Agent Plugin Review

Plugin Review

Overview

Review AI plugin PRs by running deterministic structural checks first, then applying LLM judgment for skill quality and workflow architecture. Post findings as inline PR comments.

Process

Step 1: Structural lint

Run scripts/lint_plugin.py against the plugin directory:

python scripts/lint_plugin.py <plugin-dir> --evals-dir <evals-dir> --json

The script checks:

• Every skills/*/SKILL.md has a corresponding eval file
• SKILL.md frontmatter has name and description
• No hardcoded local paths (drive letters, absolute OS paths)
• No version printing instructions
• Referenced files (references/*.md) exist
• Commands reference existing skills
• Path style consistency across commands

Report findings grouped by severity (error > warning > info).

Step 2: Eval lint

If the PR includes eval files, invoke agentv-eval-review for AgentV-specific eval quality checks.

Additionally, check each eval YAML for these structural patterns:

• File path format: Every type: file input value MUST start with a leading / (workspace-root-relative). Paths like plugins/foo/SKILL.md are wrong — correct form is /plugins/foo/SKILL.md. Scan every type: file entry and flag any missing leading slash, showing the corrected path.
• Repeated inputs: If the same file input (same type: file + value) appears identically in every test case, recommend extracting it to the top-level input field. AgentV eval files support a top-level input section that applies to all tests, eliminating per-test duplication.

Step 3: Skill quality review (LLM judgment)

For each SKILL.md, check against references/skill-quality-checklist.md:

• Description starts with "Use when..." and describes triggering conditions only (not workflow)
• Description does NOT summarize the skill's process — this causes agents to follow the description instead of reading the SKILL.md body
• Body is concise — only include what the agent doesn't already know
• Content is domain-specific (internal conventions, business patterns, context for WHY) — universal concepts AI agents already know are excluded
• Imperative/infinitive form, not second person
• Heavy reference (100+ lines) moved to references/ files
• One excellent code example beats many mediocre ones
• Flowcharts only for non-obvious decisions
• Keywords throughout for search discovery
• Cross-references use skill name with requirement markers, not @ force-load syntax
• Discipline-enforcing skills have rationalization tables, red flags lists, and explicit loophole closures
• Consistency — no contradictions within or across files (tool names, filenames, commands, rules)
• No manual routing workarounds — if AGENTS.md or instruction files contain heavy TRIGGER/ACTION routing tables or skill-chain logic, the skill descriptions are likely too weak. Good descriptions enable auto-discovery without manual routing.

Step 4: Workflow architecture review (LLM judgment)

For plugins with multi-phase workflows, check against references/workflow-checklist.md:

• Hard gates between phases (artifact existence checks)
• Artifact persistence convention (defined output directory)
• Workflow state metadata for cross-session resumption
• Resumption protocol (detect existing artifacts, skip completed phases)
• Standardized error handling with retry
• Trivial change escape hatch
• Artifact self-correction with corrections log
• Learning loop mechanism

Hard gate detection recipe — For each phase skill after the first:

1. Read the SKILL.md body
2. Check whether it verifies that the previous phase's output artifact exists before doing any work
3. If no such check exists, flag it as a missing hard gate. Recommend adding a gate at the top of the skill that checks for the prerequisite artifact (e.g., deploy-plan.md) and stops with a clear message telling the user which skill to run first if the artifact is missing

Step 5: Post review

Post findings as inline PR comments at specific line numbers. Group by severity:

• Critical — Broken references, missing evals, factual contradictions, missing hard gates
• Medium — Naming inconsistencies, hardcoded paths, missing assertions, ad-hoc error handling
• Low — Style inconsistencies, description improvements

Use a PR review (not individual comments) to batch all findings.

Skill Resources

• scripts/lint_plugin.py — Deterministic plugin linter (Python 3.11+, stdlib only)
• references/skill-quality-checklist.md — Skill quality checklist (CSO, descriptions, content, discipline skills)
• references/workflow-checklist.md — Workflow architecture checklist (OpenSpec, hard gates, artifacts)

External References

For deeper research on challenging reviews, consult these resources via web fetch, deepwiki, or clone the repo locally:

• Agent Skills specification — Official SKILL.md format, frontmatter fields, progressive disclosure rules
• Agent Skills best practices — Context spending, calibrating control, gotchas, scripts, validation loops
• Agent Skills description optimization — Trigger testing, train/validation splits, overfitting avoidance
• Agent Skills using scripts — Self-contained scripts, --help, structured output, idempotency, exit codes
• AgentV documentation — Eval YAML schema, assertion types, workspace evals, multi-provider targets
• OpenSpec — Spec-driven development framework (OPSX conventions, artifact graphs, hard gates, delta specs)
• Superpowers — Claude Code plugin with <HARD-GATE> pattern, brainstorming workflow, skill-based development phases
• Compound Engineering — Four-phase workflow (Plan/Work/Review/Compound) with learning loop pattern

Related Skills

• agentv-eval-review — Lint and review AgentV eval files (invoke for eval-specific checks)
• agent-architecture-design — Design agent architectures from scratch