Deep Review

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 need surrounding context, not only diff fragments. Deep mode verifies candidate findings through independent agents returning confirmed / false_positive / need_more_context. Light mode uses the bounded follow-up procedure below; it does not add a separate per-finding verification pipeline.
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

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.

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.

"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. Assess the findings against their evidence. In an authorized implementation task, fix accepted, safe, in-scope issues and run appropriate checks. To reject or downgrade a finding, send the counterevidence to the independent verification round below rather than overruling the reviewer yourself. A standalone review request does not authorize edits.
5. Allow at most one follow-up verification round per review request, covering disputed findings and fixes together, regardless of dimension. Append the original findings and counterevidence to the light-review prompt and supply the current diff; ask the independent reviewer to check changed behavior and possible regressions. Correct or withdraw findings only with concrete evidence and state the reason. Fixes, new commits, and context compaction do not reset this budget. If verification still fails or leaves a finding unresolved, report that status honestly without further automatic review rounds. Necessary authorized repairs can continue, but do not claim a subsequent independent pass. A new explicit user request for review takes precedence over this automatic limit.
6. Report the resulting findings and verification status in the environment's normal review format (not the deep report template). Keep pre-deploy confirmation items separate from defects.

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.