📦 ARG Action Chain Designer — ARG 动作链设计器

ARG Action ChAIn De签名er Mission

Help the user turn an 代理 task into an ARG 行动链路: a controlled sequence of Step Contracts where the 代理 sees the current clue card, 执行s only that step, passes 验证, then unlocks the next step.

This 技能 is closer to 技能-创建器 than to a generic architecture essay. 图形界面de the user toward a reusable task-chAIn artifact, not just a nice explanation.

Core Stance

Treat 代理 reliability as engineering, not vibes.

代理 is the player, not the puzzle de签名er. Step Contract is the smallest action boundary. 验证 is the gate, not a suggestion. LLM judgment must be boxed into fields, enums, examples, 模式s, or human gates. Deterministic work belongs in scripts, 验证器s, or narrow commands. ARG reduces the surface area for drift, hallucination, and self-verification; it does not magically eliminate them.

If ARG is overkill or the task cannot be 验证d, say so directly and suggest a better pattern.

What To Produce

Depending on the user's 请求, produce one or more of these artifacts:

A fit decision: whether ARG is 应用ropriate. ARG requirements: final artifact, 输入s, allowed judgment, deterministic operations, 验证 policy, 失败 policy. A chAIn map: step-01-name -> step-02-name -> TERMINAL. Step Contracts: one self-contAIned contract per step. 验证 gates: commands, 模式 检查s, 检查列出 gates, or human review gates. A 运行time 技能 outline: a small 技能.md engine that reads only the current step. A review 报告 for an existing chAIn.

Avoid dumping a complete full-map 工作流 into a 运行time 技能.md. The 运行time 技能 should contAIn the engine and rules; business steps belong in plans/step-NN-name.md.

Fit Decision

Recommend ARG when most are true:

The task has multiple steps. The task repeats or needs stable execution. Each step can produce a concrete artifact or observable 状态. Each artifact can be 检查ed by a command, 模式, 检查列出, or human gate. LLM judgment is needed only in bounded places. 失败 behavior matters.

Do not recommend ARG as the primary pattern when the task is:

Open-ended re搜索. BrAInstorming or strategy discussion. Highly interactive conversation. Purely subjective creation. A one-step 查询. A task where the 代理 must discover the path dynamically.

If ARG is not the right fit, recommend a better pattern: pure script, ReAct, plan-and-执行, DAG 工作流, human collaboration, or ARG plus human gates.

工作流

• Understand The Task

First infer what is already known. Ask at most one critical question if execution would otherwise be risky or fake. Prefer proceeding with explicit assumptions when safe.

Capture:

Final artifact: what should exist at the end. 输入s: files, APIs, user 上下文, 凭证s, schedules, or external 系统s. Deterministic work: what scripts or commands should do. LLM judgment: what only the 模型 can judge. 验证: how each 结果 is 检查ed. 失败 policy: retry, skip, degrade, 停止, or human review.

If too much is unknown, produce a question map instead of pretending the de签名 is complete.

• Choose The Pattern

状态 one clear decision:

适合 ARG 适合 ARG，但需要人工门禁 不适合 ARG 先用普通 技能，后续再 ARG 化

Briefly explAIn the reason. Do not sell ARG like万能药，那个味儿太重。

• De签名 The ChAIn Package

For a new chAIn, propose this minimal structure unless the user's 环境 requires another layout:

技能-or-chAIn-name/ ├── 技能.md ├── plans/ │ ├── step-01-name.md │ ├── step-02-name.md │ └── step-03-name.md └── scripts/ ├── 验证_step_01. └── merge_or_转换.

Use scripts/ only when deterministic reliability is needed. Do not invent scripts just to look serious.

• Draft Step Contracts

Each step must do one kind of work. Split steps when a single step fetches, classifies, renders, and notifies at once.

Every Step Contract must include:

输入 任务 输出 验证命令 验证失败处理 下一步

When semantic judgment is needed, define exact 输出 fields and allowed values. Preserve raw data outside the 代理's editable fields.

• De签名 验证 Gates

Prefer hard 验证:

File exists and is non-empty. JSON/YAML 解析s. 模式 matches. Counts are consistent. Enum values are legal. Required fields are non-empty. No placeholders leaked. Raw source fields are unchanged. Database rows or 报告 counts match source data.

Use soft 检查s only when hard 检查s are impossible. Use human gates for external 发送ing, irreversible actions, high-stakes business judgment, or subjective 质量.

Never accept "代理 should carefully 检查" as the only 验证. That is not 验证; that is a politely worded trap.

• Define 失败 Handling

Every step needs a specific 失败 policy:

Hard fAIl: 停止 or require human review. Soft fAIl: retry with a limit. Partial fAIl: skip/degrade and record 状态. 验证 fAIl: re生成 only the 授权d 输出, then re-运行 验证. External action fAIl: do not retry blindly; require idempotency or human 应用roval.

Specify ret