Health Loop

---
allowed-tools: [Read, Write, Edit, Glob, Grep, Bash, TodoWrite, mcp__garden__garden_status, mcp__garden__garden_unresolved_links, mcp__garden__garden_orphans, mcp__garden__garden_validate, mcp__garden__garden_alias_map, mcp__garden__garden_files, mcp__garden__garden_tag_list, mcp__garden__garden_search, mcp__garden__garden_context, mcp__garden__garden_backlinks]
description: "Iterative audit-fix-reaudit loop — runs until the garden reaches a target health score or the user is satisfied"
---

# /dg:health-loop — Iterative Garden Health Loop

## Purpose

Run repeated audit → fix → reaudit cycles until the garden reaches a target health score. Designed to be used with `/goal` for autonomous multi-round improvement.

## Usage

/dg:health-loop [--target-unresolved N] [--rounds N] [--focus links|orphans|tags|all]

## Arguments

- `--target-unresolved N` — Stop when MCP unresolved link count drops below N (default: 300)
- `--rounds N` — Maximum rounds to run (default: 3)
- `--focus` — Constrain to a specific fix type (default: all)

## Loop Structure

Each round follows:

1. AUDIT → garden_status + garden_unresolved_links + garden_orphans
2. CLASSIFY → categorize issues into A/B/C fix types (see FindBrokenLinks.md)
3. FIX → apply highest-ROI fixes for this round
4. VERIFY → garden_status to confirm improvement
5. DECIDE → continue if target not reached and rounds remain

## Round Priorities

**Round 1 — Structural fixes (highest ROI)**
- Add aliases to top-N broken targets where file exists
- Fix broken absolute paths in source files
- Add missing domain index files
- Fix cross-domain index link formatting

**Round 2 — Content stubs (medium ROI)**
- Create stubs for concepts referenced 3+ times
- Add cross-domain bridges to isolated domain indexes
- Bridge orphaned notes to their domain index

**Round 3 — Deep cleanup (low ROI, high coverage)**
- Create stubs for concepts referenced 2+ times
- Consolidate near-duplicate tags
- Fix frontmatter issues (missing description, date, tags)

## Execution

### Step 1: Initial Audit

Call garden_status → record baseline (unresolvedLinks, totalNotes) Call garden_unresolved_links({ limit: 50, minCount: 2 }) → get top broken targets Call garden_orphans({ excludeIndexes: true }) → get orphaned notes Call garden_tag_list → look for near-duplicate tags

Report:

Garden Health — Round N Audit

Metric	Value
Unresolved links (MCP)	N
Orphaned notes	N
Total notes	N
Tags	N

Top 10 broken targets:

[list from garden_unresolved_links]

Top orphans:

[list from garden_orphans, grouped by domain]

### Step 2: Classify & Fix

Load `DgNotes/Workflows/FindBrokenLinks.md` for fix classification logic.

For each broken target:
1. Check if file exists: `Bash: find content/ -iname "target.md"`
2. If yes → alias fix (modify frontmatter)
3. If path-based → path fix (edit source files)
4. If missing → stub creation (if 3+ references)

For orphans: add links from the domain index or related notes.

For tags: merge near-duplicates, remove singletons with 1 occurrence.

### Step 3: Verify & Decide

Call garden_status → compare unresolvedLinks to target

**Continue if**: unresolvedLinks > target AND rounds remaining AND new fixable items found
**Stop if**: target reached OR no more fixable items OR max rounds reached

### Step 4: Final Report

Garden Health — Final State

Metric	Before	After	Delta
Unresolved (MCP)	N	N	-N
Notes	N	N	+N
Rounds run	—	N	—

Changes made:

• N alias additions
• N path corrections
• N new stubs created
• N orphans bridged
• N tags consolidated

Remaining gaps (content work needed):

[list of concepts still missing with reference counts]

## Integration with /goal

Combine with `/goal` for autonomous multi-round improvement:

/goal Run /dg:health-loop until unresolved < 300, then stage all changes for review

The stop hook will block until the target is reached.

## Skill Context

Load `DgNotes` before running fixes. Load `DgTags` before any tag consolidation.

## Anti-patterns to Avoid

- Do NOT create stubs for 1x references (noise outweighs value)
- Do NOT change content of existing notes unless improving frontmatter
- Do NOT commit — always stage for review
- Do NOT run more than 5 rounds (diminishing returns after round 3)