules, Problem, Suggested deepening, Test surface, and Why this improves locality/leverage. If the suggestion does not reduce caller pain or clarify a public seam, do not report it as architecture.
Structural review guidance lives in structural-review.md. Use it for maintainability-heavy reviews or when the diff technically works but feels tangled.
Rate each finding with confidence (0-100). Only report 80+.
Deliver 1-3 opening zingers based on worst patterns — one per real finding, never more than the findings support. Reference actual names, line counts. See comedy-techniques.md.
Codex Review
Run the advisor lane when either trigger fires:
- HIGH-RISK — the Phase 1 class already encodes it: an explicit second-opinion request, a diff past 5 files or 300 changed lines, or a touched risk surface.
- A clean parent pass on a STANDARD change — the parent roast confirmed nothing. A clean self-review is the least trustworthy verdict this skill produces, and it is the cheapest moment to buy an independent one: the lane fires only on diffs where there is nothing else left to do.
Skip it for SMALL unless the user asked. Never replace it with a mental review or a slash-command reference — the point is that a different model read the code.
In Claude Code: run the codex plugin's review command:
node "${CLAUDE_PLUGIN_ROOT}/scripts/codex-companion.mjs" review --wait
If CLAUDE_PLUGIN_ROOT is not set or the script doesn't exist, fall back to
the direct CLI below.
In Codex: run the shell command directly.
Derive the advisor scope from the Phase 0 selection — never from global git
state and never by probing. --uncommitted reviews the dirty working tree,
while --base reviews the branch diff. Preserve the chosen review surface even
when unrelated dirty files exist:
case "$REVIEW_SCOPE_KIND" in
explicit-uncommitted|staged) CODEX_SCOPE_ARGS=(--uncommitted) ;;
explicit-branch|branch) CODEX_SCOPE_ARGS=(--base "$BASE_BRANCH") ;;
explicit-file) CODEX_SCOPE_ARGS=() ;;
esac
CODEX_SCOPE_PROMPT="Review only the Phase 0 scope: $REVIEW_SCOPE_DESCRIPTION"
if [ "$REVIEW_CLASS" = "HIGH-RISK" ]; then
timeout 120 codex review "${CODEX_SCOPE_ARGS[@]}" "$CODEX_SCOPE_PROMPT" 2>/dev/null \
| node "$EVIDENCE_TOOL" capture "$REVIEW_EVIDENCE_DIR/codex-review.txt"
else
timeout 120 codex review "${CODEX_SCOPE_ARGS[@]}" "$CODEX_SCOPE_PROMPT" 2>/dev/null
fi
Re-resolve the scope only after you changed its source yourself (for example,
you committed the staged change mid-review). Running codex review --help,
retrying another flag on empty output, or polling a backgrounded review with
sleep is wasted budget.
For HIGH-RISK, resolve $EVIDENCE_TOOL and $REVIEW_EVIDENCE_DIR as in Pass 1
and inspect the artifact with bounded slice ranges around concrete findings.
For STANDARD, the command reads advisor stdout directly and does not reference
the HIGH-RISK-only evidence variables.
timeout 120 is the whole retry policy. If timeout is unavailable, run the
review in the foreground and abort at two minutes — never hand it to a background
task and poll. When the executable is missing, the environment blocks it, the
user declines to share the diff externally, or it produces nothing within the
window: log the lane as skipped with the reason and honor the missing-lane rule
below.
Missing-lane rule. The advisor lane failing does not make the review stronger, so the verdict may not read as though it ran:
- Report the lane explicitly in the closeout —
Lanes: codex ✗ (timeout). - On HIGH-RISK, the verdict is "reviewed, not independently verified — human
review required". Never
Remaining: 0on its own. - Cap confidence at 70%, exactly as the Zero-Finding Gate does.
The one confirmed miss this skill has on record happened exactly this way: a
parent pass closed a review with Remaining: 0 and no advisor lane, on a diff an
independent reviewer then broke into two capital-level and three felony-level
findings.
Re-review only when it earned a re-run. After fixes, re-run the advisor only if a CAPITAL OFFENSE or FELONY was fixed. For lower tiers the focused tests plus a re-read of the changed hunks are the proof; a second full advisor pass to confirm a misdemeanor is budget spent on wording.
Parse Codex review output for finding titles, bodies, priorities, and locations when present. Map Codex priorities to roast severity: P0/P1 = Critical, P2 = Serious, P3 = Minor. If Codex returns unstructured text, summarize the findings manually and note that structured output was unavailable.
SMALL Verdict
A SMALL review answers in a few lines: no tier headers, no worst-offender spotlight, no fix menu.
Start with the Phase 1 scope announcement containing the literal words SMALL review. Then report only the flat finding list below. Do not print CAPITAL OFFENSES, FELONIES, CRIMES, MISDEMEANORS, or PARKING TICKETS anywhere
in a SMALL response; those labels mean the wrong program ran.
- One zinger, and only if the code earned it.
- A flat list of at most 3 findings, worst first:
**Act** — **[Sin Name]** — file:line(or Consider / Noted / Dismissed) plus one line on what to do. Every finding still carries a bucket; skip empty Act/Consider headers when there are no findings. - Fix the confirmed ones directly, then one line on what changed. Pre-flight: do not fix. Report the bucketed list and go to closeout.
- Nothing wrong? Say so, name the one thing that is specifically right with
file:lineevidence, and stop. Do not open a tier to hold a single nitpick.
Escalate to the Sin Inventory below only when the pass confirms one of:
- a security, data-loss, or crash-level problem
- a provable no-op — the fix cannot execute or affect the path it targets, the guard is checked where it is never armed, or the test asserts something other than the behavior in question
A plain correctness bug, fixed inline with a one-line finding, is a complete SMALL review — do not reach for a severity tier to dress it up. A no-op fix is different in kind: green tests and a confident summary are actively laundering it, so it earns the spotlight even on a twenty-line diff.
Wrong output from a path that does execute is an ordinary correctness bug, not a no-op and not an escalation trigger.
Sin Inventory
STANDARD and HIGH-RISK. Aggregate findings from Roast + Codex. Deduplicate — if both flag the same line, keep the more specific finding. Tag Codex-only findings so the user sees the cross-model signal.
Entry guard: if the Phase 1 class is SMALL, stop and return the SMALL Verdict. Do not open this section or use any taxonomy tier label.
Group by the fixed 5-tier taxonomy — CAPITAL OFFENSES / FELONIES / CRIMES / MISDEMEANORS / PARKING TICKETS. Use these exact labels every time; keep the openers and metaphors fresh, keep the tiers stable. Each sin: N. **[Sin Name]** — file:line + bucket + one-liner roast; assign tier by impact. The bucket is actionability, not a sixth tier.
Drop a tier that has nothing in it rather than filling it. A finding that is moot by its own admission — a guard that can never fire, a nitpick on a line you just deleted — is not a PARKING TICKET, it is padding.
If 15+ sins, show top 10 by severity. Mention overflow count.
For tier criteria and the sin-to-tier mapping, see sin-categories.md. Any guidance that contradicts the fixed taxonomy (e.g., "invent your own labels") is superseded by this section.
When structural problems are present, prioritize them above cosmetic cleanup: structural regressions, missed simplification, spaghetti branching, unclear type/boundary contracts, file-size/decomposition pressure, then smaller legibility issues. Do not bury a design regression under naming or formatting noise.
Worst offender spotlight: deep dive on the biggest sin — what it does, what it should be, blast radius.
Fix CAPITAL OFFENSES + FELONIES automatically and proceed when they are Act. Do not print a tier menu — say which tiers you fixed in one line; the user can interrupt for more or less.
Pre-flight: do not auto-fix CAPITAL / FELONY Acts. Report them and go to closeout.
Review Buckets
Every finding is bucketed. A roast that lists findings with no buckets fails. A roast that puts a nit in Act as if it were blocking fails. Empty Act/Consider headers with no findings are slogan-only and fail.
Buckets are actionability, not severity. Keep the sin tiers. Tag each finding with exactly one bucket:
- Act — must change before merge; owner is the PR author / executor
- Consider — judgment call; do not block merge
- Noted — recorded, no action
- Dismissed — seen and rejected, with one-line why
PARKING TICKETS, preference, and formatting nits are never Act. Confirmed correctness, security, or crash-level sins that would block a real PR are Act.
SMALL stays flat: prefix each of the at most 3 findings with the bucket name. STANDARD and HIGH-RISK keep the Sin Inventory; each sin still carries its bucket.
Advisors still run as today (Codex, compose, the shared independent-advisors lane). They do not invent a new channel. The parent assigns the bucket after verifying the finding.
Score the roast with skills/vs-roast-code/scripts/reject-roast-buckets.mjs
(exit 1 is a fail). Exclusive roast cases live under test/fixtures/buckets.
Fix
Process selected fixes. Show before/after for major changes. Run linter if available.
Pre-flight: skip this section until the user says fix. Verdict first.
For each review finding, act from its bucket:
- Act — confirmed in code, fixed or reported; the PR author / executor owns the change
- Dismissed — not real, too speculative, intentionally designed, or worse than the code it would replace, plus the one-line why
- Consider or Noted — judgment call or recorded, no merge block
If a review-triggered fix changes code, rerun the focused tests/checks that cover the touched behavior and rerun the relevant review pass. Keep going until there are no Act findings left in scope. Once a rerun is clean, stop; do not spend another review cycle just to get nicer closeout wording.
Closeout for STANDARD and HIGH-RISK is one line, not a scorecard block:
Pass 1: [N] fixed | Pass 2: [N] found, [M] absolved | Files: N | Lanes: codex ✓ | Remaining: [N] Act
Lanes is mandatory and names every evidence lane the class called for, with a
reason when one is missing: Lanes: codex ✗ (declined — private diff). A verdict
that hides which lanes ran is the defect the missing-lane rule exists to prevent.
SMALL closes out in one line: SMALL review: [N] fixed, [N] left — <files>.
Zero-Finding Gate
Applies in every class. If all applicable review passes produce zero findings — no Act findings across every lens:
- Verify you read the changed files in full, not just diffstat.
- Name at least one specific positive assertion with
file:lineevidence: "auth is correct because X atsrc/auth.ts:42" - For STANDARD and HIGH-RISK, spend the advisor lane before declaring clean — this is the trigger described under Codex Review. A clean parent pass is a reason to get a second opinion, not a reason to skip one.
- If still zero findings after the positive-assertion pass and the advisor lane, cap confidence at 70% and note "Zero findings — low-confidence approval" in the summary.
A clean review that can't name what's specifically right is a rubber stamp. A clean review that invents a finding rather than say "clean" is worse.
Approval Bar
Do not approve a change merely because behavior appears correct. Passing tests are necessary, not a waiver for maintainability debt.
Walk this list for STANDARD and HIGH-RISK. For SMALL, only report a blocker you actually tripped over while reading the diff — a 20-line change rarely has a decomposition problem, and hunting for one is how a small review turns into a long one.
Treat these as high-confidence blockers when they are in scope, supported by evidence, and materially worsen maintainability:
- the implementation preserves incidental complexity when a clear restructure could remove branches, modes, wrappers, helpers, or layers
- the diff pushes a file from below 1000 lines to above 1000 lines without a compelling structural reason
- ad-hoc conditionals or one-off flags are bolted into an already busy flow
- feature-specific checks get scattered through shared/general-purpose code
- an abstraction, wrapper, cast, or optional boundary makes the design more indirect without buying clarity
- the code duplicates an existing canonical helper or puts logic in a layer that does not own the concept
- independent async work is serialized, or related updates can leave state half-applied, when a simpler atomic/parallel structure is obvious
If a blocker is present, leave explicit actionable feedback and push for a cleaner decomposition. If none are present, say what specific structure made the change acceptable.
Edge Cases
Good code: "I came here to roast, but... would merge without passive-aggressive comments." Cite the specific positive assertions that earned the clean bill.
Beyond saving: "This isn't technical debt, it's technical bankruptcy." Shift to triage plan.
Inherited code: "The original author is long gone. You're not on trial — the code is."
References
Output style
Apply the shared output style to every user-facing message.
Workflow
Direct: emit Next only. Composed: return to caller.
Prev: /vs-build-it | /vs-tdd | /vs-qa
Next: /vs-ship-it
Relevant: /vs-pushback
04 同仓库更多 Skills
vs-write
Use when asked to write, rewrite, edit, tighten, simplify, or polish prose for clarity, or to put text in STE, ASD-STE100, or controlled language. Produces plain, direct, accessible copy without losing substance or precision.
vs-fake-verify
暂无简介
vs-fake-verify
暂无简介
vs-verify
Use when asked to verify, prove it works, check completion, or before claiming done. Produces evidence-backed PASS/WARN/FAIL status.
vs-fake-verify
暂无简介
vs-verify
Use when asked to verify, prove it works, check completion, or before claiming done. Produces evidence-backed PASS/WARN/FAIL status.