Deep Verify: Multi-Agent Verification Framework

Execute this skill to verify work through balanced expert analysis.

Execution Instructions

When invoked, you will:

0. Resolve scope and context — invoke context skill (always), then preflight if confidence is low

1. Populate verification scope from working_scope

2. Generate expert agents based on the extracted context:

   • Always spawn: Devil's Advocate, Integration Checker, Third-Party Reviewer
   • Dynamically spawn domain experts based on detected domains

   Domain expert selection via domain-registry: Read domain-registry/domains/*.md to match signals from working_scope. Minimum: 1 domain expert. No maximum.

3. Aggregate findings from all experts with proper weighting

4. Generate and save report to .outputs/verification/

---

Dependency Check

Before executing any step, verify all required skills are present:

[skills-root]/context/SKILL.md
[skills-root]/preflight/SKILL.md
[skills-root]/domain-registry/README.md

Where [skills-root] is the parent of this skill's directory. Resolve with ls ../ from this skill's location.

If any required file is missing → stop immediately and output:

⚠ Missing required skills for deep-verify:

  {missing-skill}
    Expected: {skills-root}/{missing-skill}/SKILL.md

Install the missing skill(s):
  git clone https://github.com/mikeng-io/agent-skills /tmp/agent-skills
  cp -r /tmp/agent-skills/skills/{missing-skill} {skills-root}/

Or install the full suite at once:
  cp -r /tmp/agent-skills/skills/ {skills-root}/

All dependencies present → proceed to Step 0.

---

Step 0: Scope & Context Resolution

Context (always required):

Invoke Skill("context") first. It classifies the artifact, detects domains from domain-registry, and determines routing confidence:

context_report:
  artifact_type: ""  # code | financial | marketing | creative | research | mixed
  domains: []        # matched domain names from domain-registry
  routing: ""        # parallel-workflow | debate-protocol | deep-council
  confidence: ""     # high | medium | low

Preflight (conditional — triggered by context confidence):

Invoke Skill("preflight") only if context_report.confidence == "low" OR one or more signals remain unresolved:

• Artifact is not clearly identified
• User's intent is unclear (what are we verifying?)
• Domains could not be detected
• Concerns or risks to verify are not stated

Preflight fills exactly the gaps context could not resolve (max 3 questions, one at a time):

scope_clarification:
  artifact: ""       # what to verify
  intent: "verify"
  domains: []        # supplements context_report.domains
  constraints: []    # explicit risks or concerns to focus on
  confidence: ""     # high | medium

If context_report.confidence == "high" → skip preflight entirely.

Merge into working scope:

working_scope:
  artifact: ""            # files, topics, or description of what to verify
  domains: []             # from context_report (authoritative), supplemented by preflight
  concerns: []            # from context signals and scope_clarification.constraints
  context_summary: ""     # combined description for expert agent prompts

Use working_scope throughout this skill.

---

Step 1: Populate Verification Scope

Using working_scope from Step 0, extract the verification context:

conversation_analysis:
  files: []        # from working_scope.artifact
  artifacts: []    # additional artifacts from working_scope
  topics: []       # key topics from context_report
  concerns: []     # from working_scope.concerns
  intent: ""       # from working_scope — what is being verified
  domain_inference: []  # from working_scope.domains

---

Step 2: Spawn Expert Agents

Spawn experts with dependency-aware execution for optimal analysis quality. Domain experts analyze first, then invariant experts use those findings.

Dependency-Aware Execution Strategy:

Verification has natural dependencies:

• Integration Checker needs domain findings to assess cross-system impact
• Devil's Advocate is more effective challenging concrete domain findings
• Third-Party Reviewer benefits from seeing complete analysis

Task Definitions with Dependencies:

tasks:
  # Wave 1: Domain analysis (foundation)
  - id: domain-experts
    description: "Domain-specific analysis"
    depends_on: []
    agents: [dynamic, based on conversation]

  # Wave 2: Critical analysis (needs domain findings)
  - id: integration-check
    description: "Assess system-wide integration impact"
    agent: "integration-checker"
    depends_on: [domain-experts]

  - id: devils-advocate
    description: "Challenge assumptions and find failure modes"
    agent: "devils-advocate"
    depends_on: [domain-experts]

  # Wave 3: Fresh perspective (needs complete analysis)
  - id: third-party-review
    description: "Fresh eyes review of all findings"
    agent: "third-party-reviewer"
    depends_on: [integration-check, devils-advocate]

execution:
  mode: dag-orchestrated
  waves:
    wave_1: [domain-experts]                        # N domain experts in parallel
    wave_2: [integration-check, devils-advocate]    # 2 tasks in parallel
    wave_3: [third-party-review]                    # 1 task

Why This Order Matters:

1. Domain experts first - Gather domain-specific findings and concerns
2. Integration and critique together - Integration Checker assesses impact, Devil's Advocate challenges assumptions, both using domain findings
3. Fresh eyes last - Third-Party Reviewer sees complete picture including domain analysis, integration impact, and challenges

Performance Benefit:

• Improves analysis quality - critics have concrete findings to work with
• Still maintains parallelism where possible (Wave 2 runs 2 tasks in parallel)
• Sequential where dependencies matter (Wave 3 needs everything)

Invariant Experts (Always Spawn)

Expert 1: Devil's Advocate

Weight: 40%
Purpose: Counter confirmation bias through pre-mortem analysis
Capability: highest

Use this prompt:

You are the DEVIL'S ADVOCATE. Your role is to BALANCE the verification by actively seeking what could go wrong, what we're not seeing, and what assumptions might be false.

## Your Mindset: Pre-Mortem
Imagine this work has already caused a failure 6 months from now. Work backwards: What went wrong? What did we miss?

## Focus Areas
- Hidden Assumptions: What are we assuming that might not be true?
- Failure Modes: If this fails, what happens? Who is affected?
- Silent Failures: Could this fail without anyone noticing?
- Edge Cases: What edge cases might we have missed?
- Rollback Reality: Can we undo this? How difficult?
- Negative Impacts: What does this break that we're not seeing?

## Context to Analyze
{conversation_context}

## Your Scope
{scope_description}

## Output Format (JSON)
{
  "agent": "devils-advocate",
  "pre_mortem_scenarios": [
    {
      "scenario": "What went wrong",
      "likelihood": "LOW | MEDIUM | HIGH",
      "impact": "LOW | MEDIUM | HIGH | CRITICAL",
      "evidence": "What suggests this could happen",
      "mitigation": "How to prevent (if anything)"
    }
  ],
  "hidden_assumptions": [
    {
      "assumption": "What we're assuming",
      "risk_if_false": "What happens if wrong"
    }
  ]
}

Expert 2: Integration Checker

Weight: 15%
Purpose: Assess system-wide impact
Capability: high

Use this prompt:

You are the INTEGRATION CHECKER. Your role is to assess the system-wide impact of the proposed changes.

## Focus Areas
- What other components/systems are affected?
- What coordination is needed?
- What dependencies exist?
- What could break elsewhere?

## Context
{conversation_context}

## Output Format (JSON)
{
  "agent": "integration",
  "affected_components": ["list of affected areas"],
  "dependencies": ["list of dependencies"],
  "coordination_required": ["what needs to be coordinated"],
  "risks": [{"area": "affected area", "risk": "description"}]
}

Expert 3: Third-Party Reviewer

Weight: 5%
Purpose: Fresh-eyes perspective
Capability: standard

Use this prompt:

You are a THIRD-PARTY REVIEWER seeing this for the first time. Provide fresh, unbiased feedback.

## Focus Areas
- Is the intent clear?
- Are there obvious gaps or confusion?
- What questions would a newcomer ask?

## Context
{conversation_context}

## Output Format (JSON)
{
  "agent": "third-party",
  "clarity_score": "HIGH | MEDIUM | LOW",
  "questions": ["questions a newcomer would ask"],
  "suggestions": ["constructive suggestions"]
}

Domain Experts (Spawn Based on Conversation)

For each domain detected in the conversation, generate a domain expert prompt:

Weight: 40% (shared across all domain experts)
Capability: high

Prompt Template:

You are a {DOMAIN} expert. Analyze the following work from a {DOMAIN} perspective.

## Context
{conversation_context}

## Your Focus
Identify issues, risks, and improvements specific to {DOMAIN}.

## Output Format (JSON)
{
  "agent": "{domain}",
  "findings": [
    {
      "severity": "CRITICAL | HIGH | MEDIUM | LOW",
      "description": "Issue or observation",
      "evidence": "Reference to specific content",
      "fix": "Recommendation if applicable"
    }
  ]
}

Spawn domain experts for each unique domain detected. Examples:

• "authentication" discussed → Spawn Security Expert
• "Figma design" mentioned → Spawn Design Expert
• "email campaign" discussed → Spawn Marketing Expert
• "policy update" discussed → Spawn HR/Legal Expert

Any domain can be detected - analyze the conversation and spawn appropriate experts.

---

Step 3: Aggregate Findings

After all experts complete, aggregate their findings:

Determine Verdict

• FAIL: Critical risks or blocking issues found
• CONCERNS: High risks identified
• PASS: No significant issues

Build Summary Table

Dimension	Result
Domain Correctness	✅ PASS / ⚠️ CONCERNS
Risk Assessment	Based on Devil's Advocate findings
Integration Impact	Based on Integration Checker findings

---

Step 4: Generate Report

Generate a markdown report with this structure:

# Deep Verify Report

**Verdict:** {PASS | CONCERNS | FAIL}
**Generated:** {timestamp}
**Domains Analyzed:** {list of domains}
**Experts Consulted:** {count} experts

## Summary

| Dimension | Result |
|-----------|--------|
| Domain Correctness | {result} |
| Risk Assessment | {result} |
| Integration Impact | {result} |

---

## Risk Assessment (Devil's Advocate)

### {IMPACT} Risk: {scenario}

**Pre-mortem:** {failure scenario}
**Likelihood:** {likelihood}
**Impact:** {impact}
**Evidence:** {evidence}
**Mitigation:** {mitigation}

{repeat for each scenario}

---

## Domain Expert Findings

### {Domain Name}

{findings from domain expert}

{repeat for each domain}

---

## Integration Impact

**Affected:** {affected components}
**Dependencies:** {dependencies}
**Coordination:** {coordination needed}

---

## Third-Party Perspective

**Clarity:** {clarity score}
**Questions:** {questions raised}
**Suggestions:** {suggestions}

---

Step 5: Validate Output Format

Before finalizing the report, validate it against the required format specification to ensure consistency.

Validation Gate

Spawn an output validator sub-agent using the Task tool:

Capability: standard

You are an OUTPUT VALIDATOR for deep-verify reports. Your role is to ensure format compliance.

## Files to Validate
- Markdown: {path_to_markdown_file}
- JSON: {path_to_json_file}

## Validation Instructions
Follow the validation procedure defined in: skills/deep-verify/validators/output-validator.md

## Schema Location
JSON Schema: skills/deep-verify/schemas/verification-report-schema.json

## Tasks
1. Load and validate JSON against schema
2. Validate markdown structure and required sections
3. Cross-check consistency between JSON and markdown
4. Generate validation report

## Output Format
Return validation result as JSON with:
- validation_status: PASS or FAIL
- Specific errors and warnings
- Suggestions for fixes

## Strictness
FAIL on any critical errors:
- Missing required fields
- Invalid enum values
- Type mismatches
- Missing required sections

Handling Validation Results

If validation PASSES:

• Proceed to Step 6 (Save Report)

If validation FAILS:

1. Display all errors and warnings to user
2. Provide specific suggestions for each violation
3. DO NOT save report as "latest"
4. Ask user if they want to:
   • Fix the issues and regenerate
   • Override validation (with explicit confirmation)
   • Cancel verification

Example failure output:

❌ Validation FAILED

JSON Errors:
- Missing required field: risk_assessment.scenarios
- Invalid verdict value: 'MAYBE' (must be PASS, CONCERNS, or FAIL)

Markdown Errors:
- Missing required section: ## Integration Impact
- Domain 'Security' listed in metadata but no findings section found

Suggestions:
1. Add risk_assessment.scenarios array with at least one scenario
2. Change verdict to one of the valid values
3. Add ## Integration Impact section
4. Add ## Security findings section or remove from domains_analyzed

Would you like to regenerate the report with corrections?

---

Step 6: Save Report

Artifact Output

Save to .outputs/verification/{YYYYMMDD-HHMMSS}-verification-{slug}.md with YAML frontmatter:

---
skill: deep-verify
timestamp: {ISO-8601}
artifact_type: verification
domains: [{domain1}, {domain2}]
verdict: PASS | FAIL | CONCERNS        # if applicable
context_summary: "{brief description of what was reviewed}"
session_id: "{unique id}"
---

Also save JSON companion: {timestamp}-verification-{slug}.json

No symlinks. To find the latest artifact:

ls -t .outputs/verification/ | head -1

QMD Integration (optional, progressive enhancement):

qmd collection add .outputs/verification/ --name "deep-verify-artifacts" --mask "**/*.md" 2>/dev/null || true
qmd update 2>/dev/null || true

Note: Only save reports that pass validation.

---

Step 7: Configuration (Optional)

The system uses these defaults unless overridden:

Expert Weights:

• Devil's Advocate: 40%
• Integration Checker: 15%
• Third-Party Reviewer: 5%
• Domain Experts: 40% (shared pool)

Output Directory: .outputs/verification/

These can be overridden via:

• Environment variables (e.g., DEEP_VERIFY_OUTPUT_DIR)
• Config files in .outputs/verification/config.yaml
• Command-line arguments

---

Multi-Model Second Pass (Optional)

If multi-model confidence is needed after verification:

Invoke deep-council in fallback mode with:

• review_scope: same scope as this verification
• context_summary: paste context_summary from this verification
• intensity: "standard" (or match this verification's intensity)

deep-council will run all available bridges (bridge-claude if Task tool accessible, plus any CLI bridges installed), providing cross-model confirmation of critical findings. Merge multi_model_confirmed findings from council report into this verification's final report.

---

Notes

• Model-agnostic: Uses capability levels ("highest", "high", "standard") not specific model names
• Domain-agnostic: Works for any domain detected in conversation
• Conversation-driven: All context extracted from what was discussed
• No triggers/keywords: Analyzes conversation naturally, doesn't match patterns
• Balanced: Devil's Advocate weight equals all domain experts combined to counter confirmation bias
• Multi-Model: Optionally follow with deep-council for cross-model confidence
• Context Routing: If the artifact is complex or multi-domain, invoke the context skill first to classify artifact type and determine optimal routing (parallel-workflow vs debate-protocol vs deep-council)
• DeepWiki (optional): For code artifacts, invoke Skill("deepwiki") before spawning domain experts if the codebase has a Devin-indexed wiki — provides architectural context that sharpens domain expert analysis. Non-blocking; skip if unavailable.