Maintenance — Convention Drift Audit

Audit the para-obsidian plugin for convention drift. Scans 8 domains: plugin registration, frontmatter schema, completion signals, allowed-tools accuracy, cross-references, brain routing coverage, references directory integrity, and observability gaps.

Not user-invocable — brain routes to it via keywords like "maintenance", "audit", "lint", "health check", "check conventions", "plugin health".

Step 0 — Load Conventions

Read the conventions knowledge base:

Read({ file_path: "<plugin-root>/skills/maintenance/references/conventions.md" })

Use these rules as the source of truth for all checks. The plugin root is plugins/para-obsidian relative to the repo root.

Step 1 — Plugin Registration Audit

Check that every component on disk has a matching entry in .claude-plugin/plugin.json.

1.1 — Discover components on disk:

Glob({ pattern: "plugins/para-obsidian/skills/*/SKILL.md" })
Glob({ pattern: "plugins/para-obsidian/commands/*.md" })
Glob({ pattern: "plugins/para-obsidian/agents/*.md" })

1.2 — Read plugin.json:

Read({ file_path: "<plugin-root>/.claude-plugin/plugin.json" })

1.3 — Compare:

For each component on disk, verify it has a matching entry in the appropriate plugin.json array (skills, commands, agents). Report:

• error: Component on disk missing from plugin.json
• warning: Entry in plugin.json pointing to non-existent path

Step 2 — Frontmatter Schema Audit

For each skills/*/SKILL.md, parse the YAML frontmatter and check required fields.

Required fields: name, description, user-invocable, allowed-tools

Additional checks:

• name must match the directory name
• User-invocable skills should have argument-hint (warning if missing)

Report:

• error: Missing required frontmatter field
• warning: User-invocable skill missing argument-hint

Step 3 — Completion Signal Audit

For each user-invocable skill, check if the SKILL.md body contains a SKILL_RESULT pattern.

Grep({ pattern: "SKILL_RESULT", path: "plugins/para-obsidian/skills/*/SKILL.md" })

Cross-reference with frontmatter user-invocable: true from Step 2.

Report:

• warning: User-invocable skill missing SKILL_RESULT completion signal

Step 4 — Allowed-Tools Accuracy Audit

For each skill, compare tools referenced in the SKILL.md body against the allowed-tools frontmatter declaration.

Detection patterns:

• MCP tool calls: para_create(, para_list(, para_fm_get(, etc.
• Full MCP names: mcp__plugin_para-obsidian_para-obsidian__<tool>
• Built-in tools: Read(, Write(, Edit(, Glob(, Grep(, Bash(, WebFetch(, WebSearch(
• Meta tools: AskUserQuestion(, Skill(, Task(

Exclusions:

• Tool references inside HTML comments (<!-- ... -->)
• Tool references in "Rules" or "Do NOT" sections (negative references)

Report:

• warning: Tool referenced in body but missing from allowed-tools
• info: Tool in allowed-tools but not referenced in body

Step 5 — Cross-Reference Audit

Scan all SKILL.md files for references to other skills, commands, or files, and verify they resolve.

Reference patterns to detect:

• Skill({ skill: "para-obsidian:<name>" → check skills/<name>/SKILL.md exists
• /para-obsidian:<name> → check skills/<name>/SKILL.md or commands/<name>.md exists
• [text](references/<file>) → check file exists relative to skill directory

Report:

• error: Referenced skill/command/file does not exist
• warning: Ambiguous reference that could be skill or command

Step 6 — Brain Routing Coverage Audit

Read the brain skill's routing table and verify coverage.

Read({ file_path: "<plugin-root>/skills/brain/SKILL.md" })

Checks:

• Every user-invocable skill (from Step 2) should appear in at least one routing table row
• Every skill/command referenced in the routing table should exist on disk

Report:

• warning: User-invocable skill not reachable via brain routing
• error: Routing table references non-existent skill/command

Step 7 — References Directory Audit

For each skill with a references/ subdirectory, verify files are actually referenced.

Glob({ pattern: "plugins/para-obsidian/skills/*/references/*" })

For each file found, check if the parent SKILL.md contains a reference to it (filename match in link text or path).

Report:

• warning: File in references/ not referenced from SKILL.md (possible dead content)
• error: SKILL.md references a file in references/ that doesn't exist

Step 8 — Observability Gaps Audit

Check that skills routed through brain have telemetry coverage.

8.1 — Verify brain-telemetry hook exists:

Read({ file_path: "<plugin-root>/hooks/hooks.json" })

Confirm a PostToolUse hook on Skill matcher exists and points to brain-telemetry.ts.

8.2 — Check SKILL_RESULT coverage:

Cross-reference skills in the brain routing table (Step 6) with completion signal coverage (Step 3). Skills in the routing table without SKILL_RESULT have an observability gap.

Report:

• warning: Brain-routed skill lacks SKILL_RESULT (telemetry will show undefined status)
• error: brain-telemetry hook not registered in hooks.json

Output Format

After all 8 domains are audited, produce a summary:

SKILL_RESULT:{"status":"<status>","skill":"maintenance","summary":"<N> domains audited, <M> issues found","issues":[<issue-list>]}

Status logic:

• ok — Zero errors, zero warnings
• partial — Zero errors, one or more warnings
• error — One or more errors

Issue format: Each issue in the array:

{
  "domain": "<1-8 domain name>",
  "severity": "error|warning|info",
  "file": "<relative path>",
  "message": "<what's wrong>",
  "fix": "<suggested remediation>"
}

Present the results to Nathan as a readable summary, not raw JSON. Group by domain, show severity with clear markers (e.g., [ERROR], [WARN], [INFO]). End with the SKILL_RESULT line for brain parsing.

Rules

1. Read-only audit. Never modify files. Only detect and report drift.
2. All paths relative to plugin root (plugins/para-obsidian/).
3. Conventions are the source of truth. Always read references/conventions.md first.
4. Deterministic. Same plugin state = same audit output. No LLM-dependent checks.
5. Fast. Use Glob and Grep for bulk operations. Read individual files only when needed.
6. Complete. Run all 8 domains every time. Don't short-circuit on first error.