Nav Graph

Navigator Knowledge Graph Skill

Query and manage the unified project knowledge graph. Surfaces relevant knowledge from tasks, SOPs, system docs, and experiential memories.

Why This Exists

Navigator v6.0.0 introduces the Project Knowledge Graph:

• Unified search: Query across all knowledge types with one interface
• Experiential memory: Patterns, pitfalls, decisions, learnings persist
• Context-aware retrieval: Load only relevant knowledge (~1-2k tokens)
• Relationship traversal: Find related concepts and documents

When to Invoke

Query triggers:

• "What do we know about X?"
• "Show everything related to X"
• "Any pitfalls for X?"
• "What decisions about X?"
• "Find all knowledge about X"

Memory capture triggers:

• "Remember this pattern: ..."
• "Remember this pitfall: ..."
• "Remember we decided: ..."
• "Remember this learning: ..."

Graph management triggers:

• "Initialize knowledge graph"
• "Rebuild knowledge graph"
• "Show graph stats"

Graph Location

.agent/knowledge/graph.json (~1-2k tokens, loaded on query)

Execution Steps

Step 1: Determine Action

QUERY (searching knowledge):

User: "What do we know about authentication?"
→ Query graph by concept

CAPTURE (storing memory):

User: "Remember: auth changes often break session tests"
→ Create new memory node

INIT (building graph):

User: "Initialize knowledge graph"
→ Build graph from existing docs

STATS (viewing graph):

User: "Show graph stats"
→ Display graph statistics

Step 2: Load or Initialize Graph

Check if graph exists:

if [ -f ".agent/knowledge/graph.json" ]; then
  echo "Graph exists"
else
  echo "No graph found, will initialize"
fi

Initialize if not exists:

PLUGIN_DIR="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/navigator-marketplace/navigator}"
[ -d "$PLUGIN_DIR" ] || PLUGIN_DIR="$HOME/.claude/plugins/marketplaces/navigator-marketplace"
python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_builder.py" \
  --agent-dir .agent \
  --output .agent/knowledge/graph.json

Step 3A: Query Knowledge (If QUERY Action)

Extract concept from user input:

User: "What do we know about testing?"
→ Concept: testing

User: "Any pitfalls for auth?"
→ Concept: auth (normalized to authentication)

Run query:

PLUGIN_DIR="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/navigator-marketplace/navigator}"
[ -d "$PLUGIN_DIR" ] || PLUGIN_DIR="$HOME/.claude/plugins/marketplaces/navigator-marketplace"
python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_manager.py" \
  --action query \
  --concept "testing" \
  --graph-path .agent/knowledge/graph.json

Display results:

Knowledge Graph: "testing"

TASKS (3)
  - TASK-30: Task Verification Enhancement (completed)
  - TASK-17: Visual Regression Integration (completed)
  - TASK-11: Project Skills Generation (completed)

MEMORIES (2)
  - PITFALL: "Auth changes break session tests" (90%)
  - PATTERN: "Always run unit tests before integration" (85%)

SOPs (1)
  - visual-regression-setup

FILES (5)
  - skills/backend-test/*
  - skills/frontend-test/*

Load details: "Read TASK-30" or "Show testing memories"

Step 3B: Capture Memory (If CAPTURE Action)

Parse memory from user input:

User: "Remember this pitfall: auth changes often break session tests"
→ Type: pitfall
→ Summary: "auth changes often break session tests"
→ Concepts: [auth, testing]

User: "Remember we decided to use JWT over sessions for scaling"
→ Type: decision
→ Summary: "use JWT over sessions for scaling"
→ Concepts: [auth, architecture]

Determine memory type:

Create memory:

PLUGIN_DIR="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/navigator-marketplace/navigator}"
[ -d "$PLUGIN_DIR" ] || PLUGIN_DIR="$HOME/.claude/plugins/marketplaces/navigator-marketplace"
python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_manager.py" \
  --action add-memory \
  --memory-type pitfall \
  --summary "auth changes often break session tests" \
  --concepts "auth,testing" \
  --confidence 0.9 \
  --graph-path .agent/knowledge/graph.json

Write guarantees (v6.17.0+): the backing .md is written BEFORE the graph node and failures fail loudly (no more path-points-at-nothing nodes); a failed graph save rolls the file back. Concepts are validated against the graph's concept vocabulary — an unknown concept rejects the write and lists the valid vocabulary. Pass --allow-new-concept to register genuinely new concepts instead. Graphs without a curated vocabulary skip validation.

Decisions may carry a TRIZ contradiction (TASK-72). When a decision resolved a real tension (improving A worsened B), record it so future sessions can ask "how did we resolve this kind of problem before":

python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_manager.py" \
  --action add-memory --memory-type decision \
  --summary "ship new blocking features OFF by default" \
  --concepts "release,configuration" --confidence 0.95 \
  --contradiction "feature value vs regression risk" \
  --separation condition \
  --principle "dynamization (config toggle)"

All three flags are optional and free text; keep --contradiction under ~60 chars ("A vs B") because it is appended to recall lines. Separation modes (which TRIZ move resolved it):

Query by contradiction (every filter word must match, case-insensitive, across contradiction + summary + principle; always exits 0):

python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_manager.py" \
  --action contradictions --filter "rollback"
# Contradictions "rollback" (2)
#   - DECISION: "..." (95%) ↔ clean codebase vs rollback safety [separation: time]
#       principle: prior counter-action

Tagged memories also render with a ↔ A vs B suffix in session-start and nav-brief recall. Hand-tagging an existing decision = append the three footer lines after **Concepts**: in its .md, then graph_maintenance.py --action reconcile --execute --fields-only (see Reconcile below; --fields-only keeps a pruned resolved/ archive from being re-registered under fresh ids).

Optionally create detailed memory file:

# Pitfall: Auth Changes Break Session Tests

## Summary
Auth changes often break session tests due to...

## Context
Discovered during TASK-XX when...

## Recommended Approach
When modifying auth, always run...

## Related
- TASK-12: V3 Skills-Only
- SOP: autonomous-completion

Confirm capture:

Memory captured: mem-001

Type: Pitfall
Summary: "auth changes often break session tests"
Concepts: auth, testing
Confidence: 90%

This will be surfaced when working on auth or testing topics.

Step 3C: Initialize Graph (If INIT Action)

Build from existing docs:

PLUGIN_DIR="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/navigator-marketplace/navigator}"
[ -d "$PLUGIN_DIR" ] || PLUGIN_DIR="$HOME/.claude/plugins/marketplaces/navigator-marketplace"
python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_builder.py" \
  --agent-dir .agent \
  --output .agent/knowledge/graph.json

Display results:

Knowledge Graph Initialized

Scanned:
  - Tasks: 35
  - SOPs: 12
  - System docs: 3
  - Markers: 8

Extracted:
  - Concepts: 15
  - Relationships: 47

Graph saved to .agent/knowledge/graph.json

Query with: "What do we know about [topic]?"

Rebuild safety (v6.17.0+): re-running the builder over an existing graph PRESERVES the memories and files buckets and their edges — memories carry graph-only fields no scan can reconstruct, and rebuilds used to wipe them silently. Pass --no-preserve-memories for an intentional from-scratch rebuild.

Step 3D: Show Stats (If STATS Action)

Display graph statistics:

PLUGIN_DIR="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/navigator-marketplace/navigator}"
[ -d "$PLUGIN_DIR" ] || PLUGIN_DIR="$HOME/.claude/plugins/marketplaces/navigator-marketplace"
python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_manager.py" \
  --action stats \
  --graph-path .agent/knowledge/graph.json

Output:

Knowledge Graph Statistics
==========================
Total Nodes: 65
Total Edges: 47
Memories: 5
Last Updated: 2025-01-23T10:30:00Z

By Type:
  Tasks: 35
  SOPs: 12
  System: 3
  Markers: 8
  Concepts: 15
  Memories: 5

Step 4: Find Related (Optional)

If user asks for related items:

User: "What's related to TASK-29?"

Run traversal:

PLUGIN_DIR="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/navigator-marketplace/navigator}"
[ -d "$PLUGIN_DIR" ] || PLUGIN_DIR="$HOME/.claude/plugins/marketplaces/navigator-marketplace"
python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_manager.py" \
  --action related \
  --node-id "TASK-29" \
  --max-depth 2 \
  --graph-path .agent/knowledge/graph.json

Memory Types

Pattern

"We use X for Y in this project"

• Reusable approaches
• Project conventions
• Best practices

Pitfall

"Watch out for X when touching Y"

• Common mistakes
• Gotchas
• Failure modes

Decision

"We chose X over Y because Z"

• Architecture decisions
• Technology choices
• Trade-off rationale

Learning

"X usually means Y in this codebase"

• Project-specific knowledge
• Error interpretations
• Domain insights

Confidence System

Base confidence:

• Correction-based: 0.8
• Explicit capture: 0.9

Decay:

• 1% per week since last validation

Boost:

• +5% per use (max +25%)

Threshold:

• Below 0.3: Candidate for pruning
• Above 0.7: Reliable memory

Integration with Other Skills

nav-start (Session Start)

Loads graph stats on session start:

Knowledge graph: 65 nodes, 5 memories
Relevant: 2 memories for current context

nav-task (Task Creation)

Auto-extracts concepts from new tasks:

Creating TASK-35: Project Memory
Extracted concepts: knowledge, memory, graph
Added to graph.

nav-profile (Corrections)

Corrections auto-create memories via correction_to_memory.py:

PLUGIN_DIR="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/navigator-marketplace/navigator}"
[ -d "$PLUGIN_DIR" ] || PLUGIN_DIR="$HOME/.claude/plugins/marketplaces/navigator-marketplace"
# When correction detected in nav-profile:
python3 "$PLUGIN_DIR/skills/nav-graph/functions/correction_to_memory.py" \
  --action convert-one \
  --correction-json '{"pattern": "...", "context": "...", "confidence": "high"}'

# Output:
[Correction detected]
→ Type: pitfall (based on pattern analysis)
→ Concepts: [auth, testing] (auto-extracted)
→ Created memory: mem-002
→ Added to graph

Sync all corrections:

PLUGIN_DIR="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/navigator-marketplace/navigator}"
[ -d "$PLUGIN_DIR" ] || PLUGIN_DIR="$HOME/.claude/plugins/marketplaces/navigator-marketplace"
python3 "$PLUGIN_DIR/skills/nav-graph/functions/correction_to_memory.py" \
  --action sync \
  --profile-path .agent/.user-profile.json \
  --graph-path .agent/knowledge/graph.json

nav-marker (Context Markers)

Markers reference graph state:

## Graph State
- Memories surfaced: mem-001, mem-003
- Concepts active: auth, testing

navigator-research (Codebase Exploration Agent)

The navigator-research agent emits a structured research_findings JSON block alongside its markdown summary. After the agent returns, ingest those findings as graph memories via research_to_graph.py:

PLUGIN_DIR="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/navigator-marketplace/navigator}"
[ -d "$PLUGIN_DIR" ] || PLUGIN_DIR="$HOME/.claude/plugins/marketplaces/navigator-marketplace"
# Save the JSON block from the agent output to a file (or pipe via stdin)
python3 "$PLUGIN_DIR/skills/nav-graph/functions/research_to_graph.py" findings.json

# Or from stdin
cat findings.json | python3 "$PLUGIN_DIR/skills/nav-graph/functions/research_to_graph.py" -

# Validate without writing
python3 "$PLUGIN_DIR/skills/nav-graph/functions/research_to_graph.py" findings.json --dry-run

Trigger phrases:

• "Ingest research findings"
• "Save these findings to the graph"
• (Automatic, when a navigator-research invocation completes — orchestrator may auto-ingest)

Defaults:

• Confidence: 0.7 (lower than corrections/explicit captures — research is inference)
• Memory types accepted: pattern, pitfall, decision, learning
• Invalid entries are skipped with a printed error (exit code 1 if any errors)
• Evidence path (e.g. src/auth.ts:42) is embedded into the memory summary

Schema: see the Output Format section of agents/navigator-research.md for the full JSON shape the agent emits.

nav-deep-research (Web Research Skill)

Web research runs (skills/nav-deep-research) ingest through the same path: after the ship gate passes, report_to_graph.py --run <slug> turns every typed ## Key findings bullet of the report into a memory whose evidence is the cited source URL(s), then calls research_to_graph.ingest_findings. Same 0.7 confidence and memory types; no schema change.

Configuration

In .agent/.nav-config.json:

{
  "knowledge_graph": {
    "enabled": true,
    "auto_capture_corrections": true,
    "auto_capture_decisions": true,
    "auto_surface_relevant": true,
    "max_session_memories": 5,
    "confidence_decay_rate": 0.01,
    "staleness_threshold_days": 90,
    "git_tracked": true
  }
}

Note: confidence_decay_rate and staleness_threshold_days are consumed only by the manual graph_maintenance commands (--action decay / --action stale). They are not applied automatically on session start — decaying a git-tracked file every session would create constant churn. Run decay/staleness manually when curating the graph.

Graph Maintenance

Health Check

PLUGIN_DIR="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/navigator-marketplace/navigator}"
[ -d "$PLUGIN_DIR" ] || PLUGIN_DIR="$HOME/.claude/plugins/marketplaces/navigator-marketplace"
python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_maintenance.py" --action health

Output:

Knowledge Graph Health Check
========================================
Total Nodes: 133
Total Edges: 706
Memories: 37 (37 high confidence)
Tasks: 38
Concepts: 16
Orphan Nodes: 0
Duplicate Edges: 0
Dangling Edges: 0
Confidence Out-of-Range: 0

Health Score: 100/100

No integrity issues detected!

Advisory (not scored):
  - 8 potential memory conflicts (heuristic, advisory)
  - 3 stale memories (not validated in 90+ days)

Duplicate Edges, Dangling Edges, and Confidence Out-of-Range are the integrity gate — all three should read 0 on a healthy graph. If they don't, run --action repair (below).

v6.17.0 adds disk-vs-graph checks: Broken File Links (node references a file that doesn't exist) and Unindexed Memory Files (files on disk with no node — the drift class a 2026-07 audit found at 52/84 in a consumer repo) are score-affecting; concept-vocabulary drift and archived resolved/ files without nodes are advisory. Pass --root <project-root> when running from another directory.

Reconcile Disk vs Graph (v6.17.0+)

Report drift between memory files on disk and graph nodes; --execute registers unindexed files (type from parent dir, resolved/ parent → resolved: true, frontmatter/heading parsing with conservative fallbacks — 0.5 confidence when unknown) and copies TRIZ footer fields (**Contradiction** / **Separation** / **Principle**, TASK-72) from disk onto already-indexed nodes (reported as field_updates; never clears a field removed on disk, never re-syncs summary). Broken-link nodes are never auto-deleted and concept refs are never rewritten — those two are report-and-hint only:

PLUGIN_DIR="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/navigator-marketplace/navigator}"
[ -d "$PLUGIN_DIR" ] || PLUGIN_DIR="$HOME/.claude/plugins/marketplaces/navigator-marketplace"
# Dry-run report
python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_maintenance.py" --action reconcile
# Register unindexed files + apply TRIZ field updates
python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_maintenance.py" --action reconcile --execute
# TRIZ field updates only (safe when resolved/ holds deliberately pruned files)
python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_maintenance.py" --action reconcile --execute --fields-only

Resolve / Supersede a Memory (v6.17.0+)

When a memory stops being true (bug fixed, decision reversed, guidance codified elsewhere), do NOT delete it — resolve it. The node gets resolved: true (+ superseded_by and a supersedes edge when a newer memory replaces it) and the backing file moves to the sibling resolved/ directory. Resolved memories are excluded from session-start surfacing and task-doc recall, skipped by stale/decay sweeps, and flagged [resolved] in query output:

python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_manager.py" \
  --action resolve-memory --node-id mem-012 [--superseded-by mem-045]

Memory Recall (v6.17.0+)

Deterministic relevance ranking used by the SessionStart hook and nav-task Step 2.5 — also useful standalone:

# Explicit concepts (markdown for task docs, compact for terse output)
python3 "$PLUGIN_DIR/skills/nav-graph/functions/memory_recall.py" \
  --concepts "auth,testing" --format markdown --limit 5
# Auto mode: concepts from open task nodes + active context marker
python3 "$PLUGIN_DIR/skills/nav-graph/functions/memory_recall.py" \
  --auto --agent-dir .agent --limit 5

Scoring: concept overlap (alias-resolved), then confidence; resolved memories excluded; silent (exit 0, no output) when nothing matches. Compatible with consumer graphs that use file: keys and lack a concept_index.

Repair Integrity Defects

Idempotently dedupe (from, to, type) edge rows, drop edges that reference a missing node id, and normalize out-of-range memory confidences (a value like 90.0 is treated as 90% → 0.9). Safe to re-run:

PLUGIN_DIR="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/navigator-marketplace/navigator}"
[ -d "$PLUGIN_DIR" ] || PLUGIN_DIR="$HOME/.claude/plugins/marketplaces/navigator-marketplace"
python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_maintenance.py" --action repair

Conflict Detection

Find memories that may contradict each other. Advisory only — a high-false-positive keyword heuristic that does not affect the health score:

PLUGIN_DIR="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/navigator-marketplace/navigator}"
[ -d "$PLUGIN_DIR" ] || PLUGIN_DIR="$HOME/.claude/plugins/marketplaces/navigator-marketplace"
python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_maintenance.py" --action conflicts

Stale Memory Detection

Find memories not validated in 90+ days:

PLUGIN_DIR="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/navigator-marketplace/navigator}"
[ -d "$PLUGIN_DIR" ] || PLUGIN_DIR="$HOME/.claude/plugins/marketplaces/navigator-marketplace"
python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_maintenance.py" --action stale --stale-days 90

Low Confidence Pruning

Find and optionally remove low-confidence memories:

PLUGIN_DIR="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/navigator-marketplace/navigator}"
[ -d "$PLUGIN_DIR" ] || PLUGIN_DIR="$HOME/.claude/plugins/marketplaces/navigator-marketplace"
# Preview what would be removed
python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_maintenance.py" --action prune --threshold 0.3 --dry-run

# Actually remove (use with caution)
python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_maintenance.py" --action prune --threshold 0.3 --execute

Apply Decay (experimental, manual-only)

Reduce confidence based on time since each memory's last decay. Idempotent — running it twice on the same day is a no-op (each memory tracks last_decayed). The rate defaults to knowledge_graph.confidence_decay_rate when --decay-rate is omitted. This is not wired to any hook; run it manually when curating:

PLUGIN_DIR="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/cache/navigator-marketplace/navigator}"
[ -d "$PLUGIN_DIR" ] || PLUGIN_DIR="$HOME/.claude/plugins/marketplaces/navigator-marketplace"
python3 "$PLUGIN_DIR/skills/nav-graph/functions/graph_maintenance.py" --action decay

Token Budget

Session overhead: ~1.3k tokens

Success Criteria

Graph skill succeeds when:

• Query returns relevant results across knowledge types
• Memories persist and are surfaced appropriately
• Concepts connect related items
• Confidence decay/boost works
• Graph stays under 2k tokens overhead

Best Practices

Good queries:

• "What do we know about auth?" (specific concept)
• "Any pitfalls for testing?" (scoped type)
• "Show everything related to TASK-29" (node traversal)

Good memory capture:

• "Remember: we use X for Y" (clear pattern)
• "Remember this pitfall: X breaks Y" (specific issue)
• "Remember we decided X because Y" (rationale included)

Avoid:

• Overly broad queries ("What do we know?")
• Storing code snippets in memories (use paths instead)
• Capturing obvious knowledge (focus on project-specific insights)

This skill transforms Navigator from stateless assistant to knowledge-aware team member