specialist-engineer

# Specialist Engineer — 将作监外聘专家工程 > Based on Nexus Hyper Agent Team (31 domain experts, 341-assertion contract suite) and vibecosystem (321⭐, 136 agents, 260 skills, self-learning). Adapted for 三省六部 将作监 external specialist deployment and management. ## 🚨 Red Flags: DO NOT SKIP THIS SKILL | Excuse your brain will make | Why it's wrong | |------------------------------|----------------| | "Engineer can handle this, no need for a specialist" | Generalist agents lack domain depth. A security audit by engineer misses OWASP ASI-02 violations that a security specialist catches in 30 seconds | | "I'll just add a specialist prompt to the existing agent" | Specialist agents need isolated workspaces (Git worktrees), adversarial review (challenger gate), and trust calibration. Prompt hacking doesn't provide these | | "Hiring a specialist agent is too complex for a one-off task" | The gated hiring pipeline (research→synthesis→validation→challenger→register→verify→probation→promote) ensures quality. One bad specialist can corrupt the codebase | | "I can assess specialist quality by reading their output" | Output quality ≠ process quality. A specialist might produce correct code that introduces supply chain vulnerabilities. Adversarial review catches this | ## When to Use - Complex multi-file refactoring requiring domain expertise (crypto, distributed systems, compiler internals) - Security-critical code that needs adversary-minded review - Cross-stack work (frontend + backend + infra) where no single agent has full coverage - External contractor integration: hiring and managing specialist agents from the talent pool - Code that will be deployed to production and needs 341-assertion contract validation ## Architecture ``` ┌─────────────────────────────────────────────────────────┐ │ 将作监 Specialist Hub │ │ │ │ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ │ │ │ Talent Scout │ │ Recruiter │ │ Challenger │ │ │ │ detect gaps │ │ gated hiring │ │ adversarial │ │ │ │ 5-signal conf │ │ 8 phases │ │ review gate │ │ │ └──────────────┘ └──────────────┘ └──────────────┘ │ │ │ │ Specialist Pool: │ │ ┌─────────┐ ┌─────────┐ ┌──────────┐ ┌────────────┐ │ │ │ security│ │ crypto │ │ frontend │ │ distributed│ │ │ │ auditor │ │ engineer│ │ architect│ │ systems │ │ │ └─────────┘ └─────────┘ └──────────┘ └────────────┘ │ │ │ │ Validation: 341-assertion contract test suite │ │ Trust: Bayesian per-agent trust ledger │ └─────────────────────────────────────────────────────────┘ ``` ## Core Capabilities ### 1. Gated Hiring Pipeline When a coverage gap is detected, the talent scout + recruiter hire a new specialist: ``` Phase 1: Research — Study the domain, identify required expertise Phase 2: Synthesis — Draft specialist prompt + toolset spec Phase 3: Validation — Run against known test cases (contract suite) Phase 4: Challenger — Adversarial agent tries to break the specialist Phase 5: Register — Atomic registration in agent-registry Phase 6: Verify — Post-hire smoke test on real task Phase 7: Probation — 5-task observation period with trust score Phase 8: Promote — Full specialist status, trust ledger active ``` ### 2. Coverage Gap Detection ```bash # 将作监 continuously monitors for capability gaps hermes jiangzuojian detect-gaps # Output: # Gap detected: solidity-smart-contract (confidence: 0.87) # Signal 1: 3 recent tasks required Solidity knowledge # Signal 2: engineer failed 2/3 Solidity tasks # Signal 3: no registered specialist for Solidity # Signal 4: session-sentinel co-signed the gap # Signal 5: estimated cost of gap: $45/week in failed tasks # # Recommendation: HIRE solidity-specialist # Estimated cost: $15/week # ROI: 3x (saves $45/week, costs $15/week) ``` ### 3. Adversarial Review (Challenger Gate) Before any specialist output is accepted, an adversarial agent reviews it: ```bash hermes jiangzuojian challenger-review \ --specialist crypto-engineer \ --task "implement key derivation function" \ --output /tmp/kdf-implementation.py # Challenger tries to: # 1. Inject edge cases that break the implementation # 2. Find timing side-channels # 3. Verify constant-time operations # 4. Check for known-weak PRNG usage # 5. Validate against NIST SP 800-132 ``` ### 4. Trust Ledger (Bayesian Calibration) Every specialist has a trust score calibrated by task outcomes: ```yaml specialist: solidity-expert trust_ledger: total_tasks: 15 successes: 13 failures: 1 challenger_overrides: 1 trust_score: 0.87 # Bayesian posterior domain_scores: erc20: 0.95 erc721: 0.91 custom_contracts: 0.76 # weaker area security_audit: 0.98 trend: improving (+0.03 per 5 tasks) last_incident: "2026-05-20: reentrancy vulnerability missed (caught by challenger)" ``` ### 5. Multi-Model Routing 将作监 routes tasks to the optimal specialist + model combination: ```bash hermes jiangzuojian route \ --task "audit smart contract for reentrancy vulnerabilities" \ --budget 2.00 # Router decision: # Specialist: solidity-expert (trust: 0.98 on security_audit) # Model: claude-opus-4 (reasoning_effort=high, cost: $1.50) # Fallback: deepseek-v4-pro (if opus unavailable) # Estimated: 3 tool calls, $1.80, 45s ``` ## Specialist Roster (Example) | Specialist | Domain | Trust | Model | Cost/task | |-----------|--------|-------|-------|-----------| | security-auditor | OWASP, injection, auth | 0.95 | claude-sonnet-4 | $0.80 | | crypto-engineer | ECDSA, KDF, TLS | 0.88 | claude-opus-4 | $1.50 | | distributed-systems | Raft, Paxos, CRDTs | 0.82 | deepseek-v4-pro | $0.60 | | performance-profiler | CPU, memory, I/O | 0.91 | haiku | $0.15 | | accessibility-expert | WCAG 2.2, ARIA | 0.94 | gemini-2.5-pro | $0.40 | ## Quick Start ```bash # Initialize specialist hub hermes jiangzuojian init # Hire first specialist hermes jiangzuojian hire \ --domain "security-audit" \ --from registry \ --phases 8 # Run a task with specialist hermes jiangzuojian execute \ --task "audit ~/.hermes/profiles/regent/scripts/kanban_gate.py" \ --specialist security-auditor ``` ## Reference: Upstream Projects | Project | Stars | License | Key Feature | |---------|-------|---------|-------------| | [Nexus Hyper Agent Team](https://github.com/asiflow/claude-nexus-hyper-agent-team) | - | - | 31 experts, 341-assertion suite, gated hiring | | [vibecosystem](https://github.com/vibeeval/vibecosystem) | 321 | - | 136 agents, 260 skills, self-learning | | [specialist-agent](https://github.com/HerbertJulio/specialist-agent) | 10 | MIT | 27 agents, 7 framework packs | ## Integration with 三省六部 将作监 operates as a specialist augmentation layer: ``` Kanban chain: planner → reviewer → shangshu → [engineer, 将作监] │ ┌─────────┘ ▼ 将作监 Specialist (external contractor) │ ┌─────┴─────┐ challenger trust ledger review calibration ``` 将作监 differs from 兵部 (engineer): - **兵部**: Standing department, generalist, always online - **将作监**: External specialist, domain expert, hired on-demand ## Common Pitfalls - **Specialist overload**: Hiring specialists for every task dilutes the trust ledger. Only hire when gap confidence > 0.75. - **Challenger bypass**: "The code looks fine, skip adversarial review." NEVER. The challenger gate is the last defense. - **Trust score blind spots**: A specialist with 0.95 trust might have 0.50 in a subdomain. Always check domain_scores, not just aggregate. --- ## ✅ Verification Checklist (RUN BEFORE DEPLOYING SPECIALIST OUTPUT) - [ ] Did the specialist pass all 8 phases of the gated hiring pipeline? - [ ] Did the challenger agent review the output (not just the hiring agent)? - [ ] Did I check the specialist's trust score for the SPECIFIC domain (not aggregate)? - [ ] Is the specialist working in an isolated workspace (Git worktree, not shared)? - [ ] Did I log the task outcome to update the trust ledger? **If any box is unchecked, go back.**