Skill: Review Codebase

Purpose

Run a scope-only atomic review of the current state of a given path (a single file / a directory / a repository). Paired with review-diff (which reviews git changes only), it is one of the two scope-step candidates for orchestrate-code-review — this skill looks at the snapshot, review-diff looks at the change.

Not done here: cognitive dimensions such as security / performance / architecture (the cognitive-step atomic skills review-security / review-performance / review-architecture take those), and language- or framework-specific analysis (the language / framework steps take that).

Core Objective

Primary goal: produce a scope-only findings list that identifies the structural problems in a given path (boundaries, patterns, dependencies, tech debt, interfaces).

Success criteria (all of them must hold):

1. ✅ Scope resolved: use the user's path or directory, or the repository root when no narrower scope is named
2. ✅ 5 dimensions covered: findings are emitted for module boundaries, pattern consistency, cross-module dependencies, tech debt and interface stability
3. ✅ Precise locations: every finding carries a file:line reference
4. ✅ Format conformant: every finding conforms to Findings Spec §5.1; category is scope.
5. ✅ Large scopes handled: for a repository-level scope, emit by layer (module / directory), or settle a priority subset with the user
6. ✅ No overreach: no security / performance / architecture / language / framework cognitive findings are emitted (they are flagged, pointing at the matching atomic skill)

Scope Boundaries

This skill owns:

• Structural review of the current state of the given path
• Findings across 5 dimensions: module boundaries, pattern consistency, cross-module dependency and coupling, tech debt and maintainability, interface stability

This skill does not own:

• Reviewing git changes alone (use review-diff)
• The full orchestrated review (use orchestrate-code-review)
• Language- / framework-specific conventions (use review-<lang> / review-<framework>)
• The security / performance / architecture cognitive dimensions (use review-security / review-performance / review-architecture)

Handoff point: once the findings are out, they feed the scope step of orchestrate-code-review for aggregation, or go to the user to decide what follows (refactoring / a deeper review).

Use Cases

• New module review: given src/auth/, look at the current structure and dependencies
• Legacy path audit: given a path, look at tech debt and boundary problems
• Sampled review: a file or directory a colleague names, with no diff needed
• As the scope step of orchestrate-code-review: this one or review-diff, never both

Behavior

Scope resolution

• The input defines the scope: a single file / a directory / the repository root / several paths, named by the user
• No dependence on a diff: analyze the current file content; a diff the user supplies is context only, not a requirement

Scope defaults

Use the named path without a confirmation prompt. For a repository-wide review, work by layer unless the user names a subset; ask only if the requested scope cannot be identified.

The 5 dimensions

For the code in scope (at the layer / subset the user chose), emit findings on these dimensions:

1. Module boundaries: whether module / service boundaries are clear, whether responsibilities are single, whether the dependency direction is sound
2. Pattern consistency: whether patterns are used aptly and match the repository's existing style
3. Cross-module dependency and coupling: dependency relations, cyclic dependencies, degree of coupling
4. Tech debt and maintainability: duplication, complexity, testability, the current state of docs and comments, code with no apparent caller, and deprecation markers still in place
5. Interface stability: how clear and how stable a module's outward interface is

Every finding must carry a file:line reference.

Flagging out-of-scope findings

When the analysis turns up a concrete security / performance / architecture / language / framework problem: flag it and point at the matching atomic skill, without opening the analysis. For example:

Potential SQL injection risk detected (user input concatenated without escaping); suggest running review-security

Two tech-debt observations are worth calling out because this Skill is usually the first to meet them and neither is settled by reading the code. A symbol with no apparent caller may be reached by reflection or named in configuration, so deciding it needs a runtime signal over a full business cycle; a deprecation marker is decided by the owner, removal point and replacement recorded beside it. Report the observation with its location and route the decision:

jobs/settlement.py:18 defines a class no call site in scope reaches, and config/jobs.yaml names handlers as strings; suggest running review-architecture, which decides it under ARC-010 with the evidence that item requires

api/v1/search.py:7 carries a deprecation marker recording neither an owner nor a removal point; suggest running review-architecture, which decides it under ARC-011

Input and Output

Input

• Path: one or more file / dir paths
• Optional: a focus hint ("concentrate on module boundaries", for example)

Output

• Findings list: conforms to Findings Spec §5.1; category is scope.
• Large-scope summary: when organized by layer, emit the findings count and severity distribution for each layer

Restrictions

Hard boundaries

• Do not assume "the diff only" — by default this skill reviews the complete current state of the given scope
• Emit no cognitive-dimension findings (security / performance / architecture)
• Emit no language- / framework-specific findings
• Emit no finding that lacks a file:line reference
• Use no vague language ("might be a problem", carrying neither a type nor a direction → delete it)

Skill boundaries

Not done here (other atomic skills own it):

• Reviewing git changes → review-diff
• Full-dimension orchestration → orchestrate-code-review
• Language conventions → review-<lang>
• Framework conventions → review-<framework>
• Security / performance / architecture → review-security / review-performance / review-architecture

Self-Check

• The review scope was resolved from the request or the documented default
• A large scope was emitted by layer, or a priority subset was settled
• All 5 dimensions are covered (module boundaries / patterns / dependencies / tech debt / interfaces)
• Every finding carries a file:line reference
• No cognitive / language / framework dimension finding was emitted (they are flagged only, pointing at the matching atomic skill)
• The output conforms to Findings Spec §5.1.

Examples

Example 1: a single directory

• Input: src/auth/
• Output: findings across the 5 dimensions, grouped by file, each carrying a reference of the auth.go:42 kind; a weak crypto algorithm is flagged only, pointing at review-security

Example 2: a single file

• Input: pkg/validator/validator.go
• Output: findings on module responsibility / interface clarity / test coverage / dependencies on upstream modules, and so on

Example 3: the whole repository (large scope)

• Input: the repository root
• Behavior: first emit a findings summary table by layer (top-level directory), then have the user pick a priority subset to go deeper on
• Output: the layer summary plus detailed findings for the priority subset