Coordinates bounded subagent review, intent-preserving fixes, and validation after a meaningful implementation slice; use only when the user explicitly asks to run Garfield or its review-fix-verify loop before handoff, not for standalone reviews, CI iteration, brainstorming, or skill authoring.
Instrucciones de origen · Vista previa de solo lectura
name
garfield
description
Coordinates bounded subagent review, intent-preserving fixes, and validation after a meaningful implementation slice; use only when the user explicitly asks to run Garfield or its review-fix-verify loop before handoff, not for standalone reviews, CI iteration, brainstorming, or skill authoring.
disable-model-invocation
true
spec_hash
5a1560d5f494
Garfield
Run a skeptical, concise implementation hardening loop after each meaningful code slice. Focus on concrete flaws caused by the current diff, avoid unnecessary work, and preserve the core user or PR intent.
Never activate Garfield unless the user explicitly invokes it. If review-only subagents or required capacity are unavailable, stop and report that Garfield cannot run as specified.
Frame the slice
Before review, inspect:
repository status, diff/base, and changed files
applicable repository instructions
the request, relevant specs and docs, tests, and known non-goals
generated artifacts, schemas, migrations, lockfiles, manifests, and dependencies
available validation commands and intentional tradeoffs
source-app policies from sorted policies/**/*.md, excluding any README.md or policy-template.md
Snapshot the requested behavior, intended behavior changes, compatibility expectations, touched areas, and non-goals. Scope review to the current diff and directly related files. Preserve unrelated dirty-worktree changes.
Build the effective review plan
Compare each bundled policy with discovered source-app policies by intent and scope before choosing reviewers. A repo-wide local policy supersedes a bundled policy governing substantially the same concern even when names differ. A narrower or adjacent local policy supplements it. Omit superseded bundled policies completely; never send one to a reviewer or use it for findings.
Classify every candidate below as applicable or skipped with a concrete diff-based reason. Confidence that the code looks safe is not a skip reason.
behavior/spec — always
repository instructions — always
validation sufficiency — always
specs/docs — behavior or documented contracts changed or should have changed
dead code — paths were replaced, removed, or refactored
delayering — wrappers, flags, adapters, abstractions, indirection, or ownership changed
type boundaries — typed interfaces, casts, nullable values, any, unknown, or serialization changed
generated/dependencies — an API/storage schema, migration, generated/reference output, manifest, lockfile, dependency, package artifact, CI dependency surface, or eval recording changed or should change
code comments — comments/docstrings changed or new non-obvious code makes references/code-comments.md material
implementation minimalism — guards, fallbacks, wrappers, configuration, edge cases, or supporting tests make references/implementation-minimalism.md material
interface design — public/module interfaces, lifecycle, naming, ownership, or platform boundaries make references/interface-design.md material
test quality — tests/fixtures changed or behavior creates a concrete obligation under references/test-quality.md
each effective source-app policy — its scope governs the slice
Record an omitted bundled policy as skipped — superseded by <source-policy-path>.
For each applicable native task, read only its card in references/review-lanes.md. Use the card to define what that reviewer owns, excludes, and must prove; keep severity, causality, and output rules here.
Delegate reviews
Assign exactly one no-edit subagent to every applicable task or effective policy. Give each reviewer the user goal, intent and non-goals, diff/base, changed and related files, relevant instructions/specs, validation evidence, and intentional tradeoffs. Give a native-task reviewer only its matching card plus slice context; give a policy reviewer only its one policy plus slice context.
Keep at most three Garfield review subagents open. Spawn to available capacity, wait, collect and close or release completed reviewers, then refill. Drain all review agents before independent verification. Never silently drop a task, rely on implicit queuing, repeatedly retry a capacity error, or allow a reviewer to edit.
Use direct, spec, policy, test, validation, missing, or inferred evidence. Use a changed-code location when available; otherwise use a command, test, policy, spec, artifact, manifest, or lockfile locator. Never accept inferred alone as a blocker.
Tell reviewers to return only findings caused or made material by the current diff, including required omissions. They must not propose speculative hardening, compatibility changes, API or permission changes, normalization, precedence changes, new abstractions, or unrelated cleanup. If none exist, they return none.
Coordinate and repair
Treat reviewer output as advice. Validate every finding yourself.
Cluster reports by stable locator and normalized underlying cause before deciding severity or work. Assign one owning lane to each concern; treat other reports as corroborating evidence, a better locator, or an alternate remedy. A broad obsolete-workflow concern owns the members it already names.
Decide whether the demonstrated defect is valid separately from whether the suggested remedy is proportionate. Accepting a concern never makes its proposed fix authoritative.
blocker: fix before proceeding when the smallest fix preserves intent.
high: fix when the smallest fix preserves intent or repairs a slice regression.
medium: fix only when current-diff-caused and it removes introduced bloat, repairs stale local evidence, or is a one-hop edit to changed code.
low: optional; never loop solely for it.
Reject vague, preference-only, or evidence-free findings. Defer valid pre-existing debt, adjacent hardening, unclear behavior changes, and anything whose smallest fix expands the patch or core intent.
Apply only the smallest accepted fixes. Before adding a validator, trigger, cache, compatibility layer, wrapper, state field, or elaborate test harness, compare a smaller boundary, type, transaction, assertion, deletion, or existing-test repair. Preserve permission, security, idempotency, lock-order, migration, and durable-integrity checks unless you can prove what establishes the invariant and how long it remains valid.
Never introduce behavior changes, speculative hardening, broad rewrites, unrelated cleanup, or duplicate fixes. Do not add comments, tests, wrappers, compatibility paths, persistent mechanisms, or test machinery solely to satisfy overlapping review suggestions without a distinct current-diff contract and proportionate need. Do not change accepted inputs, error behavior, permissions, parameter precedence, defaults, serialization, validation policy, or public API semantics unless explicitly requested or required to repair a regression introduced by the slice.
Validate and repeat
After fixes, run the smallest relevant tests, type checks, linters, builds, schema checks, generated-artifact checks, or record an explicit blocker. Recompute applicability from the repair delta and rerun only affected concern owners and newly applicable lanes. Always rerun behavior/spec after a behavior-changing repair; do not restart the full review for a narrow comment, test, or artifact-only edit.
Stop successfully when no current-diff-caused blocker, high, or eligible medium concern remains and targeted validation passes. Stop blocked when:
After draining review agents, use a separate no-edit verification advisor when validation is uncertain, the slice is risky, generated or dependency artifacts changed, or final readiness needs independent evidence. Ask it only whether the final diff preserves intent, validation matches the touched behavior and files, required checks ran or have explicit blockers, and deferrals have concrete evidence and valid reasons. Do not ask it to re-review the whole diff.
Address any material verification gap using the same intent and severity gates.
Handoff
Report only:
garfield: pass or garfield: blocked
validation commands and results
independent verification result when run, or when skipped for a non-obvious reason
residual accepted or deferred blocker, high, or medium concerns
deferred adjacent improvements or unclear behavior changes
