Tn Nemesis Scan

Nemesis Scan

Deep combined Feynman + State Inconsistency audit orchestrated across 8 specialized agents, with dynamic domain pattern discovery.

Activation

• User says /nemesis-scan or nemesis scan or deep combined audit
• User wants maximum-depth business logic + state inconsistency coverage

Invocation Format

/nemesis-scan [target] [optional: --hints "domain description"] [optional: --fresh]

Examples:

• /nemesis-scan src/contracts/ — audit contracts, auto-discover domain
• /nemesis-scan src/consensus/ --hints "DAG-based BFT consensus" — audit with domain hints
• /nemesis-scan src/ --fresh — force domain pattern regeneration

Execution

1. Determine the target scope from the user's request (files, directories, or modules to audit)

1. Parse optional arguments:

• Domain hints: Extract text after --hints (quoted string). If absent, set to "none".
• Fresh flag: If --fresh is present, skip cache check and force domain discovery.

1. Create the output directory:

``bash mkdir -p .audit/nemesis-scan/research .audit/findings ``

1. Cache check — determine if domain discovery is needed:

``bash # Check if cached domain-patterns exist if [ -f .audit/domain-patterns.md ]; then head -5 .audit/domain-patterns.md fi ``

Read the header of .audit/domain-patterns.md if it exists. The cache is valid when ALL of:

• The file exists
• The _Git hash:_ line matches current git rev-parse --short HEAD
• The _Target scope:_ line matches the requested scope
• The _User hints:_ line matches the provided hints
• The --fresh flag was NOT passed

Set discovery_needed = true if cache is invalid or missing. Set discovery_needed = false if cache is valid.

1. Resolve the path to this skill's references/ directory for agent prompts.

1. Spawn the nemesis-orchestrator agent:

Agent({
  subagent_type: "nemesis-orchestrator",
  description: "Run nemesis-scan pipeline",
  prompt: "Run the full nemesis-scan pipeline on the following target scope:

Target: [target files/directories]
Domain hints: [user hints or 'none']
Discovery needed: [true/false]

References are at: [absolute path to skills/tn-nemesis-scan/references/]

All agents MUST read core-rules.md and language-adaptation.md.

[If discovery_needed is true:]
Run Phase -1 (Domain Discovery) FIRST:
- Phase -1a: Spawn nemesis-strategy with target scope, hints, and research-guide.md
- Phase -1b: Spawn parallel nemesis-researcher agents (one per research topic from strategy plan)
- Phase -1c: Compile fragments into .audit/domain-patterns.md with cache metadata

[If discovery_needed is false:]
Skip Phase -1. Domain patterns are cached at .audit/domain-patterns.md

Then execute all 8 phases in order, with all agents reading .audit/domain-patterns.md for domain context:
- Phase 0+1 (parallel): nemesis-recon + nemesis-mapper
- Phase 2: nemesis-feynman (full mode)
- Phase 3: nemesis-state-check (full mode)
- Phase 4: feedback loop (max 3 iterations)
- Phase 5: nemesis-journey
- Phase 6: nemesis-verifier
- Phase 7: nemesis-reporter

Additionally, load bundled domain references based on the target scope:
- If target includes epoch boundaries, system calls, validator lifecycle, ConsensusRegistry, StakeManager, region, shuffle, geographic, diversity, slash, applySlashes, burn, confiscate, committee size, or setNextCommitteeSize → also read references/protocol-contract-patterns.md
- If target includes consensus state, DAG, certificates, aggregators, state sync, network/peer management, batch, builder, pool, transaction ordering, subscriber, finality, CVV, NVV, observer, storage, persistence, archive, recovery, peer score, Kademlia, or reputation → also read references/consensus-dag-patterns.md

Write phase outputs to .audit/nemesis-scan/
Write final reports to .audit/findings/nemesis-scan-verified.md and .audit/findings/nemesis-scan-raw.md

Present a concise summary when complete."
})

1. After the orchestrator returns, relay its summary to the user with paths to the report files.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: grantkee
• Source: grantkee/claude-extensions
• License: MIT

Install and usage instructions live in the source repository linked above.