multi-agent-e2e-validation（多 Agent 端到端验证）

Multi-Agent E2E Validation

Self-Evolving Skill: This skill improves through use. If instructions are wrong, parameters drifted, or a workaround was needed — fix this file immediately, don't defer. Only update for real, reproducible issues.

Overview

Prescriptive workflow for spawning parallel validation agents to comprehensively test database refactors. Successfully identified 5 critical bugs (100% system failure rate) in QuestDB migration that would have shipped in production.

When to Use This Skill

Use this skill when:

• Database refactors (e.g., v3.x file-based → v4.x QuestDB)
• Schema migrations requiring validation
• Bulk data ingestion pipeline testing
• System migrations with multiple validation layers
• Pre-release validation for database-centric systems

Key outcomes:

• Parallel agent execution for comprehensive coverage
• Structured validation reporting (VALIDATION_FINDINGS.md)
• Bug discovery with severity classification (Critical/Medium/Low)
• Release readiness assessment

Core Methodology

1. Validation Architecture (3-Layer Model)

Layer 1: Environment Setup

• Container orchestration (Colima/Docker)
• Database deployment and schema application
• Connectivity validation (ILP, PostgreSQL, HTTP ports)
• Configuration file creation and validation

Layer 2: Data Flow Validation

• Bulk ingestion testing (CloudFront → QuestDB)
• Performance benchmarking against SLOs
• Multi-month data ingestion
• Deduplication testing (re-ingestion scenarios)
• Type conversion validation (FLOAT→LONG casts)

Layer 3: Query Interface Validation

• High-level query methods (get_latest, get_range, execute_sql)
• Edge cases (limit=1, cross-month boundaries)
• Error handling (invalid symbols, dates, parameters)
• Gap detection SQL compatibility

2. Agent Orchestration Pattern

Sequential vs Parallel Execution:

Agent 1 (Environment) → [SEQUENTIAL - prerequisite]
  ↓
Agent 2 (Bulk Loader) → [PARALLEL with Agent 3]
Agent 3 (Query Interface) → [PARALLEL with Agent 2]

Dependency Rule: Environment validation must pass before data flow/query validation

Dynamic Todo Management:

• Start with high-level plan (ADR-defined phases)
• Prune completed agents from todo list
• Grow todo list when bugs discovered (e.g., Bug #5 found by Agent 3)
• Update VALIDATION_FINDINGS.md incrementally

3. Validation Script Structure

Each agent produces:

1. Test Script (e.g., test_bulk_loader.py)

- 5+ test functions with clear pass/fail criteria - Structured output (test name, result, details) - Summary report at end

1. Artifacts (logs, config files, evidence)
2. Findings Report (bugs, severity, fix proposals)

Example Test Structure:

def test_feature(conn):
    """Test 1: Feature description"""
    print("=" * 80)
    print("TEST 1: Feature description")
    print("=" * 80)

    results = {}

    # Test 1a: Subtest name
    print("\n1a. Testing subtest:")
    result_1a = perform_test()
    print(f"   Result: {result_1a}")
    results["subtest_1a"] = result_1a == expected_1a

    # Summary
    print("\n" + "-" * 80)
    all_passed = all(results.values())
    print(f"Test 1 Results: {'✓ PASS' if all_passed else '✗ FAIL'}")
    for test_name, passed in results.items():
        print(f"  - {test_name}: {'✓' if passed else '✗'}")

    return {"success": all_passed, "details": results}

4. Bug Classification and Tracking

Severity Levels:

• 🔴 Critical: 100% system failure (e.g., API mismatch, timestamp corruption)
• 🟡 Medium: Degraded functionality (e.g., below SLO performance)
• 🟢 Low: Minor issues, edge cases

Bug Report Format:

#### Bug N: Descriptive Name (**SEVERITY** - Status)

**Location**: `file/path.py:line`

**Issue**: One-sentence description

**Impact**: Quantified impact (e.g., "100% ingestion failure")

**Root Cause**: Technical explanation

**Fix Applied**: Code changes with before/after

**Verification**: Test results proving fix

**Status**: ✅ FIXED / ⚠️ PARTIAL / ❌ OPEN

5. Release Readiness Decision Framework

Go/No-Go Criteria:

BLOCKER = Any Critical bug unfixed
SHIP = All Critical bugs fixed + (Medium bugs acceptable OR fixed)
DEFER = >3 Medium bugs unfixed OR any High-severity bug

Example Decision:

• 5 Critical bugs found → all fixed ✅
• 1 Medium bug (performance 55% below SLO) → acceptable ✅
• Verdict: RELEASE READY

Workflow: Step-by-Step

Step 1: Create Validation Plan (ADR-Driven)

Input: ADR document (e.g., ADR-0002 QuestDB Refactor) Output: Validation plan with 3-7 agents

Plan Structure:

## Validation Agents

### Agent 1: Environment Setup

- Deploy QuestDB via Docker
- Apply schema.sql
- Validate connectivity (ILP, PG, HTTP)
- Create .env configuration

### Agent 2: Bulk Loader Validation

- Test CloudFront → QuestDB ingestion
- Benchmark performance (target: >100K rows/sec)
- Validate deduplication (re-ingestion test)
- Multi-month ingestion test

### Agent 3: Query Interface Validation

- Test get_latest() with various limits
- Test get_range() with date boundaries
- Test execute_sql() with parameterized queries
- Test detect_gaps() SQL compatibility
- Test error handling (invalid inputs)

Step 2: Execute Agent 1 (Environment)

Directory Structure:

tmp/e2e-validation/
  agent-1-env/
    test_environment_setup.py
    questdb.log
    config.env
    schema-check.txt

Validation Checklist:

• ✅ Container running
• ✅ Ports accessible (9009 ILP, 8812 PG, 9000 HTTP)
• ✅ Schema applied without errors
• ✅.env file created

Step 3: Execute Agents 2-3 in Parallel

Agent 2: Bulk Loader

tmp/e2e-validation/
  agent-2-bulk/
    test_bulk_loader.py
    ingestion_benchmark.txt
    deduplication_test.txt

Agent 3: Query Interface

tmp/e2e-validation/
  agent-3-query/
    test_query_interface.py
    gap_detection_test.txt

Execution:

# Terminal 1
cd tmp/e2e-validation/agent-2-bulk
uv run python test_bulk_loader.py

# Terminal 2
cd tmp/e2e-validation/agent-3-query
uv run python test_query_interface.py

Step 4: Document Findings in VALIDATION_FINDINGS.md

Template:

# E2E Validation Findings Report

**Validation ID**: ADR-XXXX
**Branch**: feat/database-refactor
**Date**: YYYY-MM-DD
**Target Release**: vX.Y.Z
**Status**: [BLOCKED / READY / IN_PROGRESS]

## Executive Summary

E2E validation discovered **N critical bugs** that would have caused [impact]:

| Finding | Severity | Status | Impact       | Agent   |
| ------- | -------- | ------ | ------------ | ------- |
| Bug 1   | Critical | Fixed  | 100% failure | Agent 2 |

**Recommendation**: [RELEASE READY / BLOCKED / DEFER]

## Agent 1: Environment Setup - [STATUS]

...

## Agent 2: [Name] - [STATUS]

...

Step 5: Iterate on Fixes

For each bug:

1. Document in VALIDATION_FINDINGS.md with 🔴/🟡/🟢 severity
2. Apply fix to source code
3. Re-run failing test
4. Update bug status to ✅ FIXED
5. Commit with semantic message (e.g., fix: correct timestamp parsing in CSV ingestion)

Example Fix Commit:

git add src/gapless_crypto_clickhouse/collectors/questdb_bulk_loader.py
git commit -m "fix: prevent pandas from treating first CSV column as index

BREAKING CHANGE: All timestamps were defaulting to epoch 0 (1970-01)
due to pandas read_csv() auto-indexing. Added index_col=False to
preserve first column as data.

Fixes #ABC-123"

Step 6: Final Validation and Release Decision

Run all tests:

/usr/bin/env bash << 'SKILL_SCRIPT_EOF'
cd tmp/e2e-validation
for agent in agent-*; do
    echo "=== Running $agent ==="
    cd $agent
    uv run python test_*.py
    cd ..
done
SKILL_SCRIPT_EOF

Update VALIDATION_FINDINGS.md status:

• Count Critical bugs: X fixed, Y open
• Count Medium bugs: X fixed, Y open
• Apply decision framework
• Update Status field to ✅ RELEASE READY or ❌ BLOCKED

Real-World Example: QuestDB Refactor Validation

Context: Migrating from file-based storage (v3.x) to QuestDB (v4.0.0)

Bugs Found:

1. 🔴 Sender API mismatch - Used non-existent Sender.from_uri() instead of Sender.from_conf()
2. 🔴 Type conversion - number_of_trades sent as FLOAT, schema expects LONG
3. 🔴 Timestamp parsing - pandas treating first column as index → epoch 0 timestamps
4. 🔴 Deduplication - WAL mode doesn't provide UPSERT semantics (needed DEDUP ENABLE UPSERT KEYS)
5. 🔴 SQL incompatibility - detect_gaps() used nested window functions (QuestDB unsupported)

Impact: Without this validation, v4.0.0 would ship with 100% data corruption and 100% ingestion failure

Outcome: All 5 bugs fixed, system validated, v4.0.0 released successfully

Common Pitfalls

1. Skipping Environment Validation

❌ Bad: Assume Docker/database is working, jump to data ingestion tests ✅ Good: Agent 1 validates environment first, catches port conflicts, schema errors early

2. Serial Agent Execution

❌ Bad: Run Agent 2, wait for completion, then run Agent 3 ✅ Good: Run Agent 2 & 3 in parallel (no dependency between them)

3. Manual Test Reporting

❌ Bad: Copy/paste test output into Slack/email ✅ Good: Structured VALIDATION_FINDINGS.md with severity, status, fix tracking

4. Ignoring Medium Bugs

❌ Bad: "Performance is 55% below SLO, but we'll fix it later" ✅ Good: Document in VALIDATION_FINDINGS.md, make explicit go/no-go decision

5. No Re-validation After Fixes

❌ Bad: Apply fix, assume it works, move on ✅ Good: Re-run failing test, update status in VALIDATION_FINDINGS.md

Resources

scripts/

Not applicable - validation scripts are project-specific (stored in tmp/e2e-validation/)

references/

• example_validation_findings.md - Complete VALIDATION_FINDINGS.md template
• agent_test_template.py - Template for creating validation test scripts
• bug_severity_classification.md - Detailed severity criteria and examples

assets/

Not applicable - validation artifacts are project-specific

Troubleshooting

Post-Execution Reflection

After this skill completes, reflect before closing the task:

1. Locate yourself. — Find this SKILL.md's canonical path before editing.
2. What failed? — Fix the instruction that caused it.
3. What worked better than expected? — Promote to recommended practice.
4. What drifted? — Fix any script, reference, or dependency that no longer matches reality.
5. Log it. — Evolution-log entry with trigger, fix, and evidence.

Do NOT defer. The next invocation inherits whatever you leave behind.