⚠️ MANDATORY ENTRY POINT — Execute Before ANY Other Action

Step 1: Read schema.sql

# ALWAYS read the schema first to understand tables, constraints, views
cat .claude/skills/audit-flow/schema.sql

Step 2: Check if DB exists — NEVER recreate

# Check for existing database
ls -la .audit/audit.db 2>/dev/null && echo "DB EXISTS - DO NOT RECREATE" || echo "No DB - safe to init"

Step 3: If DB exists, show current state

python .claude/skills/audit-flow/scripts/audit.py list

🚫 FORBIDDEN ACTIONS

Rule: If .audit/audit.db exists, ONLY use audit.py list, show, export, or INSERT operations. NEVER recreate.

Audit Flow

Interactive tracing of system flows with SQLite persistence. Supports multiple named flows per session, non-linear flows (branching/merging), and multi-format exports.

Organization Principles

Directory structure by purpose:

• Audits/Documentation/Compliance: docs/audits/{name}-{YYYY-MM-DD}/
• Ideation/Brainstorming: docs/ideation/{name}-{YYYY-MM-DD}.md (single file, no subdirectory unless artifacts needed)
• Debugging/Incident Review: docs/audits/{name}-{YYYY-MM-DD}/ (same as audits — captures evidence)
• Architecture Review: docs/audits/{name}-{YYYY-MM-DD}/ (same as audits — captures structural analysis)

Required files:

• INDEX.md (manifest, entry point)
• README.md (executive summary)
• {name}-audit.md (flow trace)

Lazy initialization: Create subdirectories only when artifacts exist

• screenshots/ network-traces/ diagrams/ code-samples/ test-results/ evidence/

Naming: {audit-name}-{type}.md

DB-First Discipline

Invariant: SQLite = sole source of truth. Context window: volatile, compacts without notice, hallucinates state.

🚨 CRITICAL: NEVER DESTROY EXISTING DATA

• If .audit/audit.db exists → it contains irreplaceable audit history
• NEVER run init when DB exists — use list to see what's there
• NEVER delete, drop, or recreate — only append

Constraints:

Checkpoint: Every 5 tuples → audit.py show <session> <flow>

Interactive Workflow - ALWAYS ASK USER

1. Session Start - Ask:

Name: ___
Purpose: security-audit | documentation | compliance | ideation | brainstorming | debugging | architecture-review | incident-review
Description: ___ (optional)

Initialize directory immediately. Lazily create subdirectories when artifacts are generated.

2. Granularity - Ask:

[fine]   Function-level trace (~50-200 tuples)
         Use: Security audits, debugging

[coarse] Boundary-level trace (~10-30 tuples)
         Use: Documentation, high-level flows

Choose: fine / coarse

3. During Trace:

Ask at decision points: trace deeper? mark concern? add finding (severity)? note?

4. On Export:

Ask format: json | yaml | md | mermaid | all

Post-export: Generate INDEX.md manifest. Organize artifacts by type. Prune empty directories.

Quick Reference

Layers: CODE | API | NETWORK | AUTH | DATA

Relations: TRIGGERS | READS | WRITES | VALIDATES | TRANSFORMS | BRANCHES | MERGES

Semantic Rules for Relations

CRITICAL: BRANCHES Must Have Conditions. Every BRANCHES edge requires a condition describing which path. Example: BRANCHES [token expired] vs BRANCHES [token valid].

Observations vs Flow Steps

Flow steps = things the SYSTEM DOES (function calls, data reads, network requests). Verified by tracing code.

Observations = things the ANALYST NOTES (missing features, potential risks). Record as findings, not tuples.

Wrong pattern:

T50 "NO cross-tab sync"       ← observation, not a system action
T51 "React state NOT shared"  ← observation
T50 --VALIDATES--> T51         ← chaining observations as flow

Correct pattern:

-- Record as finding instead:
INSERT INTO findings (flow_id, session_id, severity, category, description)
VALUES (?, ?, 'medium', 'state-management',
        'No cross-tab sync: React state not shared across tabs');

Rule: NEVER chain observations with VALIDATES. If describing what the system DOESN'T do, use a finding.

Data Model

Session (audit container)
  └── Flow (named DAG with entry point)
       └── Tuple (node: layer + action + subject)
            └── Edge (relation + optional condition)

Storage & CLI

# Core
python .claude/skills/audit-flow/scripts/audit.py init              # Initialize DB
python .claude/skills/audit-flow/scripts/audit.py list              # List sessions
python .claude/skills/audit-flow/scripts/audit.py show <session>    # Show flows
python .claude/skills/audit-flow/scripts/audit.py show <session> <flow>  # Show flow details
python .claude/skills/audit-flow/scripts/audit.py export <session>  # Export all
python .claude/skills/audit-flow/scripts/audit.py export <session> -f <flow>  # Export one flow
python .claude/skills/audit-flow/scripts/audit.py validate <session>         # Validate flows

# Git integration
python .claude/skills/audit-flow/scripts/audit.py git-setup         # Configure merge/diff drivers (once)
python .claude/skills/audit-flow/scripts/audit.py db-merge %O %A %B # Git merge driver (auto-called)

# CSV backup/portability (optional)
python .claude/skills/audit-flow/scripts/audit.py csv-export               # DB → .audit/csv/*.csv
python .claude/skills/audit-flow/scripts/audit.py csv-import               # .audit/csv/*.csv → DB
python .claude/skills/audit-flow/scripts/audit.py csv-merge <theirs_dir>   # Merge two CSV sets

Non-Linear Flows

Branching: One tuple → multiple outgoing edges with conditions

INSERT INTO edges (from_tuple, to_tuple, relation, condition)
VALUES (5, 6, 'BRANCHES', 'token valid'),
       (5, 7, 'BRANCHES', 'token expired');

Merging: Multiple tuples → one tuple

INSERT INTO edges (from_tuple, to_tuple, relation)
VALUES (6, 8, 'TRIGGERS'),
       (9, 8, 'MERGES');  -- refresh path merges back

Files

• [scripts/audit.py](scripts/audit.py) - CLI for all commands (init, list, show, export, validate, db-merge, git-setup, csv-\*)
• [COMMANDS.md](COMMANDS.md) - Detailed SQL reference
• [EXAMPLES.md](EXAMPLES.md) - Full examples with non-linear flows
• [schema.sql](schema.sql) - Database schema
• .gitattributes - Git merge/diff driver config for audit.db

Git Context

Capture on session start: commit hash, branch, working tree status. Include in all exports.

Mermaid Validation

Run python .claude/skills/audit-flow/scripts/audit.py validate <session> before export.

Post-export features (automatic):

• Step numbers: BFS topological order from entry point (1. action, 2. action)
• Entry point: Stadium shape with green styling
• Edge arrows: --> solid (TRIGGERS/VALIDATES/TRANSFORMS/BRANCHES/MERGES), -.-> dotted (READS), ==> thick (WRITES)
• Observations: Separated into dashed-border OBSERVATIONS subgraph
• Direction: --direction LR flag for horizontal layouts

Diagram Readability Requirements

All diagrams MUST be produced by audit.py export. Never hand-craft mermaid. The exporter enforces:

1. Entry point marker — Green stadium-shape node ([label]):::entryPoint
2. Step numbers — BFS topological order: 1. action, 2. action, 3. action
3. Legend block — classDef styles for entryPoint, concern, observation
4. Observation separation — Concern-only chains go to OBSERVATIONS subgraph, not main flow
5. Label safety — HTML entities for (), "", <>, |, [] characters (auto-sanitized)

Reading flow must be obvious. A reader opening the diagram cold must immediately see:

• WHERE to start (green entry node)
• WHAT ORDER to read (step numbers)
• WHICH PATH is happy vs error (branch conditions on edges)
• WHAT THE COLORS MEAN (legend)

Node label rules:

• Use concrete nouns/verbs: handleCallback(), exchangeCodeForTokens()
• NOT bare verbs: "Configure", "Select", "Enable"
• Disambiguate duplicates: auto-suffixed with subject when action repeats

Size limits:

• 40+ nodes → warning, consider splitting
• 60+ nodes → error, MUST split into sub-flows
• If flow has 5+ independent sub-flows → split by purpose

Git Workflow — Custom Merge Driver

Problem: SQLite is binary — git merge can't auto-resolve .audit/audit.db.

Solution: Custom git merge driver. audit.db stays in git (small, single file). On conflict, git calls audit.py db-merge to auto-merge using SQL.

One-Time Setup

python .claude/skills/audit-flow/scripts/audit.py git-setup

This configures (in .git/config):

• Merge driver: merge.sqlite-audit — calls audit.py db-merge %O %A %B on conflict
• Diff driver: diff.sqlite — sqlite3 .dump for readable git diff output

Also requires .gitattributes (already in repo):

.audit/audit.db diff=sqlite merge=sqlite-audit

How It Works

1. You commit audit.db normally — git add .audit/audit.db && git commit
2. git diff shows SQL text (via textconv)
3. On git merge with conflict → git calls the merge driver automatically
4. Driver opens both DBs, merges sessions by name (later updated_at wins), remaps IDs
5. Result written to ours — merge completes cleanly

Merge Strategy

All INTEGER PKs remapped sequentially. Foreign keys updated.

CSV Backup (Optional)

CSV export/import still available for portability and backup:

python .claude/skills/audit-flow/scripts/audit.py csv-export   # DB → .audit/csv/*.csv (QUOTE_ALL)
python .claude/skills/audit-flow/scripts/audit.py csv-import   # CSV → DB (recreates from scratch)

Output Quality

Principles:

• ASCII sequence diagrams for complex flows
• Side-by-side tables for alternatives
• Real code from traced files with file:line references
• What/Why/Example pattern
• No generic templates

Completion Checklist

• Directory exists with ISO date suffix
• INDEX.md manifest generated
• README.md (executive summary)
• Naming convention: {name}-{type}.md
• Artifacts in typed subdirectories (lazy init)
• Git context captured
• Git merge driver configured (audit.py git-setup)
• No orphaned files
• Diagrams pass audit.py validate

Anti-Patterns

Flat structure, empty directories, orphan nodes, unlabeled branches, generic identifiers, missing git context, no manifest, hand-crafted mermaid, bare-verb labels.