Imported from wan-huiyan/claude-ecosystem-hygiene (plugins/ecosystem-audit/SKILL.md). Install upstream with npx skills add wan-huiyan/claude-ecosystem-hygiene --skill ecosystem-audit. Copyright stays with the author.

Ecosystem Audit

A comprehensive audit of the entire ~/.claude/ ecosystem that scans 9 categories of persistent artifacts, scores each on a health scale, and produces an interactive HTML dashboard with prioritized cleanup recommendations.

When to Use

• User wants to know what skills are installed and which are actually used
• User wants to clean up their Claude Code environment
• User asks about ecosystem health, dormant artifacts, or storage bloat
• Monthly hygiene check (recommend running every 30 days)
• Before/after installing a batch of new skills
• When memory-hygiene or schliff:doctor findings suggest broader issues

Architecture

This skill orchestrates parallel subagents for speed. The audit runs in 3 phases:

Phase 1: Scan (parallel subagents)
  ├── Skills: parse JSONL logs for invocations, classify by domain
  ├── Memory: check all project dirs, apply memory-hygiene thresholds
  ├── Handoffs: count, classify lifecycle state, find orphans
  └── ADRs + Docs: read status fields, check links, classify

Phase 2: Score (sequential)
  ├── Calculate health % per category
  ├── Apply lifecycle scoring for worktrees
  ├── Cross-reference with memory-hygiene thresholds
  └── Generate prioritized recommendations (P0/P1/P2)

Phase 3: Report (write files)
  ├── Markdown summary → docs/handoffs/ecosystem_audit_report.md
  └── Interactive HTML → docs/handoffs/ecosystem_audit_report.html

Phase 1: Scan

Launch 4 parallel subagents, each responsible for one scan domain. Each subagent reports structured findings back to the orchestrator.

1A. Skill Usage Scan

Parse JSONL session logs to determine which skills are actually invoked.

Data source: ~/.claude/projects/<dir>/<uuid>.jsonl

What to extract:

• Lines where message.content[].name == "Skill" — these are invocations
• Extract input.skill (skill name), input.args, and timestamp
• Cross-reference with toolUseResult.commandName and toolUseResult.success

Classification rules:

• Active: invoked in the session log window (default: 30 days)
• Relevant-Dormant: not invoked but domain-aligned with user's work (check user profile in memory files for role/domain context)
• Niche-Dormant: specialized domain unlikely to be needed (bio/chem databases, quantum computing, wet-lab integrations, astrophysics, game dev, embedded systems)
• DevOps-Dormant: IaC/CI/CD tools (terraform, helm, k8s, ansible, dockerfile, jenkins, github-actions generators and validators)
• Not-a-Skill: workspace directories, cloned repos, .DS_Store, README files
• Deprecated: SKILL.md explicitly says "DEPRECATED" or "merged into"

Also check:

• Duplicate/overlap clusters: skills with similar trigger phrases
• Adware: skills that say "ALWAYS run" or inject into every session
• Skills without SKILL.md files

1B. Memory & Lessons Scan

Delegation contract with memory-hygiene (single source of truth).

memory-hygiene (v3.0+) is the authoritative spec for the Memory tier. To prevent drift, the Memory subagent does NOT re-implement memory-hygiene's detection logic. Instead:

1. Invoke memory-hygiene Phase 1 (Discover) only — do not execute Phase 4 fixes.
2. Consume its structured audit report (the headings under ## Memory Hygiene Audit) as the input to ecosystem-audit's Memory score.
3. Version-pin check: read memory-hygiene's SKILL.md frontmatter. If the major version differs from what this skill was calibrated against (currently v3.3 — Phase 1 Discover + Phase 1h ADRs/MADR 4.0 interface verified compatible 2026-06-01), fail loudly and ask the user to re-sync ecosystem-audit before scoring.

If for any reason memory-hygiene cannot run (missing, errored, version mismatch), render the Memory axis as N/A in the radar chart (see Phase 3) and list blockers. Never fabricate a score.

Scan targets covered by memory-hygiene (full tier map T0→T3):

Staleness signals (memory-hygiene §1d — all four, not just dates):

• Broken references (grep codebase for identifier; not found = stale)
• Relative dates ("last week", "recently") that should be absolute
• Codebase contradictions (memory says lib X; package.json says otherwise)
• Contradicted lessons (global vs project rules conflict)

Agency-aware staleness: read user_role.md before thresholding. Agency / multi-client users measure dormancy in calendar time across portfolio; single-long-project users measure in session count within the project.

Per-file cross-cutting checks (detection only; resolution belongs to memory-hygiene):

• Orphan detection (file exists but not in MEMORY.md index)
• Cross-project duplicates (same feedback file in multiple projects) — report clusters; memory-hygiene Phase 4 owns keep/move/merge decisions

1C. Handoffs & Session Docs Scan

Count and classify all session handoff documents.

Scan locations:

• <project>/docs/handoffs/ — handoff and prompt files
• <project>/.claude/worktrees/*/docs/handoffs/ — worktree duplicates

Classification:

• Current: references in-progress or planned work
• Historical: work completed, value is purely archival
• Orphaned prompt: *_prompt.md whose session already happened or never started
• Worktree duplicate: handoff file inside a stale worktree (pure waste)

1D. ADRs, Plans, Reviews, Findings, Tasks Scan

Read and classify structured project documentation.

ADR checks (per memory-hygiene MADR 4.0):

• Status field validation (Accepted/Proposed/Superseded/Resolved)
• Bidirectional supersession links
• README.md link integrity (do filenames match?)
• Duplicate content detection
• Numbering gaps and sequence issues

Plans/Reviews/Findings/Tasks:

• Classify as Active / Historical / Stale
• Check for completed items still marked as in-progress

Phase 2: Score

Health Percentage Calculation

Each category gets a health score from 0-100%. The methodology differs by category because "healthy" means different things for different artifact types.

Skills: active_count / total_installed * 100 Simple utilization rate. Active = invoked in the log window.

Memory: Weighted composite of sub-checks, computed from memory-hygiene's Discover report:

• MEMORY.md line count vs threshold (weight: 25%)
• Frontmatter compliance rate, incl. type enum validity (weight: 15%)
• Axiom health — count ≤12 AND zero Phase items in axioms.md (weight: 15%)
• Phase-template hygiene — templates exist, no dead paths: globs (weight: 10%)
• Staleness rate of topic files (all 4 signals, agency-aware) (weight: 20%)
• Session compression backlog (weight: 15%)

If any sub-check is un-computable (missing input, memory-hygiene unavailable), set the Memory axis to N/A rather than averaging over a partial signal.

Handoffs: current_count / total_count * 100 Only current handoffs are "healthy" — historical ones should be archived.

ADRs: active_count / total_count * 100 Active = still governing current behavior (Accepted status, not superseded).

Docs (plans + reviews + findings + tasks): active_count / total_count * 100

Worktrees — Lifecycle Score (not binary active/stale): Each worktree is scored individually based on its lifecycle state, then averaged:

• EXPECTED (unmerged branch, ≤7 days old): 100%
• ACCEPTABLE (unmerged branch, 7-14 days old): 75%
• NEEDS_CLEANUP (branch already merged): 25%
• ABANDONED (unmerged branch, >14 days old): 0%

If 0 worktrees exist, score is 100% (nothing to manage, nothing mismanaged).

Priority Classification

Actions are classified by urgency:

• P0 (Do Now): Security risks (exposed keys), broken automation, data integrity issues
• P1 (This Week): Stale artifact cleanup, threshold violations, worktree removal
• P2 (This Month): Archival, consolidation, lessons triage, ongoing monitoring

T1 Promotion Criteria (v1.2.0)

T1 means evidence-backed load-bearing, not merely "frequently referenced." High reference count alone does not predict A/B contribution — the v3 layered ablation (240 cells, n=15) found that lessons.md (highest-ref layer) had a Δ within noise, while skills+plugins (lower ref count) cleanly separated from the noise floor (pitfall-avoided rate 80% → 43% when stripped).

To promote a layer to T1, require one of:

If neither condition is met, score_health.py emits a t1_warnings entry. The audit report must surface this warning — a silently-promoted T1 erodes trust in the tier system.

Cross-reference: ab-harness v1.2.0 §"Noise floor: design a no-op cell" and §"C11 saturation: everything-stripped ties the no-op control."

Correctness-vs-Latency Annotation (v1.2.0)

The v3 finding that skills+plugins (cell C4) was the fastest cell on generic tasks (7.2 turns / $0.30 vs C0's 18.8 turns / $0.57) means skills add latency on routine work even when they help on pitfall-prone tasks. Do not blanket-recommend "install more skills."

For each skill-type layer, annotate with:

Mismatch pattern: ref_count > 10 AND delta_vs_noop_se < 1.0. This is the v3 signature of a skill that's frequently seen but adds latency on generic tasks without reducing pitfalls on the user's measured workload.

Skills with trigger_surface_match == "mismatched" feed the new "Skills with mismatched trigger surface" report section (see Phase 3).

Updated recommendation engine logic:

• Mine recent session prompts for dominant task-type patterns before recommending any skill
• Recommend only skills whose trigger surface matches the user's common pitfall patterns
• Flag high-ref / low-A/B-signal skills as "noise" — they add latency without pitfall benefit
• Do NOT blanket-recommend installing more skills

Phase 3: Report

Generate two output files:

Markdown Report

Save to docs/handoffs/ecosystem_audit_report.md with:

• Executive summary with key metrics
• Per-category findings table
• Prioritized recommendations
• Cleanup script (ready-to-run shell commands)

Interactive HTML Report

Save to docs/handoffs/ecosystem_audit_report.html with:

• Read assets/report_template.html for the base template
• The template uses a dark terminal theme (GitHub-dark background, Inter headings, Fira Code / IBM Plex Mono data, teal accent #2DD4A8)
• Populate the template by replacing data placeholders with audit results
• Features: sticky sidebar nav, tabbed data views, radar chart (6 axes), animated counters, hover tooltips, bar charts, priority-coded action cards

Skills with Mismatched Trigger Surface (v1.2.0)

When trigger_surface_mismatches is non-empty, add a section to both markdown and HTML reports listing each mismatched skill with:

• Skill name
• ref_count (how often it appeared in session logs)
• delta_vs_noop_se (A/B signal strength vs no-op control)
• latency_cost (estimated turn overhead on generic tasks)
• Recommendation: review trigger description, narrow scope to pitfall-prone tasks only, or deprecate

If no mismatches, omit the section entirely.

Also emit a T1 warning block when t1_warnings is non-empty, listing each affected layer by name with the remediation options (add ab_evidence or set evidence: "reference-count-only").

Radar Chart Data

The radar chart shows 6 axes: Skills, Memory, Handoffs, ADRs, Docs, Worktrees. Each axis shows the health percentage from Phase 2 scoring.

Two overlaid polygons:

• Teal (active/healthy): the health score per category
• Red (dormant/stale): the complement (100% - health)

Never fabricate a score. If any sub-check for an axis could not be computed (e.g., memory-hygiene unavailable, user_role.md missing, version mismatch), render that axis as N/A with a hatched pattern and list blockers in the report's "Coverage" section. A partial average is worse than a missing reading.

Complementary Skills

This skill is designed to work alongside, not replace:

• ab-harness (v1.2.0+): Counterfactual A/B + layered-ablation harness. Run to generate the ab_evidence fields this skill uses for T1 validation and latency annotation. Key sections: §"Noise floor: design a no-op cell" (threshold calibration) and §"C11 saturation: everything-stripped ties the no-op control" (why all-stripped ≠ signal).
• schliff:doctor: Structural quality scoring of individual SKILL.md files (7-dimension scoring). Run after ecosystem-audit identifies dormant skills to assess whether they're worth keeping based on quality.
• schliff:mesh: Cross-skill trigger overlap detection. Run after ecosystem-audit identifies overlap clusters to get detailed conflict analysis.
• memory-hygiene: Deep persistent knowledge store audit. Run after ecosystem-audit flags memory health issues to get specific fixes (tiered loading, contradiction detection, axiom promotion).
• skill-portfolio-audit: Portfolio-wide README/badge standardization. Run after ecosystem-audit cleanup to polish the remaining skills.

Edge Cases

• No session logs: If ~/.claude/projects/ has no JSONL files, skip skill usage analysis and note "no session data available" in the report. Other categories still work.
• No worktrees: Score is 100% (clean state). Note "no worktrees found" in report.
• Single project: Still audit global files (axioms, lessons, skills).
• Very large ecosystems (>500 skills): Use subagent parallelism aggressively. Batch skill classification into groups of 50.