AgentSkillsCN

comprehensive-test

仅运行 5 个基础代理，进行快速质量检查。适用于开发过程中的快速反馈，跳过阶段验证器。

SKILL.md
--- frontmatter
name: comprehensive-test
description: Orchestrates 14 testing agents for complete SIOPV verification. Use when running full quality gates, before releases, or after major changes. Generates reports to ~/siopv/claude-verification-reports/.
disable-model-invocation: true
allowed-tools: Read, Write, Bash, Task

Comprehensive Test

Runs all 14 verification agents in parallel phases for complete SIOPV quality gates.

Anthropic Best Practices Applied (Feb 2026)

PracticeImplementation
Model routingHaiku for simple, Sonnet for analysis
Tool guidanceEach prompt specifies TOOLS TO USE / DO NOT USE
Effort budgets5-80 tool calls per agent
Parallel executionFoundation (5) and Phase (8) agents run in parallel
Extended thinkingComplex agents use REASONING steps
LLM-as-judgeOrchestrator verifies report format
CheckpointingMANIFEST.md tracks status
Progressive disclosureAgent prompts in separate files

Workflow

Step 1: Setup (YOU must do directly)

bash
TIMESTAMP=$(date +%Y-%m-%d-%H-%M)
mkdir -p ~/siopv/claude-verification-reports/$TIMESTAMP

Write MANIFEST using templates/manifest.md. Replace {TIMESTAMP} with actual value.

Save to: ~/siopv/claude-verification-reports/{TIMESTAMP}/MANIFEST.md

Step 2: Foundation Agents (Parallel)

Launch ALL 5 in parallel. Read each prompt file and pass to Task tool:

AgentPrompt FileModelSubagent
Best Practicesprompts/best-practices.mdhaikuExplore
Securityprompts/security.mdsonnetExplore
Hallucinationprompts/hallucination.mdsonnetExplore
Code Reviewprompts/code-review.mdsonnetExplore
Coverageprompts/coverage.mdhaikuBash

Invocation pattern:

code
Task(subagent_type="[Subagent]", model="[Model]", prompt="[Content from prompt file with {TIMESTAMP} replaced]")

Step 3: Phase Validators (Parallel)

Launch ALL 8 in parallel:

PhaseNameStatusPrompt
1IngestionActiveprompts/phase-1-ingestion.md
2RAG/CRAGActiveprompts/phase-2-rag.md
3ML ClassificationActiveprompts/phase-3-ml.md
4OrchestrationActiveprompts/phase-4-orchestration.md
5AuthorizationActiveprompts/phase-5-authorization.md
6DLPStubprompts/phase-stub.md
7HITLStubprompts/phase-stub.md
8OutputStubprompts/phase-stub.md

For stubs (6-8): Write stub report directly, no agent needed.

Step 3.5: Verify Report Quality (LLM-as-Judge)

After agents complete, YOU verify each report:

  1. •Read each report file
  2. •Check: Has **Status:** PASS or **Status:** FAIL?
  3. •Check: Has ## Quality Gate section?
  4. •Check: Non-empty and follows format?

If any check fails:

  • •Mark as ERROR in MANIFEST
  • •Note "Report format invalid"
  • •Continue with other agents

Step 4: Generate Summary (YOU must do directly)

Read all 13 reports, extract status, generate summary using templates/summary.md.

Save to: ~/siopv/claude-verification-reports/{TIMESTAMP}/00-COMPREHENSIVE-SUMMARY.md

Step 5: Finalize

  1. •Update MANIFEST.md status column (Pending → PASS/FAIL/ERROR)
  2. •Display summary to user
  3. •Report execution time

Quality Gates

See reference/quality-gates.md for threshold definitions.

Anti-Autopilot Rules

  1. •MANIFEST created by orchestrator - Not delegated
  2. •Agent prompts are self-contained - No external file reading
  3. •Output format embedded in prompt - Templates for consistency
  4. •Summary created by orchestrator - Not delegated
  5. •Explicit file paths - Full paths, not relative