ami-audit-quality
Performs a deep code quality, security, data schema consistency, and structure audit on modified files using both test execution and code inspection.
SKILL.mdskills/ami-audit-quality/SKILL.md
GitHubSkill: Push Quality Auditor
When invoked, act as a strict Code Reviewer focused on code verification by execution, security, data schema consistency, structural integrity, and linguistic consistency.
Everything else reviews by reading; this skill begins by reviewing through execution.
Workflow
1. Scope and Execution Verification (Run What the Repo Declares)
Before reading code heuristically, verify whether the repository's own tooling passes on the modified changes:
Discover Declared Tooling: Inspect the repository's configuration manifests for declared verification commands:
package.json(scripts:typecheck,lint,format:check,test)pyproject.tomlorsetup.cfg([tool.mypy],[tool.ruff],[tool.pytest])Makefile(check,lint,test)Cargo.toml(cargo check,cargo test,cargo clippy).pre-commit-config.yamlRule: Run only what the repository declares. Do not invent or globally install unconfigured linters. Prefer local project binaries (node_modules/.bin/,.venv/Scripts/).
Run in Increasing Order of Cost:
- Type-Check / Compile: The cheapest non-gameable signal.
- Lint / Format Check: Run in check mode only (do not auto-format mid-verification).
- Test Suite: Run the tests covering modified components. Note whether failures pre-existed on the base commit before reporting them.
- Mechanical Dead Code / Duplication: Run available static analyzers (
ts-prune,vulture,jscpd) if declared.
Execution Signal Requirements:
- Every detected failure MUST provide the exact one-line command to reproduce it.
- Mandatory
gaps:line: Any check that could not run (tool missing, test suite skipped) MUST be explicitly listed as a gap. A check that did not run is never a silent pass.
2. Static Security & Structural Quality Check
Inspect the code diffs (git diff or changed files) across core dimensions:
Security & Hygiene:
- Ensure there are no exposed credentials, API keys, private tokens, or dangerous raw shell executions.
- Guarantee exception handlers and diagnostic loggers do not log sensitive data (PII, authorization tokens).
- Ensure no local machine absolute paths (e.g., user home directory paths) leak into committed files.
Failure Path Resilience & State Safety:
- State Cleanup & Rollback: Verify that failures mid-execution (timeouts, thrown exceptions, network drops) trigger proper cleanup to avoid corrupted state.
- Retry Idempotency: Ensure operations subjected to retries do not produce duplicated side effects or inconsistent state.
Duplication & Dead Code:
- Verify that new code does not duplicate existing utilities or domain helpers. Enforce centralizing reusable logic.
Data Schema, ORM & Migration Consistency (When Data Models Are Modified):
- If the diff modifies database schemas, ORM models, raw SQL queries, or serialized API payloads, verify that all referencing queries, connections, and callers structurally match the updated definitions.
- Temporal Compatibility & Two-Phase Migrations: Confirm that renamed or dropped fields do not break in-flight/cached payloads, enforce validation at entry boundaries (no de-facto gating), and check whether destructive schema changes require a dual read/write migration phase.
Agentic AI & Declarative Code Integrity (.md / .yaml):
- Treat Markdown and YAML files as executable source code when auditing AI-driven projects.
- Structural Debt: Scan for duplicated instructions or contradictory rules across related files.
- DRY Principle: Procedural steps belong in skills, while agent profiles contain high-level roles and constraints.
- Token Efficiency & Bloat: Ensure prompt instructions are concise and skimmable.
UI & Design System Consistency (When Interface Files Are Modified):
- If the diff touches UI files (
.tsx,.jsx,.xaml,.vue,.svelte,.html,.dart), adhere toskills/ami-audit-ui-consistency/SKILL.md. - Zero Raw Hex in New Code: Reject unconfigured hexadecimal or RGB colors outside theme files.
- Component & Asset Hygiene: Enforce max 1 primary CTA button per view, standard interactive states (
hover,focus-visible,disabled), single icon family, zero duplicate glyphs for the same action, and accessible labels on icon buttons.
- If the diff touches UI files (
3. Adversarial Verification Gate (Fresh-Context Review)
For changes touching critical domain logic, algorithms, or API boundaries, conduct or invoke an adversarial check (ami-adversarial-verifier):
- Evaluate the change strictly against the stated requirement and its existing callers, ignoring self-justifications.
- Verify boundary conditions: empty inputs, null values, concurrent executions, and caller compatibility.
- Ensure the change does not introduce behavioral regressions in existing interfaces.
4. Linguistic & Comment Consistency
- Comments & Docstrings:
- Ensure comments explain why something is done, not what. Avoid commenting obvious lines.
- Actively audit docstrings for conciseness; avoid narrative bloat.
- Zero Internal Planning References:
- Comments and docstrings MUST NEVER reference internal phases, stages, steps, or design options (e.g., 'Phase 1', 'Stage 2', 'Etapa 3', 'Option A', 'Opción B').
- All code documentation must describe technical rationale, domain invariants, and contracts from a permanent domain perspective.
- Language Consistency:
- Source code, docstrings, and inline comments must strictly be in English.
5. Reporting
Emit a structured verdict:
- If failures or defects are found: Reject the check. Provide:
- Exact command to reproduce execution failures.
- Line numbers and triggers for code or security defects.
- Concrete remediations required.
- Await explicit user approval before applying any fixes.
- If all checks pass: Emit the clean verdict including executed checks and reported
gaps:, then terminate the skill.
Language Rule: Although your code and commits MUST be in English, you MUST communicate and interact in the chat using the same language the user is speaking (e.g., Spanish, French, etc.).