Deep Review

Multi-dimensional code review built on independent subagents. Review breadth comes from parallel per-dimension reviewers; precision comes from an adversarial verification pass that falsifies findings before they 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. Review must run in independent subagents with a third-party reviewer stance. Never silently degrade deep mode to "the main agent reviews and then verifies its own findings".
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	Main agent reviews inline against the Quick checklist section of each applicable dimension file. No subagents.
Deep	Explicit only: /deep-review, "run deep review", "full multi-agent review"	Full orchestration: per-dimension review subagents → pipelined verify subagents → 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.

Dimensions

Rules live in one place: references/dimensions/, one file per dimension. Both modes read the same files (light mode reads only the Quick checklist section; deep-mode subagents read the full file plus its listed rule sources).

Dimension	id prefix	Covers	Verified?
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
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
code-style, logic, business-logic, reuse-architecture	never (skip only for docs/lockfile-only diffs)
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	never in deep mode (cheap external-state checks)
skill-freshness	never in deep mode (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 read.

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

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; read the Quick checklist section of each surviving dimension file, plus extension-pack counterparts.
3. Review inline. Findings must cite a rule source or code evidence; respect the codebase-calibration principle.
4. Label every finding on the same two axes deep mode uses, and let them drive the recommendation:
   • Introduced or pre-existing — if the problem lives in code this diff did not change, do not fix it here. Say who introduced it (git blame / git log -L on the location) and offer to file a Linear issue for that owner. The only exception is an old bug this diff makes reachable or harmful for the first time, at P0.
   • Likelihood — high / medium / low, judged on the real production path. State it next to the severity. A low-likelihood, non-blocking finding is a follow-up, not a fix-now item; on a repeat review of the same change, list it in one line and move on instead of re-arguing it.
5. Output in your environment's normal review format (light mode does NOT use the deep report template). When release-risk produced pre-deploy confirmation items, list them as a short checklist separate from the findings — they are questions about production state, not defects. Mention that deep mode exists if findings suggest the diff deserves a full pass.

Deep mode 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 degrade to a single-agent pass (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.