gh-cipherstash-cipherpowers-plugin/commands/verify.md

Verify

Generic dual-verification dispatcher for high-confidence verification across all verification types.

Core principle: Agents cannot be trusted. Two independent agents + systematic collation = confidence.

Usage

/cipherpowers:verify <type> [scope] [--model=<sonnet|opus|haiku>]

Model guidance:

• opus - Deep analysis, security-critical verification, complex codebases
• sonnet - Balanced quality/speed (default for most verification types)
• haiku - Quick checks, simple verifications, execute adherence checks

Algorithmic Workflow

Decision tree (follow exactly, no interpretation):

1. What verification type is requested?

   • code → Dispatch to code verification workflow
   • plan → Dispatch to plan verification workflow
   • execute → Dispatch to execute verification workflow
   • research → Dispatch to research verification workflow
   • docs → Dispatch to documentation verification workflow
   • OTHER → Error: Unknown verification type. Valid types: code, plan, execute, research, docs

2. MANDATORY: Skill Activation

Load skill context: @${CLAUDE_PLUGIN_ROOT}skills/dual-verification/SKILL.md

Step 1 - EVALUATE: State YES/NO for skill activation:

• Skill: "cipherpowers:dual-verification"
• Applies to this task: YES/NO (reason)

Step 2 - ACTIVATE: If YES, use Skill tool NOW:

Skill(skill: "cipherpowers:dual-verification")

⚠️ Do NOT proceed without completing skill evaluation and activation.

3. FOLLOW THE SKILL EXACTLY:

   • Phase 1: Dispatch 2 specialized agents in parallel (see dispatch table)
   • Phase 2: Dispatch review-collation-agent to compare findings
   • Phase 3: Present collated findings to user with confidence levels

4. STOP when verification is complete.

Dispatch Table

Type	Agent	Focus	Default Model
code	cipherpowers:code-review-agent + cipherpowers:code-agent	Heterogeneous review (Standards + Engineering)	sonnet
plan	cipherpowers:plan-review-agent + cipherpowers:code-agent	Plan quality + Technical feasibility	sonnet
execute	cipherpowers:execute-review-agent ×2	Plan adherence, implementation match	haiku
research	cipherpowers:research-agent ×2	Information completeness, accuracy	sonnet
docs	cipherpowers:technical-writer + cipherpowers:code-agent	Docs structure + Code example accuracy	haiku

Model parameter rules:

• If user specified --model=X → pass model: X to ALL dispatched agents
• If no model specified → use default model from table above
• Collation agent always uses haiku (simple comparison task)

Verification Types

Code Verification

When to use: Before merging, after significant implementation.

What it checks:

• Code quality and standards compliance
• Testing coverage and quality
• Security considerations
• Performance implications
• Maintainability

Workflow:

/verify code [scope] [--model=<sonnet|opus|haiku>]

→ Dispatches 1 code-review-agent and 1 code-agent in parallel
  (with model parameter if specified, otherwise sonnet)
→ Each agent independently reviews:
  - Read code changes
  - Run tests and checks
  - Review against standards
→ Dispatches review-collation-agent (always haiku)
→ Produces collated report with confidence levels

Plan Verification

When to use: Before executing implementation plans.

What it checks:

• 35 quality criteria (security, testing, architecture, etc.)
• Blocking issues that must be fixed
• Non-blocking improvements to consider

Workflow:

/verify plan [plan-file] [--model=<sonnet|opus|haiku>]

→ Dispatches 1 plan-review-agent and 1 code-agent in parallel
  (with model parameter if specified, otherwise sonnet)
→ Each agent independently evaluates against criteria
→ Dispatches review-collation-agent (always haiku)
→ Produces collated report with confidence levels

Execute Verification

When to use: After each batch during /execute workflow.

What it checks:

• Each task implemented exactly as plan specified
• No skipped requirements
• No unauthorized deviations
• No incomplete implementations

What it does NOT check:

• Code quality (that's code verification)
• Testing strategy (that's code verification)
• Standards compliance (that's code verification)

Workflow:

/verify execute [batch-number] [plan-file] [--model=<sonnet|opus|haiku>]

→ Dispatches 2 execute-review-agent agents in parallel
  (with model parameter if specified, otherwise haiku)
→ Each agent independently verifies:
  - Read plan tasks for batch
  - Read implementation changes
  - Verify each task: COMPLETE / INCOMPLETE / DEVIATED
→ Dispatches review-collation-agent (always haiku)
→ Produces collated report with confidence levels

Research Verification

When to use: When exploring unfamiliar topics, APIs, patterns, or codebases.

What it checks:

• Information completeness (did we find everything relevant?)
• Accuracy (are findings correct?)
• Multiple perspectives (different angles covered?)
• Gaps identified (what's missing?)

Examples:

• "How does authentication work in this codebase?"
• "What are the patterns for Bevy 0.17 picking?"
• "How should we structure the API layer?"

Workflow:

/verify research [topic] [--model=<sonnet|opus|haiku>]

→ Dispatches 2 research-agent agents in parallel
  (with model parameter if specified, otherwise sonnet)
→ Each agent independently explores:
  - Different entry points
  - Multiple sources (codebase, web, docs)
  - Different perspectives
→ Dispatches review-collation-agent (always haiku)
→ Produces collated report:
  - Common findings (high confidence)
  - Unique insights (worth knowing)
  - Divergences (needs clarification)

Documentation Verification

When to use: Auditing documentation accuracy.

What it checks:

• File paths exist
• Commands work
• Examples accurate
• Structure complete

Workflow:

/verify docs [files] [--model=<sonnet|opus|haiku>]

→ Dispatches 1 technical-writer and 1 code-agent in parallel
  (with model parameter if specified, otherwise haiku)
→ Each agent independently verifies against codebase
→ Dispatches review-collation-agent (always haiku)
→ Produces collated report with confidence levels

Why Dual Verification?

Problem: Single agent can miss issues, hallucinate, or confirm biases.

Solution: Two independent agents catch what one misses.

Confidence levels:

• VERY HIGH: Both agents found → Act on this
• MODERATE: One agent found → Consider carefully
• INVESTIGATE: Agents disagree → User decides

Example (research):

Agent #1: "Auth uses JWT with 1-hour expiry"
Agent #2: "Auth uses JWT with 24-hour refresh tokens"

→ Collation: Both partially correct (access vs refresh)
→ Higher confidence understanding than single agent

Integration with Other Commands

Execute workflow uses verify for batch verification:

/execute workflow:
  → Batch 1 (3 tasks)
  → /verify code (quality/standards)
  → /verify execute (plan adherence)
  → Fix all BLOCKING issues
  → Repeat for next batch

Related Commands

• /cipherpowers:execute - Plan execution workflow (uses /cipherpowers:verify for batch verification)

Related Skills

• dual-verification - Core pattern for all dual-verification
• executing-plans - Plan execution workflow integrating verification

Related Agents

• code-review-agent & code-agent - Code quality verification
• plan-review-agent & code-agent - Plan quality verification
• execute-review-agent - Plan adherence verification
• research-agent - Research verification
• technical-writer & code-agent - Documentation verification
• review-collation-agent - Generic collation (works for all types)

Remember

• All verification types use dual-verification pattern
• Dispatch table determines which agents to use
• Collation agent is always the same (generic)
• Confidence levels guide user decisions
• Agents cannot be trusted - that's why we use two