| name | json-firs-fnnc-cli-patt |
| description | Absorb agent-facing patterns from trading/financial CLIs such as kraken-cli without blindly integrating live execution. Use when evaluating exchange CLIs, paper-trading shells, MCP-enabled finance tools, or when porting their JSON-first subprocess contracts, stable error taxonomies, paper-first workflow patterns, and machine-readable tool catalogs into Hermes or trading research systems like ict-engine. |
| version | 1.0.0 |
| author | Hermes Agent |
| license | MIT |
JSON-first Financial CLI Patterns
适用:
- 外部金融 CLI / 交易 CLI / paper trading CLI 评估
- 吸收 exchange CLI 的 agent contract,而非直接接 live trading
- 为 ict-engine / Hermes 设计 json-first adapter、stable error taxonomy、paper-first workflow
不适用:
- 直接下真实单
- 凭 README 盲信 live-trading 接入
核心原则
- 先安全审查,再看功能。
- 默认吸收“agent contract”,不默认吸收“live execution”。
- 优先抽取:
- JSON-first subprocess contract
- stable error categories
- paper/sim before live
- machine-readable tool catalog
- workflow skill/recipe packaging
- 若工具可触真实资金、账户、提现、杠杆:视为高风险。
吸收清单
1. JSON-first subprocess contract
目标:stdout 只放结构化 payload;stderr 只放诊断;退出码表达成功/失败。
建议模式:
somecli <command> ... -o json 2>/dev/null
要求:
- success -> stdout valid JSON
- failure -> stdout JSON error envelope
- exit code 0 success, non-zero failure
- 不靠人类提示文字做路由
2. Stable error taxonomy
至少抽成稳定类别:
api
auth
network
rate_limit
validation
config
io
parse
规则:
- 路由看 category,不看 message 文案
- retry 只对明确可重试类开放
3. Paper-first workflow
若外部工具有 live 与 paper/sim 双面:
- 默认先 paper
- 任何 live path 都需额外确认与权限隔离
- 在研究系统中,paper/sim 是 first-class,不是附属脚本
4. Machine-readable tool contract
优先寻找:
tool-catalog.json
error-catalog.json
CONTEXT.md
AGENTS.md
CLAUDE.md
吸收用途:
- 参数 schema
- dangerous flags
- auth requirements
- output contracts
- retry guidance
5. Workflow skills / recipes
若上游已把多步流程打包成 skills/recipes:
- 抽 workflow boundary
- 抽 safety boundary
- 抽 reusable sequence
- 不盲搬交易策略本身
针对 ict-engine 的落地建议
优先可吸收面:
- 外部 market-data / exchange adapter 的 JSON envelope 统一
- adapter 错误分类统一
- paper/backtest/sim 与任何未来 live path 对称化
- 将常见研究流打包为 skills/prompts,而非只堆 CLI flags
- 若未来暴露工具目录,做 machine-readable catalog
不建议默认吸收:
- live order execution
- withdrawal / transfer / staking
- shared MCP trading server
- 任何无用户明确批准的危险命令
评估结论模板
- Verdict: learn / absorb / install / reject
- Absorb:
- json contract
- error taxonomy
- paper-first symmetry
- skill packaging
- Reject:
- unsafe live automation by default
产物建议
若任务命中 repo/documentation 吸收:
- 写 repo docs note
- 写 architecture boundary rule
- 只在用户明确要求时再做实际 live integration