Dependency Checker

Analyze the Agent Alchemy plugin ecosystem to detect dependency issues, broken paths, orphaned components, and documentation drift. Produces a health report with severity-ranked findings.

CRITICAL: Complete ALL 5 phases. The workflow is not complete until Phase 5: Report is finished. After completing each phase, immediately proceed to the next phase without waiting for user prompts.

Critical Rules

AskUserQuestion is MANDATORY

IMPORTANT: You MUST use the AskUserQuestion tool for ALL questions to the user. Never ask questions through regular text output.

• Report presentation -> AskUserQuestion
• Action selection -> AskUserQuestion
• View mode selection -> AskUserQuestion

Text output should only be used for:

• Displaying progress updates between phases
• Presenting intermediate analysis summaries (inventory counts, graph stats)
• Phase transition markers

If you need the user to make a choice or provide input, use AskUserQuestion.

NEVER do this (asking via text output):

Would you like to view findings by severity or by plugin group?
1. By severity
2. By plugin group

ALWAYS do this (using AskUserQuestion tool):

AskUserQuestion:
  questions:
    - header: "View Mode"
      question: "How would you like to view the findings?"
      options:
        - label: "By severity"
          description: "Group findings from critical to low"
        - label: "By plugin group"
          description: "Filter findings to a specific plugin group"
      multiSelect: false

Plan Mode Behavior

CRITICAL: This skill performs an interactive analysis workflow, NOT an implementation plan. When invoked during Claude Code's plan mode:

• DO NOT create an implementation plan for how to build the analysis
• DO NOT defer analysis to an "execution phase"
• DO proceed with the full analysis workflow immediately
• DO write report files as normal if --report-file is specified

Phase Overview

Execute these phases in order, completing ALL of them:

1. Load & Discover — Parse arguments, load settings, build component inventory
2. Build Dependency Graph — Parse every component file to extract dependency edges
3. Analyze — Run 7 detection passes over the dependency graph
4. Cross-Reference Documentation — Compare graph against CLAUDE.md/README docs for drift
5. Report — Present findings interactively; optionally export

Phase 1: Load & Discover

Goal: Parse arguments, load settings, build a complete component inventory of the plugin ecosystem.

Step 1: Parse Arguments

Parse $ARGUMENTS for:

• --plugin <group> — Filter to one plugin group by short directory name (e.g., core-tools). Default: analyze all groups.
• --verbose — Include healthy/passing entries in the report, not just issues. Default: false.
• --report-file <path> — Export the full report as a markdown file to the given path. Default: none (interactive only).

Set variables:

• FILTER_GROUP from --plugin value (default: null = all groups)
• VERBOSE_MODE from --verbose flag (default: false)
• REPORT_FILE from --report-file value (default: null)

Step 2: Load Settings

Read settings from .claude/agent-alchemy.local.md if it exists. Look for the plugin-tools.dependency-checker section in the YAML frontmatter.

If the settings file doesn't exist or the section is missing, use defaults.

Step 3: Load Marketplace Registry

Read the marketplace registry:

Read: ${CLAUDE_PLUGIN_ROOT}/../../.claude-plugin/marketplace.json

Build a registry map: { plugin_name -> { version, source_dir, description } } for each entry in the plugins array. The source_dir is derived from the source field (strip leading ./).

Step 4: Discover Components

For each plugin group directory under claude/ (or only FILTER_GROUP if set), enumerate all components using Glob:

Skills:

Glob: claude/{group}/skills/*/SKILL.md

For each found skill, read its YAML frontmatter to extract:

• name, description, user-invocable, disable-model-invocation
• allowed-tools list
• skills list (if present — for agent-like skill composition)

Agents:

Glob: claude/{group}/agents/*.md

For each found agent, read its YAML frontmatter to extract:

• name, description, model
• tools list
• skills list (these are skill bindings that must resolve to real skills)

Shared references (plugin-level):

Glob: claude/{group}/references/**/*.md

Skill-local references:

Glob: claude/{group}/skills/*/references/**/*.md

Hooks:

Glob: claude/{group}/hooks/hooks.json

If found, read and parse the JSON to extract hook entries.

Hook scripts:

Glob: claude/{group}/hooks/*.sh

Step 5: Display Inventory Summary

Display a text summary of what was discovered:

[Phase 1/5] Plugin Ecosystem Inventory

| Group | Skills | Agents | Shared Refs | Skill Refs | Hooks | Scripts |
|-------|--------|--------|-------------|------------|-------|---------|
| core-tools | N | N | N | N | N | N |
| dev-tools | N | N | N | N | N | N |
| ... | ... | ... | ... | ... | ... | ... |
| **Total** | **N** | **N** | **N** | **N** | **N** | **N** |

Phase 2: Build Dependency Graph

Goal: Parse every component file to extract all dependency edges, building a directed graph of the ecosystem.

Step 1: Define Edge Types

The graph has the following edge types:

Step 2: Extract Edges from Skills

For each SKILL.md file discovered in Phase 1, scan the full file content (not just frontmatter) using these regex patterns:

Cross-plugin skill loads:

\$\{CLAUDE_PLUGIN_ROOT\}/\.\./([^/]+)/skills/([^/]+)/SKILL\.md

Creates edge: skill-loads-skill-cross from current skill to {group}:{skill_name}

Same-plugin skill loads:

\$\{CLAUDE_PLUGIN_ROOT\}/skills/([^/]+)/SKILL\.md

Match only if the path does NOT contain /../ before it. Creates edge: skill-loads-skill from current skill to {skill_name} within the same plugin.

Shared reference loads (same plugin):

\$\{CLAUDE_PLUGIN_ROOT\}/references/([^\s"'`)+]+\.md)

Creates edge: skill-loads-shared-ref

Skill-local reference loads:

\$\{CLAUDE_PLUGIN_ROOT\}/skills/([^/]+)/references/([^\s"'`)+]+\.md)

Creates edge: skill-loads-local-ref

Cross-plugin reference loads:

\$\{CLAUDE_PLUGIN_ROOT\}/\.\./([^/]+)/(skills/[^/]+/)?references/([^\s"'`)+]+\.md)

Creates edge: skill-loads-cross-ref

Agent spawning:

subagent_type[:\s]*["']?([^"'\s,}]+)

Creates edge: skill-spawns-agent from current skill to the resolved agent name.

Registry reads:

\$\{CLAUDE_PLUGIN_ROOT\}/\.\./\.\./\.claude-plugin/

Creates edge: skill-reads-registry

Step 3: Extract Edges from Agents

For each agent .md file, parse its YAML frontmatter and extract the skills: list.

Each entry in the skills: list creates an agent-binds-skill edge. The skill name must be resolved to a skill within the same plugin group as the agent (since agent frontmatter skill bindings are group-local).

Step 4: Extract Edges from Hooks

For each hooks.json file, parse the JSON and scan all command strings for script references:

\$\{CLAUDE_PLUGIN_ROOT\}/hooks/([^\s"']+)

Creates edge: hook-runs-script from the hook config to the referenced script file.

Step 5: Display Graph Summary

[Phase 2/5] Dependency Graph Built

Nodes: N total (N skills, N agents, N shared refs, N skill refs, N hooks, N scripts)
Edges: N total
  - skill-loads-skill: N (same-plugin: N, cross-plugin: N)
  - skill-loads-ref: N (shared: N, local: N, cross-plugin: N)
  - skill-spawns-agent: N
  - skill-reads-registry: N
  - agent-binds-skill: N
  - hook-runs-script: N

Phase 3: Analyze (7 Detection Passes)

Goal: Run 7 detection passes over the dependency graph to find issues. Each pass produces findings with a severity level.

Filter findings by severity-threshold setting — only retain findings at or above the configured threshold.

Pass 1: Circular Dependencies (Critical)

Run DFS cycle detection on the skill-loads-skill and skill-loads-skill-cross edges (the skill-to-skill subgraph).

Algorithm:

1. For each skill node, perform DFS tracking the visit stack
2. If a node is encountered that is already in the current visit stack, a cycle is detected
3. Record the full cycle path: skill_A -> skill_B -> ... -> skill_A

Each cycle found is a Critical severity finding.

Pass 2: Missing Dependencies (High)

For every edge in the graph, verify that the target node exists on disk:

• skill-loads-skill / skill-loads-skill-cross: Check the target SKILL.md file exists
• skill-loads-shared-ref / skill-loads-local-ref / skill-loads-cross-ref: Check the target .md file exists
• skill-spawns-agent: Resolve the agent name to a file. Agent names may be:
  • Bare names (e.g., "code-explorer") — look in the same plugin group's agents/ directory
  • Qualified names (e.g., "agent-alchemy-core-tools:code-explorer") — map marketplace name to source dir, then look in that group's agents/
• agent-binds-skill: Check the skill exists in the same plugin group's skills/ directory
• hook-runs-script: Check the script file exists

Each missing target is a High severity finding. Include the source file, expected target path, and edge type.

Pass 3: Broken Cross-Plugin Paths (Medium)

Scan all SKILL.md and agent .md files for path anti-patterns:

Anti-pattern 1 — Marketplace name in path:

agent-alchemy-[a-z-]+/

Paths should use short directory names (e.g., core-tools), not marketplace names (e.g., agent-alchemy-core-tools).

Anti-pattern 2 — Hardcoded absolute paths:

/Users/|/home/|/tmp/.*claude/

Paths should use ${CLAUDE_PLUGIN_ROOT} variable, not absolute paths.

Anti-pattern 3 — Incorrect nesting depth:

\$\{CLAUDE_PLUGIN_ROOT\}/\.\./\.\./\.\./

Triple ../ or deeper should not appear (maximum is ../../ for registry access).

Each finding is Medium severity.

Pass 4: Orphaned Components (Low)

Find components with zero inbound edges (nothing references them):

• Skills: Only flag non-user-invocable skills with zero inbound edges. User-invocable skills are entry points and are expected to have no inbound edges.
• Agents: Flag agents with zero skill-spawns-agent inbound edges (no skill spawns them) AND zero agent-binds-skill outbound edges isn't sufficient — check that nothing spawns them.
• Shared references: Flag .md files in references/ directories with zero skill-loads-shared-ref or skill-loads-cross-ref inbound edges.
• Skill-local references: Flag .md files in skills/*/references/ with zero skill-loads-local-ref inbound edges.
• Hook scripts: Flag .sh files in hooks/ with zero hook-runs-script inbound edges.

Each orphaned component is a Low severity finding. Note: files in references/adapters/ subdirectories may be dynamically loaded based on arguments and should be flagged with a caveat noting they may be loaded dynamically.

Pass 5: Agent-Skill Mismatches (High)

For each agent's skills: list in its frontmatter:

1. Unresolvable skill names: The skill name doesn't match any skill in the same plugin group. Flag as High severity with the agent name, unresolvable skill name, and a suggestion of possible matches (fuzzy).

2. Ambiguous cross-plugin bindings: If the skill name matches skills in multiple plugin groups but no qualifier is provided, flag as High severity.

Note: Agent skills: bindings in Claude Code are resolved within the same plugin group. A skill name in an agent's skills: list refers to a skill directory name under the agent's own plugin group's skills/ directory.

Pass 6: Marketplace Consistency (Medium)

Compare the marketplace registry against actual directories:

1. Missing registry entries: Directories under claude/ that look like plugin groups (contain skills/ or agents/ subdirectories) but have no corresponding entry in marketplace.json. Exclude .claude-plugin/ itself.

2. Stale registry entries: Entries in marketplace.json whose source path doesn't resolve to an existing directory.

3. Source path mismatches: Registry entry source field doesn't match the actual directory name.

Each finding is Medium severity.

Pass 7: Hook Integrity (Low)

For each hooks.json file:

1. Invalid matcher patterns: Check that matcher values reference valid Claude Code tool names. Known valid tools include: Read, Write, Edit, Bash, Glob, Grep, WebFetch, WebSearch, Task, AskUserQuestion, NotebookEdit, SendMessage, TodoWrite. Matchers using | for OR are valid. Flag unrecognized tool names.

2. Missing scripts: Already covered by Pass 2 (missing dependencies), but if the script reference uses a non-standard path pattern (not ${CLAUDE_PLUGIN_ROOT}/hooks/), flag it.

3. Timeout values: If timeout is set, verify it's a positive number. Flag zero or negative timeouts.

4. Invalid hook types: Verify type field is a recognized value (command, prompt).

Each finding is Low severity.

Step: Display Analysis Summary

[Phase 3/5] Analysis Complete

| Pass | Check | Severity | Findings |
|------|-------|----------|----------|
| 1 | Circular dependencies | Critical | N |
| 2 | Missing dependencies | High | N |
| 3 | Broken cross-plugin paths | Medium | N |
| 4 | Orphaned components | Low | N |
| 5 | Agent-skill mismatches | High | N |
| 6 | Marketplace consistency | Medium | N |
| 7 | Hook integrity | Low | N |
| | **Total** | | **N** |

Phase 4: Cross-Reference Documentation

Goal: Compare the dependency graph against documentation files (CLAUDE.md, plugin README files) to detect drift. Skip this phase if the check-docs-drift setting is false.

If check-docs-drift is false, display:

[Phase 4/5] Documentation cross-referencing skipped (check-docs-drift: false)

Then proceed to Phase 5.

Check 1: CLAUDE.md Plugin Inventory Table

Read the root CLAUDE.md file and parse the Plugin Inventory table.

For each row in the table:

• Skill count: Compare the comma-separated skill names in the "Skills" column against the actual skills discovered in Phase 1. Flag missing or extra skills.
• Agent count: Compare the comma-separated agent names in the "Agents" column against actual agents. Flag missing or extra agents.
• Version: Compare the version in the table against the version in marketplace.json. Flag mismatches.

Each drift finding is Medium severity with type docs-drift.

Check 2: CLAUDE.md Composition Chains

Read the Key Skill Composition Chains section in CLAUDE.md. For each documented chain:

• Verify the source skill exists
• Verify the target skill/agent exists
• Check that the edge exists in the dependency graph (the source actually loads/spawns the target)

Flag documented chains that don't match the actual graph as Medium severity docs-drift.

Do NOT flag undocumented chains — CLAUDE.md only documents "key" chains, not all of them.

Check 3: CLAUDE.md Critical Plugin Files Table

Read the Critical Plugin Files table. For each row:

• Verify the file exists at the listed path
• Get the actual line count using Bash: wc -l
• Compare against the documented line count. Apply the line-count-tolerance setting (default 10%). Flag if the actual count differs by more than the tolerance percentage.

Each finding is Low severity with type docs-drift.

Check 4: Plugin README Files

For each plugin group, read its README.md (if it exists):

• Parse the skills table and compare listed skills against actual skills in that group
• Parse the agents table and compare listed agents against actual agents in that group

Flag missing or extra entries as Medium severity with type docs-drift.

Display Documentation Summary

[Phase 4/5] Documentation Cross-Reference

| Source | Check | Findings |
|--------|-------|----------|
| CLAUDE.md | Plugin Inventory | N |
| CLAUDE.md | Composition Chains | N |
| CLAUDE.md | Critical Files | N |
| README files | Skill/Agent tables | N |
| **Total** | | **N** |

Phase 5: Report

Goal: Present findings interactively; optionally export as a markdown file.

Step 1: Compute Health Score

health_score = (components_without_issues / total_components) × 100

Where total_components is the count of all skills, agents, references, and scripts discovered in Phase 1. A component "has issues" if it appears in any finding from Phase 3 or Phase 4.

Determine health indicator:

• 90-100%: "Healthy"
• 70-89%: "Needs Attention"
• 50-69%: "Outdated"
• 0-49%: "Critical Issues"

Step 2: Auto-Export (if --report-file)

If REPORT_FILE is set, write the full report before presenting the interactive menu. The markdown report should include:

1. Header with ecosystem name, date, health score
2. Inventory summary table from Phase 1
3. Dependency graph statistics from Phase 2
4. Dependency graph diagram — Load Read ${CLAUDE_PLUGIN_ROOT}/../core-tools/skills/technical-diagrams/SKILL.md, then generate a Mermaid flowchart showing plugin groups as subgraphs and key cross-plugin dependency edges. Limit to 15-20 nodes max. Highlight nodes involved in findings using danger (critical) or warning (high) classDef styles.
5. All findings grouped by severity (critical → low), each with:
   • Component involved
   • Issue type and description
   • Expected vs actual (where applicable)
   • Suggested fix
6. Documentation drift findings from Phase 4 (if run)
7. Health score and recommendation

Write: {REPORT_FILE}

Display: Report exported to: {REPORT_FILE}

Step 3: Present Interactive Report

Present the summary via AskUserQuestion:

AskUserQuestion:
  questions:
    - header: "Dependency Health Report"
      question: |
        ## Plugin Ecosystem — {health_indicator}

        **Health Score: {health_score}%** ({components_without_issues}/{total_components} components clean)

        | Severity | Count |
        |----------|-------|
        | Critical | N |
        | High | N |
        | Medium | N |
        | Low | N |

        {total_findings} findings across {groups_with_issues} plugin groups.

        How would you like to explore the results?
      options:
        - label: "View all findings"
          description: "Findings grouped by severity, from critical to low"
        - label: "View by plugin group"
          description: "Select a plugin group to see its findings"
        - label: "View dependency graph"
          description: "Text-based graph showing all dependency edges"
        - label: "Done"
          description: "Exit the dependency checker"
      multiSelect: false

Step 4: Handle User Selections

View all findings: Display all findings grouped by severity (critical first), each showing:

• Component: The file or component involved
• Issue: Description of the problem
• Expected: What was expected (if applicable)
• Actual: What was found (if applicable)
• Fix: Suggested remediation

After displaying, loop back to the interactive menu (Step 3).

View by plugin group: Present a follow-up AskUserQuestion with the plugin groups as options. After the user selects a group, display only findings involving components in that group. Then loop back to the interactive menu.

View dependency graph: Display a text-based representation of the dependency graph:

core-tools/
  skills/
    deep-analysis
      -> spawns: code-explorer (core-tools), code-synthesizer (core-tools)
      <- loaded by: codebase-analysis (core-tools), feature-dev (dev-tools), docs-manager (dev-tools), create-spec (sdd-tools)
    codebase-analysis
      -> loads: deep-analysis (core-tools)
      -> loads ref: report-template.md, actionable-insights-template.md
  agents/
    code-explorer
      -> binds: project-conventions, language-patterns
      <- spawned by: deep-analysis
...

Also generate a Mermaid flowchart version showing plugin groups as subgraphs with color-coded finding severity (danger classDef for critical, warning for high, primary/secondary for clean nodes). Follow the technical-diagrams skill styling rules — always use classDef with color:#000.

After displaying, loop back to the interactive menu.

Done: Exit the workflow. Display a final one-line summary:

Dependency check complete. {total_findings} findings ({critical_count} critical, {high_count} high).

Error Handling

No Plugin Groups Found

If no plugin group directories are found under claude/ (or the filtered group doesn't exist):

1. Display an error message listing what was searched
2. If --plugin was used, suggest valid group names from the marketplace registry
3. Exit the workflow

Marketplace Registry Missing

If marketplace.json doesn't exist or can't be parsed:

1. Log a warning: Warning: marketplace.json not found or invalid. Skipping marketplace consistency checks (Pass 6).
2. Continue with all other passes — the registry is only needed for Pass 6 and documentation drift checks.

Malformed Component Files

If a SKILL.md or agent .md file can't have its YAML frontmatter parsed:

1. Log a warning for that specific file
2. Skip edge extraction for that file
3. Add a Medium severity finding: "Malformed frontmatter: {file_path}"
4. Continue processing other files

Large Ecosystem Performance

The analysis reads every component file. For performance:

• Read frontmatter-only when only frontmatter fields are needed (agents)
• Read full file content when body scanning is needed (skills — for subagent_type and path patterns)
• Use Glob results from Phase 1 to avoid redundant file discovery

Finding Schema

Every finding produced by Phase 3 and Phase 4 follows this structure:

{
  id: "{pass_number}.{index}",
  severity: "critical" | "high" | "medium" | "low",
  type: "circular-dep" | "missing-dep" | "broken-path" | "orphaned" |
        "agent-skill-mismatch" | "marketplace-inconsistency" |
        "hook-integrity" | "docs-drift",
  component: "{group}/{component_type}/{name}",
  message: "Human-readable description of the issue",
  expected: "What was expected (optional)",
  actual: "What was found (optional)",
  fix: "Suggested remediation"
}

Dependency Patterns Reference

These are the regex patterns used in Phase 2 for edge extraction. They are provided here as a quick reference — the authoritative patterns are in Phase 2 above.