Paths: File paths (shared/, references/, ../ln-*) are relative to skills repo root. If not found at CWD, locate this SKILL.md directory and go up one level for repo root.

Persistence & Performance Auditor (L2 Coordinator)

Coordinates 4 specialized audit workers to perform database efficiency, transaction correctness, runtime performance, and resource lifecycle analysis.

Purpose & Scope

• Coordinates 4 audit workers (ln-651, ln-652, ln-653, ln-654) running in parallel
• Research current best practices for detected DB, ORM, async framework via MCP tools ONCE
• Pass shared context to all workers (token-efficient)
• Aggregate worker results into single consolidated report
• Write report to docs/project/persistence_audit.md (file-based, no task creation)
• Manual invocation by user; not part of Story pipeline
• Independent from ln-620 (can be run separately or after ln-620)

Workflow

1. Discovery: Load tech_stack.md, package manifests, detect DB/ORM/async framework, auto-discover Team ID
2. Research: Query MCP tools for DB/ORM/async best practices ONCE
3. Build Context: Create contextStore with best practices + DB-specific metadata
4. Prepare Output: Create output directory
5. Delegate: 4 workers in PARALLEL
6. Aggregate: Collect worker results, calculate scores
7. Write Report: Save to docs/project/persistence_audit.md

Phase 1: Discovery

Load project metadata:

• docs/project/tech_stack.md - detect DB, ORM, async framework
• Package manifests: requirements.txt, pyproject.toml, package.json, go.mod
• Auto-discover Team ID from docs/tasks/kanban_board.md

Extract DB-specific metadata:

Metadata	Source	Example
Database type	tech_stack.md, docker-compose.yml	PostgreSQL 16
ORM	imports, requirements.txt	SQLAlchemy 2.0
Async framework	imports, requirements.txt	asyncio, FastAPI
Session config	grep create_async_engine, sessionmaker	expire_on_commit=False
Triggers/NOTIFY	migration files	pg_notify('job_events', ...)
Connection pooling	engine config	pool_size=10, max_overflow=20

Scan for triggers:

Grep("pg_notify|NOTIFY|CREATE TRIGGER", path="alembic/versions/")
  OR path="migrations/"
→ Store: db_config.triggers = [{table, event, function}]

Phase 2: Research Best Practices (ONCE)

For each detected technology:

Technology	Research Focus
SQLAlchemy	Session lifecycle, expire_on_commit, bulk operations, eager/lazy loading
PostgreSQL	NOTIFY/LISTEN semantics, transaction isolation, batch operations
asyncio	to_thread, blocking detection, event loop best practices
FastAPI	Dependency injection scopes, background tasks, async endpoints

Build contextStore:

{
  "tech_stack": {"db": "postgresql", "orm": "sqlalchemy", "async": "asyncio"},
  "best_practices": {"sqlalchemy": {...}, "postgresql": {...}, "asyncio": {...}},
  "db_config": {
    "expire_on_commit": false,
    "triggers": [{"table": "jobs", "event": "UPDATE", "function": "notify_job_events"}],
    "pool_size": 10
  },
  "codebase_root": "/project",
  "output_dir": "docs/project/.audit/ln-650/{YYYY-MM-DD}"
}

Phase 3: Prepare Output Directory

mkdir -p {output_dir}   # No deletion — date folders preserve history

Phase 4: Delegate to Workers

CRITICAL: All delegations use Task tool with subagent_type: "general-purpose" for context isolation.

Prompt template:

Task(description: "Audit via ln-65X",
     prompt: "Execute ln-65X-{worker}-auditor. Read skill from ln-65X-{worker}-auditor/SKILL.md. Context: {contextStore}",
     subagent_type: "general-purpose")

Anti-Patterns:

• ❌ Direct Skill tool invocation without Task wrapper
• ❌ Any execution bypassing subagent context isolation

Workers (ALL 4 in PARALLEL):

#	Worker	Priority	What It Audits
1	ln-651-query-efficiency-auditor	HIGH	Redundant queries, N-UPDATE loops, over-fetching, caching scope
2	ln-652-transaction-correctness-auditor	HIGH	Commit patterns, trigger interaction, transaction scope, rollback
3	ln-653-runtime-performance-auditor	MEDIUM	Blocking IO in async, allocations, sync sleep, string concat
4	ln-654-resource-lifecycle-auditor	HIGH	Session scope mismatch, streaming resource holding, pool config, cleanup

Invocation (4 workers in PARALLEL):

FOR EACH worker IN [ln-651, ln-652, ln-653, ln-654]:
  Task(description: "Audit via " + worker,
       prompt: "Execute " + worker + ". Read skill. Context: " + JSON.stringify(contextStore),
       subagent_type: "general-purpose")

Worker Output Contract (File-Based):

Workers write full report to {output_dir}/{worker_id}.md per shared/templates/audit_worker_report_template.md.

Workers return minimal summary in-context (~50 tokens):

Report written: docs/project/.audit/ln-650/{YYYY-MM-DD}/651-query-efficiency.md
Score: 6.0/10 | Issues: 8 (C:0 H:3 M:4 L:1)

Phase 5: Aggregate Results (File-Based)

Workers wrote reports to {output_dir}/ and returned minimal summaries. Aggregation uses return values for numbers and file reads for findings tables.

Aggregation steps:

1. Parse scores/counts from worker return strings (already in context)
2. Read worker report files from {output_dir}/ for findings tables
3. Calculate overall score: average of 4 category scores
4. Sum severity counts across all workers
5. Sort findings by severity (CRITICAL → HIGH → MEDIUM → LOW)
6. Context Validation (Post-Filter)

Context Validation:

MANDATORY READ: Load shared/references/context_validation.md

Apply Rules 1, 6 to merged findings:

FOR EACH finding WHERE severity IN (HIGH, MEDIUM):
  # Rule 1: ADR/Planned Override
  IF finding matches ADR → advisory "[Planned: ADR-XXX]"

  # Rule 6: Execution Context
  IF finding.check IN (blocking_io, redundant_fetch, transaction_wide, cpu_bound):
    context = 0
    - Function in __init__/setup/bootstrap/migrate → context += 1
    - File in tasks/jobs/cron/                      → context += 1
    - Has timeout/safeguard nearby                  → context += 1
    - Small data (<100KB file, <100 items dataset)  → context += 1
    IF context >= 3 → advisory
    IF context >= 1 → severity -= 1

Downgraded findings → "Advisory Findings" section in report.
Recalculate overall score excluding advisory findings from penalty.

Exempt: Missing rollback CRITICAL, N-UPDATE loops in hot paths.

Output Format

## Persistence & Performance Audit Report - [DATE]

### Executive Summary
[2-3 sentences on overall persistence/performance health]

### Compliance Score

| Category | Score | Notes |
|----------|-------|-------|
| Query Efficiency | X/10 | ... |
| Transaction Correctness | X/10 | ... |
| Runtime Performance | X/10 | ... |
| Resource Lifecycle | X/10 | ... |
| **Overall** | **X/10** | |

### Severity Summary

| Severity | Count |
|----------|-------|
| Critical | X |
| High | X |
| Medium | X |
| Low | X |

### Findings by Category

#### 1. Query Efficiency

| Severity | Location | Issue | Recommendation | Effort |
|----------|----------|-------|----------------|--------|
| HIGH | job_processor.py:434 | Redundant entity fetch | Pass object not ID | S |

#### 2. Transaction Correctness

| Severity | Location | Issue | Recommendation | Effort |
|----------|----------|-------|----------------|--------|
| CRITICAL | job_processor.py:412 | Missing intermediate commits | Add commit at milestones | S |

#### 3. Runtime Performance

| Severity | Location | Issue | Recommendation | Effort |
|----------|----------|-------|----------------|--------|
| HIGH | job_processor.py:444 | Blocking read_bytes() in async | Use aiofiles/to_thread | S |

#### 4. Resource Lifecycle

| Severity | Location | Issue | Recommendation | Effort |
|----------|----------|-------|----------------|--------|
| CRITICAL | sse_stream.py:112 | DbSession held for entire SSE stream | Scope session to auth check only | M |

### Recommended Actions (Priority-Sorted)

| Priority | Category | Location | Issue | Recommendation | Effort |
|----------|----------|----------|-------|----------------|--------|
| CRITICAL | Transaction | ... | Missing commits | Add strategic commits | S |
| HIGH | Query | ... | Redundant fetch | Pass object not ID | S |

### Sources Consulted
- SQLAlchemy best practices: [URL]
- PostgreSQL NOTIFY docs: [URL]
- Python asyncio-dev: [URL]

Phase 6: Write Report

Write consolidated report to docs/project/persistence_audit.md with the Output Format above.

Critical Rules

• Single context gathering: Research best practices ONCE, pass contextStore to all workers
• Parallel execution: All 4 workers run in PARALLEL
• Trigger discovery: Scan migrations for triggers/NOTIFY before delegating (pass to ln-652)
• Metadata-only loading: Coordinator loads metadata; workers load full file contents
• Do not audit: Coordinator orchestrates only; audit logic lives in workers

Definition of Done

• Tech stack discovered (DB type, ORM, async framework)
• DB-specific metadata extracted (triggers, session config, pool settings)
• Best practices researched via MCP tools
• contextStore built with output_dir = docs/project/.audit/ln-650/{YYYY-MM-DD}
• Output directory created (no deletion of previous runs)
• All 4 workers invoked in PARALLEL and completed; each wrote report to {output_dir}/
• Results aggregated from return values (scores) + file reads (findings tables)
• Compliance score calculated per category + overall
• Executive Summary included
• Report written to docs/project/persistence_audit.md
• Sources consulted listed with URLs

Workers

• ln-651-query-efficiency-auditor
• ln-652-transaction-correctness-auditor
• ln-653-runtime-performance-auditor
• ln-654-resource-lifecycle-auditor

Reference Files

• Tech stack: docs/project/tech_stack.md
• Kanban board: docs/tasks/kanban_board.md

---

Version: 1.0.0 Last Updated: 2026-02-04