health

Runtime Configuration (Step 0 — before any processing)

Read these files to configure domain-specific behavior:

1. ops/derivation-manifest.md — vocabulary mapping, folder names, platform hints

   • Use vocabulary.notes for the notes folder name
   • Use vocabulary.inbox for the inbox folder name
   • Use vocabulary.note for the note type name in output
   • Use vocabulary.topic_map for MOC/topic map references
   • Use vocabulary.topic_maps for plural form

2. ops/config.yaml — processing depth, thresholds

3. Three-space reference — ${CLAUDE_PLUGIN_ROOT}/reference/three-spaces.md for boundary rules (load only for full and three-space modes)

4. Templates — read template files to understand required schema fields for validation

If these files don't exist (pre-init invocation or standalone use), use universal defaults:

• notes folder: notes/
• inbox folder: inbox/
• topic map: topic maps in notes/

---

EXECUTE NOW

Target: $ARGUMENTS

Parse the invocation mode immediately:

Input	Mode	Categories Run
empty or quick	Quick	1 (Schema), 2 (Orphans), 3 (Links)
full	Full	All 8 categories
three-space	Three-Space	5 (Three-Space Boundaries) only

Execute these steps:

1. Detect mode from arguments
2. Scan the vault — inventory all note files, {vocabulary.topic_map} files, inbox items, ops files
3. Run each applicable diagnostic category in order (1-8)
4. Classify each result as PASS, WARN, or FAIL using the thresholds below
5. Surface condition-based maintenance signals (check against threshold table)
6. Generate the health report with specific file paths, counts, and recommended actions
7. Write report to ops/health/YYYY-MM-DD-report.md

START NOW. Reference below explains each diagnostic category in detail.

Platform Adaptation

Checks adapt to what the platform supports:

• If semantic search (qmd) is not configured, skip semantic-dependent checks and note their absence
• If hooks are not available, note that validation is convention-only (no automated enforcement)
• If self/ directory does not exist (disabled by config), skip self-space checks but verify ops/ absorbs self-space content correctly

Report what CAN be checked, not what the platform lacks.

---

The 8 Diagnostic Categories

Category 1: Schema Compliance (quick, full)

What it checks: Every {vocabulary.note} in {vocabulary.notes}/ and self/memory/ (if self/ is enabled) has valid YAML frontmatter with required fields.

How to check:

# Find all note files (exclude topic maps)
for f in {vocabulary.notes}/*.md; do
  [[ -f "$f" ]] || continue
  # Check: YAML frontmatter exists
  head -1 "$f" | grep -q '^---$' || echo "FAIL: $f — no YAML frontmatter"
  # Check: description field present
  rg -q '^description:' "$f" || echo "WARN: $f — missing description field"
  # Check: topics field present and links to at least one topic map
  rg -q '^topics:' "$f" || echo "WARN: $f — missing topics field"
done

Additional checks:

• Domain-specific enum fields have valid values (check against template _schema blocks if templates exist)
• description field is non-empty (not just present)
• topics field contains at least one wiki link

If validate-kernel.sh exists in ${CLAUDE_PLUGIN_ROOT}/reference/, run it and include results.

Thresholds:

Condition	Level
Any note missing YAML frontmatter	FAIL
Any note missing description field	WARN
Any note missing topics field	WARN
Any invalid enum value	WARN
All notes pass all checks	PASS

Output format:

[1] Schema Compliance ............ WARN
    2 notes missing description:
      - notes/example-note.md
      - notes/another-note.md
    1 note missing topics:
      - notes/orphaned-claim.md
    12/15 notes fully compliant

Category 2: Orphan Detection (quick, full)

What it checks: Every {vocabulary.note} has at least one incoming wiki link from another file.

How to check:

# For each note file, check if ANY other file links to it
for f in {vocabulary.notes}/*.md; do
  [[ -f "$f" ]] || continue
  basename=$(basename "$f" .md)
  # Search for [[basename]] in all other files
  count=$(rg -l "\[\[$basename\]\]" --glob '*.md' | grep -v "$f" | wc -l | tr -d ' ')
  if [[ "$count" -eq 0 ]]; then
    echo "WARN: $f — no incoming links (orphan)"
  fi
done

Nuance: Orphans are not automatically failures. A note created today that hasn't been through /{vocabulary.cmd_reflect} yet is expected to be orphaned temporarily. Check file age:

Condition	Level
Orphan note created < 24 hours ago	INFO (expected — awaiting reflect phase)
Orphan note created 1-7 days ago	WARN
Orphan note older than 7 days	FAIL (persistent orphan needs attention)
No orphans detected	PASS

Output format:

[2] Orphan Detection ............. WARN
    3 orphan notes detected:
      - notes/new-claim.md (created 2h ago — awaiting reflect) [INFO]
      - notes/old-observation.md (created 5d ago) [WARN]
      - notes/forgotten-insight.md (created 14d ago) [FAIL]
    Recommendation: run /reflect on forgotten-insight.md and old-observation.md

Category 3: Link Health (quick, full)

What it checks: Every wiki link [[target]] in every file resolves to an existing file.

How to check:

# Extract all wiki links from all markdown files
# For each unique link target, verify a file with that name exists
rg -oN '\[\[([^\]]+)\]\]' --glob '*.md' -r '$1' | sort -u | while read target; do
  # Search for file matching this name
  found=$(find . -name "$target.md" -not -path "./.git/*" 2>/dev/null | head -1)
  if [[ -z "$found" ]]; then
    echo "FAIL: dangling link [[${target}]] — no file found"
    # Show which files contain this dangling link
    rg -l "\[\[$target\]\]" --glob '*.md'
  fi
done

Thresholds:

Condition	Level
Any dangling link (target does not exist)	FAIL
All links resolve	PASS

Why FAIL not WARN: Dangling links are broken promises. Every [[link]] a reader follows that leads nowhere erodes trust in the graph. Fix these immediately.

Output format:

[3] Link Health .................. FAIL
    2 dangling links found:
      - [[nonexistent-note]] referenced in:
        - notes/some-claim.md (line 14)
        - notes/another-claim.md (line 8)
      - [[removed-topic]] referenced in:
        - notes/old-note.md (line 22)
    Recommendation: create missing notes or remove broken links

Category 4: Description Quality (full only)

What it checks: Every {vocabulary.note}'s description adds genuine information beyond the title — not just a restatement.

How to check:

For each {vocabulary.note}:

1. Read the title (filename without extension)
2. Read the description field from YAML frontmatter
3. Evaluate: does the description add scope, mechanism, or implication that the title does not cover?

Quality heuristics:

Check	Threshold
Description length	50-200 chars ideal. < 30 chars = too terse. > 250 chars = too verbose
Restatement detection	If description uses >70% of the same words as the title = restatement
Information added	Description should mention mechanism, scope, or implication not in title

Thresholds:

Condition	Level
Description is a clear restatement of the title	WARN
Description is < 30 characters	WARN
Description is missing entirely	WARN (also caught by Category 1)
Description adds genuine new information	PASS

This check requires judgment. Use semantic understanding, not just string matching. A description that uses different words but says the same thing as the title is still a restatement.

Output format:

[4] Description Quality .......... WARN
    2 descriptions are restatements:
      - notes/quality-matters.md
        Title: "quality matters more than quantity"
        Description: "quality is more important than quantity in knowledge work"
        Issue: restates title without adding mechanism or implication
      - notes/structure-helps.md
        Title: "structure without processing provides no value"
        Description: "having structure without processing it is not valuable"
        Issue: exact restatement
    Recommendation: rewrite descriptions to add scope, mechanism, or implication

Category 5: Three-Space Boundary Check (full, three-space)

What it checks: Content respects the boundaries between self/, {vocabulary.notes}/, and ops/. Each space has a purpose — conflating them degrades search quality, navigation, and trust.

Before running this check: Read ${CLAUDE_PLUGIN_ROOT}/reference/three-spaces.md for the full boundary specification.

Six conflation patterns to detect:

5a. Ops into Notes (Infrastructure Creep)

Queue state, health metrics, task files, or processing artifacts appearing in {vocabulary.notes}/ directory.

Detection:

# Check for ops-pattern YAML fields in notes
rg '^(current_phase|completed_phases|batch|source_task|queue_id):' {vocabulary.notes}/ --glob '*.md'
# Check for task file patterns in notes
rg '## (Create|Reflect|Reweave|Verify|Enrich)$' {vocabulary.notes}/ --glob '*.md'

Found	Level
Any ops-pattern content in {vocabulary.notes}/	WARN

5b. Self into Notes (Identity Pollution)

Agent identity or methodology content mixed into user's knowledge graph. Agent's operational observations appearing in user's {vocabulary.notes}/ space.

Detection:

# Check for agent-reflection patterns in notes (methodology observations, workflow assessments)
rg -i '(my methodology|I observed that|agent reflection|session learning|I learned)' {vocabulary.notes}/ --glob '*.md'

Found	Level
Any agent-reflection content in {vocabulary.notes}/	WARN

5c. Notes into Ops (Trapped Knowledge)

Genuine insights trapped in session logs, observations, or ops files that should be promoted to {vocabulary.notes}/.

Detection:

# Check for claim-like content in ops that could be notes
# Look for files with description fields in ops/observations/ or ops/methodology/
rg '^description:' ops/observations/*.md ops/methodology/*.md 2>/dev/null

Found	Level
Content in ops/ with note-like schema (description + topics)	INFO (may be intentional, flag for review)

5d. Self into Ops / Ops into Self

Temporal state stored in self/ (should be in ops/), or identity content stored in ops/ (should be in self/).

Detection:

# Check for temporal/queue content in self/
rg '^(current_phase|status|queue):' self/ --glob '*.md' 2>/dev/null
# Check for identity content in ops/
rg -i '(my identity|I am|who I am|my personality)' ops/ --glob '*.md' 2>/dev/null

5e. Notes into Self

Domain knowledge stored in self/ instead of {vocabulary.notes}/.

Detection:

# Check self/memory/ for notes that have topics linking to notes-space topic maps
rg '^topics:.*\[\[' self/memory/*.md 2>/dev/null | grep -v 'identity\|methodology\|goals\|relationships'

5f. Self Space Absence Effects (when self/ is disabled)

When self/ is disabled, verify ops/ absorbs self-space content correctly. Check that goals and handoffs are in ops/, not floating in {vocabulary.notes}/.

Detection:

# If self/ doesn't exist, check that goals/handoffs are in ops/
if [[ ! -d "self/" ]]; then
  # Check for goals or handoff content in notes/
  rg -i '(my goals|current goals|handoff|session handoff)' {vocabulary.notes}/ --glob '*.md'
fi

Thresholds:

Condition	Level
Any boundary violation (5a, 5b, 5d, 5e)	WARN
Trapped knowledge in ops (5c)	INFO
Self-absence effects (5f)	WARN
All boundaries intact	PASS

Output format:

[5] Three-Space Boundaries ....... WARN
    1 boundary violation detected:
      Ops into Notes (infrastructure creep):
        - notes/task-tracking.md contains queue state fields (current_phase, batch)
        - Should be in ops/queue/ not notes/
    1 potential trapped knowledge:
      - ops/observations/interesting-pattern.md has note-like schema
        Consider promoting to notes/ via /reduce or direct creation
    Recommendation: move task-tracking.md to ops/queue/

Category 6: Processing Throughput (full only)

What it checks: The inbox-to-knowledge ratio. High ratios indicate collector's fallacy — capturing more than processing.

How to check:

# Count items in each space
INBOX_COUNT=$(find {vocabulary.inbox}/ -name '*.md' -not -path '*/archive/*' 2>/dev/null | wc -l | tr -d ' ')
NOTES_COUNT=$(find {vocabulary.notes}/ -name '*.md' 2>/dev/null | wc -l | tr -d ' ')
QUEUE_COUNT=$(find ops/queue/ -name '*.md' -not -path '*/archive/*' 2>/dev/null | wc -l | tr -d ' ')

# Calculate ratio
if [[ $((INBOX_COUNT + NOTES_COUNT)) -gt 0 ]]; then
  RATIO=$((INBOX_COUNT * 100 / (INBOX_COUNT + NOTES_COUNT)))
else
  RATIO=0
fi

echo "Inbox: $INBOX_COUNT | Notes: $NOTES_COUNT | In-progress: $QUEUE_COUNT | Ratio: ${RATIO}%"

Thresholds:

Condition	Level
Inbox-to-total ratio > 75% (3:1 inbox to notes)	FAIL
Inbox-to-total ratio > 50%	WARN
Inbox items > 20	WARN (regardless of ratio)
Inbox-to-total ratio <= 50% AND inbox <= 20	PASS

Output format:

[6] Processing Throughput ........ WARN
    inbox: 12 | notes: 8 | in-progress: 3 | ratio: 60%
    Inbox items outnumber processed notes — collector's fallacy risk
    Recommendation: run /reduce on oldest inbox items or /pipeline for end-to-end processing

Category 7: Stale Note Detection (full only)

What it checks: Notes not modified recently AND with low link density. Staleness is condition-based (low link density + no activity since N new notes were added), not purely calendar-based.

How to check:

For each {vocabulary.note}:

1. Check last modified date (stat or git log)
2. Count incoming links (same as orphan detection, but counting instead of boolean)
3. Flag notes with: modified > 30 days ago AND < 2 incoming links

for f in {vocabulary.notes}/*.md; do
  [[ -f "$f" ]] || continue
  basename=$(basename "$f" .md)

  # Last modified (days ago)
  mod_days=$(( ($(date +%s) - $(stat -f %m "$f" 2>/dev/null || stat -c %Y "$f" 2>/dev/null)) / 86400 ))

  # Incoming link count
  incoming=$(rg -l "\[\[$basename\]\]" --glob '*.md' | grep -v "$f" | wc -l | tr -d ' ')

  if [[ $mod_days -gt 30 ]] && [[ $incoming -lt 2 ]]; then
    echo "STALE: $f — $mod_days days old, $incoming incoming links"
  fi
done

Thresholds:

Condition	Level
Any note > 30 days old with < 2 incoming links	WARN
Any note > 90 days old with 0 incoming links	FAIL
All notes either recent or well-connected	PASS

Prioritization: Sort stale notes by:

1. Age (oldest first)
2. Link density (least connected first)
3. Topic relevance (notes in active topic areas are higher priority)

Output format:

[7] Stale Notes .................. WARN
    4 stale notes detected (>30d old, <2 incoming links):
      - notes/old-observation.md (92d, 0 links) [FAIL — consider archiving or reweaving]
      - notes/early-claim.md (45d, 1 link) [WARN]
      - notes/setupial-thought.md (38d, 1 link) [WARN]
      - notes/first-draft.md (31d, 0 links) [WARN]
    Recommendation: run /reweave on these notes to find connections, or archive if no longer relevant

Category 8: {vocabulary.topic_map} Coherence (full only)

What it checks: Each {vocabulary.topic_map} has a healthy number of linked notes, and notes in the same topic area are actually linked.

How to check:

For each {vocabulary.topic_map} file:

1. Count notes that link TO this {vocabulary.topic_map} (notes with this in their topics footer/field)
2. Check if there are notes in the same topic area NOT linked to the {vocabulary.topic_map} (coverage gaps)
3. Verify context phrases exist on Core Ideas links (not bare links)

# For each topic map
for moc in {vocabulary.notes}/*.md; do
  [[ -f "$moc" ]] || continue
  # Check if this is a topic map (has type: moc in frontmatter)
  rg -q '^type: moc' "$moc" || continue

  moc_name=$(basename "$moc" .md)

  # Count notes linking to this topic map
  note_count=$(rg -l "\[\[$moc_name\]\]" {vocabulary.notes}/ --glob '*.md' | grep -v "$moc" | wc -l | tr -d ' ')

  echo "$moc_name: $note_count notes"
done

Thresholds:

Condition	Level
{vocabulary.topic_map} with < 5 notes	WARN (underdeveloped — consider merging into parent)
{vocabulary.topic_map} with > 50 notes	WARN (oversized — consider splitting into sub-{vocabulary.topic_maps})
{vocabulary.topic_map} with > 40 notes	INFO (approaching threshold)
Notes exist in topic area but not linked to {vocabulary.topic_map}	WARN (coverage gap)
All {vocabulary.topic_maps} in 5-50 range with good coverage	PASS

Context phrase check:

# Check for bare links in topic map Core Ideas (links without context phrases)
for moc in {vocabulary.notes}/*.md; do
  rg -q '^type: moc' "$moc" || continue
  # Look for "- [[note]]" without " — " context
  rg '^\s*- \[\[' "$moc" | grep -v ' — ' | grep -v '^\s*- \[\[.*\]\].*—'
done

Bare links without context phrases are address book entries, not navigation. Every link in a {vocabulary.topic_map} should explain WHY to follow it.

Output format:

[8] MOC Coherence ................ WARN
    3 topic maps checked:
      - knowledge-work: 28 notes [PASS]
      - graph-structure: 52 notes [WARN — oversized, consider splitting]
        3 bare links without context phrases
      - agent-cognition: 3 notes [WARN — underdeveloped, consider merging]
    Recommendation: split graph-structure into sub-topic-maps; add context phrases to bare links

---

Condition-Based Maintenance Signals

After running all applicable diagnostic categories, check these condition-based triggers. These are NOT the 8 categories above — they are cross-cutting signals that suggest specific skill invocations.

Condition	Threshold	Recommendation
Pending observations	>= 10 files in ops/observations/	Consider running /rethink
Open tensions	>= 5 files in ops/tensions/	Consider running /rethink
Inbox items	>= 3 items	Consider /reduce or /pipeline
Unprocessed sessions	>= 5 files in ops/sessions/	Consider /remember --mine-sessions
Orphan notes	Any persistent (> 7d)	Run /reflect on orphaned notes
Dangling links	Any	Fix broken references immediately
Stale notes	Low links + old	Consider /reweave
{vocabulary.topic_map} oversized	> 40 notes	Consider splitting
Queue stalled	Tasks pending > 2 sessions without progress	Surface as blocked
Trigger coverage gap	Known maintenance condition has no configured trigger	Flag gap itself

How to check condition counts:

# Pending observations
OBS_COUNT=$(find ops/observations/ -name '*.md' 2>/dev/null | wc -l | tr -d ' ')

# Open tensions
TENSION_COUNT=$(find ops/tensions/ -name '*.md' 2>/dev/null | wc -l | tr -d ' ')

# Inbox items
INBOX_COUNT=$(find {vocabulary.inbox}/ -name '*.md' -not -path '*/archive/*' 2>/dev/null | wc -l | tr -d ' ')

# Unprocessed sessions
SESSION_COUNT=$(find ops/sessions/ -name '*.md' 2>/dev/null | wc -l | tr -d ' ')

# Queue stalled
PENDING_TASKS=$(jq '[.tasks[] | select(.status=="pending")] | length' ops/queue/queue.json 2>/dev/null || echo 0)

The meta-trigger: Include a "trigger coverage" check. Compare known maintenance conditions against what is actually being checked. If a maintenance condition has no corresponding check or trigger configured, that gap itself is a finding. This prevents the failure mode where maintenance debt accumulates undetected.

---

Output Format

The complete health report follows this structure. Every report, regardless of mode, uses this format.

=== HEALTH REPORT ===
Mode: [quick | full | three-space]
Date: YYYY-MM-DD
Notes scanned: N | Topic maps: N | Inbox items: N

Summary: N FAIL, N WARN, N PASS

FAIL:
- [Category N]: [brief description of failure]
  [specific files and details]

WARN:
- [Category N]: [brief description of warning]
  [specific files and details]

PASS:
- [Category N]: [confirmation]

---

[1] Schema Compliance ............ PASS | WARN | FAIL
    [details — specific files, specific issues]

[2] Orphan Detection ............. PASS | WARN | FAIL
    [details — specific files, age, link count]

[3] Link Health .................. PASS | WARN | FAIL
    [details — specific dangling links, where referenced]

[4] Description Quality .......... PASS | WARN | FAIL  (full mode only)
    [details — specific files, title vs description comparison]

[5] Three-Space Boundaries ....... PASS | WARN | FAIL  (full/three-space mode)
    [details — specific boundary violations by type]

[6] Processing Throughput ........ PASS | WARN | FAIL  (full mode only)
    inbox: N | notes: N | in-progress: N | ratio: N%

[7] Stale Notes .................. PASS | WARN | FAIL  (full mode only)
    [N notes older than 30d with <2 incoming links, sorted by priority]

[8] MOC Coherence ................ PASS | WARN | FAIL  (full mode only)
    [details — note count per topic map, coverage gaps, bare links]

---

Maintenance Signals:
    [condition-based triggers from table above, if any thresholds met]
    - observations: N pending (threshold: 10) [TRIGGERED | OK]
    - tensions: N pending (threshold: 5) [TRIGGERED | OK]
    - inbox: N items (threshold: 3) [TRIGGERED | OK]
    - sessions: N unprocessed (threshold: 5) [TRIGGERED | OK]

---

Recommended Actions (top 3, ranked by impact):
1. [Most impactful action — specific command + specific file]
2. [Second priority — specific command + specific file]
3. [Third priority — specific command + specific file]
=== END REPORT ===

Report Storage

Write every health report to ops/health/YYYY-MM-DD-report.md. If multiple reports are run on the same day, append a counter: YYYY-MM-DD-report-2.md.

This creates a health history that /architect can reference when proposing evolution. Trends across reports reveal systemic patterns that individual reports miss.

---

Quality Standards

Be Specific

Every finding MUST name the specific file(s) involved.

Bad: "some notes lack descriptions" Good: "notes/example-note.md and notes/another-note.md are missing the description field"

Bad: "there are dangling links" Good: "[[nonexistent-claim]] is referenced in notes/old-note.md (line 14) and notes/related.md (line 22) but no file with that name exists"

Prioritize by Impact

Not all issues are equal. The recommended actions section ranks by impact:

Impact Tier	Examples	Why
Highest	Dangling links, persistent orphans	Broken promises in the graph — readers hit dead ends
High	Schema violations, boundary violations	Structural integrity — compounds into larger problems
Medium	Description quality, stale notes	Retrieval quality — degraded but not broken
Low	{vocabulary.topic_map} size warnings, throughput ratio	Maintenance debt — matters at scale

Don't Overwhelm

• Focus on the top 5-10 issues in the recommended actions
• Group related issues (3 notes missing descriptions = 1 finding, not 3)
• For large vaults, cap per-category detail at 10 specific files, then summarize: "...and 15 more"

Distinguish FAIL from WARN

Level	Meaning	Action Required
FAIL	Structural issue — something is broken	Fix before it compounds
WARN	Improvement opportunity — something is suboptimal	Address when convenient
PASS	Category is healthy	No action needed
INFO	Noteworthy but not actionable	Context for understanding

FAIL is reserved for: Dangling links (broken graph), persistent orphans (> 7 days), severe schema violations (missing frontmatter entirely), critical boundary violations.

WARN is for everything else that is suboptimal but not broken.

---

Mode-Specific Behavior

Quick Mode (default)

Runs categories 1-3 only: Schema, Orphans, Links.

Use when:

• Session start health check
• Quick pulse on vault integrity
• After a batch of note creation

Runtime: Should complete in < 30 seconds for vaults up to 200 notes.

Full Mode

Runs all 8 categories plus maintenance signals.

Use when:

• Periodic comprehensive health review
• After significant vault changes (large batch processing, restructuring)
• When something "feels wrong" about the vault
• Before a /architect or /reseed session

Runtime: May take 1-3 minutes for larger vaults due to description quality analysis and {vocabulary.topic_map} coherence checks.

Three-Space Mode

Runs category 5 only: boundary violation checks.

Use when:

• After /reseed to verify boundaries are intact
• When search results seem contaminated (ops content in knowledge queries)
• When self/ is being enabled or disabled
• Debugging "why does my search return weird results?"

Runtime: Should complete in < 30 seconds.

---

Edge Cases

Empty Vault

A vault with 0 notes is not unhealthy — it is new. Report:

Notes scanned: 0 | Topic maps: 0 | Inbox items: N
All categories PASS (no notes to check)
Maintenance Signal: inbox has N items — consider /reduce to start building knowledge

Self Space Disabled

When self/ does not exist:

• Skip self-space boundary checks (5b, 5d, 5e inbound)
• But DO check that ops/ correctly absorbs self-space content (5f)
• Note in the report: "self/ is disabled — boundary checks adapted accordingly"

No Semantic Search

When qmd/MCP tools are unavailable:

• Skip any checks that depend on semantic search
• Note in the report: "Semantic search unavailable — some checks skipped"
• All file-based checks (schema, orphans, links, boundaries) still run

Large Vaults (500+ notes)

• Cap per-category file listings at 10, then summarize
• Consider running checks in batches if performance degrades
• Focus recommended actions on highest-impact items

---

Integration with Other Skills

Health report findings feed into other skills:

Finding	Feeds Into	How
Orphan notes	/reflect	Run reflect to find connections for orphaned notes
Stale notes	/reweave	Run reweave to update old notes with new connections
Description quality issues	/verify or manual rewrite	Fix descriptions to improve retrieval
Schema violations	/validate	Run validation to fix specific schema issues
Boundary violations	Manual restructuring	Move files to correct space
Processing throughput	/reduce or /pipeline	Process inbox items to improve ratio
{vocabulary.topic_map} oversized	Manual split or /architect	Split oversized {vocabulary.topic_maps} into sub-{vocabulary.topic_maps}
Accumulated observations	/rethink	Review and triage observations
Accumulated tensions	/rethink	Resolve or dissolve tensions

The health-to-action loop:

/health (diagnose) -> specific findings -> specific skill invocation -> /health (verify fix)

Health is diagnostic only — it measures state without prescribing changes. /architect reads health reports and proposes changes with research backing. The separation matters: health tells you WHAT is wrong, architect tells you WHY and HOW to fix it.