Architect

Runtime Configuration (Step 0 — before any processing)

Read these files to configure domain-specific behavior:

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

   • Use vocabulary.notes for the notes folder name
   • Use vocabulary.note / vocabulary.note_plural for note type references
   • Use vocabulary.topic_map / vocabulary.topic_map_plural for MOC references
   • Use vocabulary.inbox for the inbox folder name
   • Use vocabulary.cmd_reflect for connection-finding command name
   • Use vocabulary.cmd_reweave for backward-pass command name
   • Use vocabulary.cmd_verify for verification command name
   • Use vocabulary.architect for the command name in output

2. ops/config.yaml — processing depth, pipeline chaining, automation settings

3. ops/derivation.md — original derivation record (the design intent baseline)

If these files don't exist, use universal defaults and warn the user.

EXECUTE NOW

Target: $ARGUMENTS

Parse immediately:

• If target names a specific area (e.g., "schema", "processing", "MOC structure"): focus analysis on that area
• If target is empty: run full-system analysis across all dimensions
• If target is --dry-run: run analysis but do not offer implementation

Execute these phases sequentially:

1. Locate system files and detect platform
2. Read derivation record for design intent
3. Analyze health data (recent report or live check)
4. Scan for friction patterns across operational surfaces
5. Consult research to ground evidence in specific claims
6. Generate 3-5 ranked recommendations with full evidence chains
7. Present to user and implement on approval

START NOW. Reference below defines the seven-phase workflow.

Philosophy

Evidence beats intuition. Research beats habit.

Every rule in the context file was a hypothesis. Every skill workflow was a design choice. Hypotheses need testing against operational reality. This skill connects three evidence streams — health data, friction patterns, and research claims — to produce specific, actionable recommendations.

You are not guessing what might help. You are diagnosing what IS happening (health + friction) and prescribing what research says SHOULD happen. Every recommendation traces to specific evidence. "I think this would be better" is not a recommendation. "Health shows 12 orphans, friction log shows repeated orphan complaints, research claim [[orphan notes decay faster than connected ones]] supports adding a condition-triggered reweave pass" — THAT is a recommendation.

The 25% meta-work budget: In a 60-minute session, at most 15 minutes should be spent on system evolution. If a recommendation estimates >15 minutes to implement, the recommendation should be "defer to next session." The system serves the work, not the other way around.

INVARIANT: Architect NEVER auto-implements. Every recommendation requires explicit user approval before any files are modified. This prevents the cognitive outsourcing failure mode — the human must remain in the judgment loop for system evolution.

PHASE 1: Locate

Automated. No user interaction needed.

Detect the platform and find the system's key files:

Check filesystem:
  .claude/ directory exists         -> platform = "claude-code"
  Neither                           -> platform = "minimal"

Locate these files (paths vary by domain vocabulary):

If ops/derivation.md does not exist: Warn the user: "No derivation record found. Recommendations will be based on current state analysis only, without historical context of design decisions."

If ops/config.yaml does not exist: Warn: "No config file found. Using observed behavior to infer current configuration."

Record all file locations for use in subsequent phases.

PHASE 2: Read Derivation

Read ops/derivation.md to understand the system's original design intent. This is the baseline against which drift is measured.

Extract from derivation:

Also read ops/config.yaml — this is the live operational config that may have drifted from derivation. Compare dimension positions between derivation and config:

For each of the 8 dimensions:
  derivation_value = [from ops/derivation.md]
  config_value = [from ops/config.yaml]
  drifted = derivation_value != config_value

Record any drift for Phase 6. Drift is not inherently bad — it may represent healthy evolution. But UNRECOGNIZED drift creates incoherence.

If a specific focus area was requested ($ARGUMENTS), note which dimensions and failure modes are most relevant to that area and prioritize them in subsequent phases.

PHASE 3: Health Analysis

Check for a recent health report in ops/health/:

# Find health reports from the last 7 days
find ops/health/ -name "*.md" -mtime -7 2>/dev/null | sort -r | head -1

If a recent report exists: Read it fully. Extract every FAIL and WARN item as structured evidence:

For each FAIL/WARN:
  category: [schema | orphan | link | description | stale | moc | boundary | throughput]
  severity: [FAIL | WARN]
  detail: [specific finding — which notes, which fields, which links]
  count: [how many instances]

If no recent report exists: Run a live health assessment. Check each category:

Live check implementation:

# Count total notes
NOTE_COUNT=$(ls -1 {vocabulary.notes}/*.md 2>/dev/null | wc -l | tr -d ' ')

# Find orphans (notes with no incoming links)
for f in {vocabulary.notes}/*.md; do
  NAME=$(basename "$f" .md)
  LINKS=$(grep -rl "\[\[$NAME\]\]" {vocabulary.notes}/ 2>/dev/null | wc -l | tr -d ' ')
  [[ "$LINKS" -eq 0 ]] && echo "ORPHAN: $NAME"
done

# Find dangling links
grep -ohP '\[\[([^\]]+)\]\]' {vocabulary.notes}/*.md | sort -u | while read -r link; do
  NAME=$(echo "$link" | sed 's/\[\[//;s/\]\]//')
  [[ ! -f "{vocabulary.notes}/$NAME.md" ]] && echo "DANGLING: $NAME"
done

# Count inbox items
INBOX_COUNT=$(ls -1 {vocabulary.inbox}/ 2>/dev/null | wc -l | tr -d ' ')

# MOC sizes
grep -rl '^type: moc' {vocabulary.notes}/*.md 2>/dev/null | while read -r moc; do
  COUNT=$(grep -c '^\- \[\[' "$moc" 2>/dev/null)
  echo "MOC: $(basename "$moc" .md) = $COUNT notes"
done

# Missing descriptions
grep -rL '^description:' {vocabulary.notes}/*.md 2>/dev/null

Record all findings as evidence for Phase 6. Every finding must include the specific notes, links, or fields affected — not just counts.

PHASE 4: Read Friction

Scan for friction patterns across multiple operational surfaces. Friction is the gap between how the system should work and how it actually works.

4a. Observation Notes

# Find all pending observations
grep -rl '^status: pending' ops/observations/ 2>/dev/null

Read each pending observation. Categorize by type:

Count occurrences per category. 3+ items in the same category = a friction pattern worth investigating.

4b. Methodology Notes

# Read all methodology notes
ls ops/methodology/*.md 2>/dev/null

Read each methodology note. Look for:

• Recurring themes across multiple notes
• Tensions between stated methodology and actual practice
• Behavioral corrections that suggest the original design was wrong
• Methodology notes with multiple evidence entries (strong signal)

4c. Session Logs

# Read recent session logs (last 5-10 sessions)
ls -t ops/sessions/*.md 2>/dev/null | head -10

Scan session logs for:

• Repeated complaints or workarounds
• Steps that get skipped consistently
• Patterns in what the agent does vs what the system recommends
• Error patterns or tool failures

4d. Self Space

If self/ exists, read:

• self/methodology.md — how the agent describes its own process
• self/goals.md — whether current priorities align with system capabilities
• self/memory/ — any notes about workflow frustrations

4e. Build Friction Inventory

Compile all friction evidence into a structured inventory:

Friction Inventory:
  1. [category]: [description]
     Frequency: [N observations across M sessions]
     Sources: [filenames]
     Impact: [what breaks or degrades]
     System area: [which dimension/skill/workflow affected]

  2. [category]: [description]
     ...

Rank by frequency * impact. The most frequent, highest-impact friction patterns get priority in Phase 6.

PHASE 5: Consult Research

Ground the evidence in research. This is what separates /architect from ad-hoc troubleshooting — every recommendation connects to specific research claims, not general intuition.

Read the reference files:

Research Matching Process

For each friction pattern or health issue identified in Phases 3-4:

Step 1: Check dimension-claim-map.md

• Which research claims relate to this issue?
• What dimension positions do those claims support?
• Does the current configuration conflict with what research recommends?

Step 2: Check interaction-constraints.md

• Is this friction a CASCADE effect from a dimension mismatch?
• Would changing one dimension to fix this create pressure on another?
• Are there compensating mechanisms documented?

Step 3: Check failure-modes.md

• Does this match a known failure mode pattern?
• Was this failure mode flagged as HIGH risk during derivation?
• What does the failure mode document say about mitigation?

Step 4: Search the knowledge graph

• Use mcp__qmd__deep_search to find claims that address the specific friction
• Fall back to mcp__qmd__vector_search if hybrid search is unavailable
• If MCP is unavailable, fall back to qmd CLI (qmd query then qmd vsearch)
• Fall back to reading bundled reference files directly only if both MCP and qmd CLI are unavailable
• Search for the friction pattern described in natural language
• Search for the system area affected
• Search for proposed solution concepts

mcp__qmd__deep_search  query="[friction description in natural language]"  limit=10

Read the top results. For each relevant claim:

• Note the claim title (for citation in recommendations)
• Note what the claim argues
• Note how it applies to the current friction

The goal is to connect EVERY recommendation to specific research, not general intuition.