Deep Review

Multi-dimensional code review built on independent reviewers. Review breadth comes from parallel dimension coverage; precision comes from adversarial verification and global duplicate consolidation before findings reach the report.

Core principles

Every design choice below serves one of these. When unsure how to execute a step, come back here.

1. Anti-hallucination — reviewers that only see diff fragments invent bugs. Candidate findings are therefore falsified one by one by an independent verify subagent that reads full context and returns a three-way verdict (confirmed / false_positive / need_more_context). Three-way verdicts beat confidence percentages: calibrated-sounding scores are unreliable as hard filters.
2. Anti self-approval — an agent that just wrote the code is grading its own homework and will pass it. Both modes use an independent reviewer with a third-party stance. Light: one reviewer that is not the main agent. Deep: independent subagents per the environment manual. Never silently degrade either mode to "the main agent reviews the diff itself".
3. Rules over model — review quality comes from fine-grained, executable dimension rules, not from a smarter model. Subagents run on balanced/fast model tiers; each dimension file tells them exactly how to check, what counts, and what does not.
4. Calibrate to codebase and lifespan — hold the diff to the standard the codebase already meets, not an idealized one. If a pattern is widespread in the existing code and this diff does not make it worse, it is not a finding. Declared-temporary code (time-boxed campaign, experiment, one-off script) is judged against its lifespan: hardcoding and low-extensibility shortcuts are the intended trade-off for shipping fast, and "delete the code at expiry" is a valid plan — do not demand configurability from code built to be deleted. (Security is exempt from all calibration — see the dimension file.)
5. Speed is a feature — one wave of parallel reviewers, verification pipelined per dimension (never a global barrier), irrelevant dimensions pruned up front.

Two entry modes

Mode	Trigger	What runs
Light (default)	Any ordinary review ask: "review this PR", an informal "look at this change for problems", a diff pasted for review — but not explain-only questions about a change	One independent reviewer against the Quick checklist of each applicable dimension. Prefer a different-vendor harness when the session can dispatch to one; otherwise one same-harness subagent. No verify pass. The main agent does not review.
Deep	Explicit only: /deep-review, "run deep review", "full multi-agent review"	Full orchestration: dimension review agents → pipelined verification → global consolidation → structured report → interactive fix flow.

Do not auto-escalate light to deep. Do not run deep mode for a casual "看看这个改动" — that is light mode.

Deep-mode budget

Within one logical requirement (the same requirement, PR, or branch), run Deep mode at most once by default. Follow-up reviews after fixes, CI work, rebases, cleanup, context compaction, or a resumed session use Light mode, even when they review the whole diff again. These events do not reset the budget, and the review/verify/consolidate waves inside one Deep run all count as that single run.

Run Deep again only when the user explicitly requests another Deep pass, or when a genuinely new logical requirement begins. A casual follow-up such as "review again", "复核一下", or "再看看" is a Light request, not implicit authorization for another Deep run.

Dimensions

Rules live in one place: references/dimensions/, one file per dimension. Both modes use the same files. The light reviewer reads the full Quick checklist section, including nested example subsections; deep-mode agents read the full dimension file plus only the rule sources and routed references applicable to the touched surface.

Dimension	id prefix	Covers	Verified?
ai-coding-bad-habits	ai	narrow special cases in generic code, partial refactors, redundant type guards/aliases, comment narration, and precedent-blind implementation	yes
code-style	style	naming, readability, dead code, comments, i18n hardcoding, UI-library and styling conventions	yes
logic	logic	logic correctness: edge cases, null, races, error handling, state machines, requirement deviation, test coverage	yes
business-logic	design	design judgment: framework misuse, best-practice violations, solution-weight mismatch, self-inflicted complexity	yes
new-feature	feature	new-capability design: domain vocabulary and structure, durable data model and migration evidence, product analytics and operational observability	yes
reuse-architecture	reuse	duplicate implementations, unused existing patterns, extensibility, architectural boundaries	yes
performance	perf	N+1, blocking calls, resource leaks, render-path waste, DB migration locking and idempotency	yes
release-risk	risk	ship/no-ship gate: irreversible persisted state (incl. schemaless shape drift), dev-cycle migration residue, in-flight work & config at deploy time, irreversible outbound effects, prompt/tool-description behavior shifts, shared-surface and high-frequency-UI blast radius, PR purity	yes + checks
security	sec	injection, auth bypass, secret/PII leakage, business-slot confidentiality	yes
compatibility	compat	light/dark theme, desktop app / web (desktop, mobile) / RN, released-client API compatibility, client vs server agent runtime (gateway on/off), Vercel vs Docker deploys, paired router configs	yes
ux	ux	empty/loading/error states, async feedback, confirmation flows, design-value adherence	yes
observability	obs	debug: unexplained fixes, uncommented hacks, silent catches, missing logs; product: analytics on qualifying new capabilities; perf: monitoring on high-frequency / high-traffic / polling paths	yes
workflow	flow	issue tracking state, PR description freshness, undocumented key decisions, CI / preview build status	no (objective state)
skill-freshness	skill	agent skills invalidated by this diff, knowledge worth distilling into a new skill	no (advisory)

Verified? no means findings from that dimension are objective state checks or advisories — they skip the verify pass and go straight to the report. yes + checks means the dimension's findings are verified normally, but it also emits a second, unverified output — release_checks, pre-deploy confirmation items about production state the repo cannot answer (see its dimension file).

Pruning table (deep mode)

Before spawning, the main agent prunes dimensions that cannot apply to the diff. List pruned dimensions and the one-line reason in the report header. When in doubt, run the dimension.

Dimension	Skip when
ai-coding-bad-habits, code-style, logic, business-logic, reuse-architecture	never (skip only for docs/lockfile-only diffs)
new-feature	diff does not introduce a new product, operator, or platform capability
performance	no server/db/loop/render-path code touched (e.g. docs, copy, pure type changes)
release-risk	no DB schema/migration, no schemaless persisted payload, no queue/cron/config/outbound dependency, no prompt or tool-description text, no shared component or package public API, no high-frequency user surface, and the diff has one obvious purpose
security	lockfile/generated-only diff — docs and copy still run it (text is a leak vector: secrets, internal URLs, commercial details)
compatibility	diff touches no UI theming/routing, no API contract, no deployment config, no runtime-branching code
ux	no user-facing surface changed (components, styles, copy, interaction flows)
observability	no error handling, async flow, server code, new user-facing capability, or perf-sensitive path touched
workflow	light: no PR/issue/CI context is available; deep: never (cheap external-state checks)
skill-freshness	light: the diff neither changes agent instructions nor changes behavior/conventions covered by an existing skill; deep: never (cheap)

"Docs-only" means human-facing prose only. Files that are executable instructions for agents — .agents/skills/**, AGENTS.md / CLAUDE.md, prompt templates, orchestration manuals — count as code for pruning purposes: their "prose" carries control flow, contracts, and rules whose contradictions are exactly what logic / business-logic / reuse-architecture exist to catch. A diff touching them is never docs-only.

Light mode applies the same table to decide which Quick checklists to pass to the reviewer.

Extension packs

A wrapping repository (e.g. a private deployment that vendors this repo as a submodule) can extend the rule set without forking this skill: any sibling skill directory in the active skills root matching deep-review-* (for example .agents/skills/deep-review-cloud/) is an extension pack.

• Extension packs contain dimensions/*.md files in the same format; a file named after a built-in dimension extends it (load both), a new name adds a dimension.
• An extension file declaring extends: <name> where no such built-in dimension exists (typically because the wrapping repo pins an older submodule commit) falls back to standing on its own. Such a file must therefore carry its own id_prefix / verify / skip_when so it stays usable either way; when the built-in is present, the built-in's frontmatter wins.
• Both modes must check for extension packs at startup and load whatever is present. Absence is normal — this skill is self-sufficient.
• Extension packs may carry rules that must not live in this open-source repo; never copy their content into files under this directory.

Light mode procedure

If your user message is already an instantiated light-review prompt, you are the reviewer — follow that prompt and do not re-enter this procedure.

1. Determine review scope exactly as deep mode step 0 does (see the environment manual's scope rules — three-dot diff from a base that does not lag the fork point, submodule diffs included), but skip the background-hunting extras when context already tells you what changed.
2. Apply the pruning table; collect the surviving dimension files plus extension-pack counterparts. Do not read the checklists yourself.
3. Instantiate references/light-review-prompt.md with the surviving {dimensions}, {dimension_files}, {scope_summary}, and {changes}. Dispatch exactly one independent reviewer, trying in this order:
   1. If this session can dispatch a self-contained task to a different vendor's coding-agent harness (not a same-harness subagent), use that. Prefer a preset whose job is independent code review when the dispatcher exposes one.
   2. Otherwise spawn one same-harness subagent. Disable context inheritance when the spawn API allows it.
   3. If the chosen path fails, try the next. If none work, stop and tell the user why. Do not review the diff yourself.
4. Relay the reviewer's findings in your environment's normal review format (light mode does NOT use the deep report template). Do not drop, downgrade, or re-verify findings. When the reviewer listed pre-deploy confirmation items, keep them as a short checklist separate from defects. Mention that deep mode exists if findings suggest the diff deserves a full pass.
   • On a follow-up round (verifying that earlier findings were fixed), dispatch again on the newly changed code as a fresh diff — never just confirm the requested edits landed. Fix commits introduce new logic (guards, parsers, refactors) whose bugs a checkbox pass will miss.

Deep mode procedure

Before entering this procedure, check the logical requirement's Deep-mode budget above. If Deep has already run and the user did not explicitly request another Deep pass, use the Light procedure.

Pick the manual for the current environment and follow it end to end:

• Claude Code → references/claude-code/main.md
• Codex → references/codex/main.md

If the environment is not listed, tell the user deep mode does not support it yet and offer light mode instead. Do not improvise another environment's mechanics, and do not simulate deep in the main agent or as one packed reviewer (see principle 2).

Keeping this skill sharp

The skill-freshness dimension and the workflow_feedback channel in the subagent return schema exist to feed observations back into these files. When a review surfaces a rule gap, an outdated rule, or a recurring team preference, update the relevant dimension file in the same PR or a follow-up — that is how calibration stays current.