Claude Architect · 课时

PostToolUse 与外发调用钩子

拦截工具结果，并阻止违反策略的操作。

PostToolUse 与外发调用钩子 是 CoddyKit 上的免费 Claude Architect 课时。这是第 1 节课，共 4 节。你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。这是 Claude Architect 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 Claude Architect 课程共包含 4 节课。

本课时的部分内容尚未翻译，以英文显示。

Why Hooks Exist

Prompts steer the model, but they are probabilistic — roughly 90% reliable. For most behavior that is fine. But some actions must never slip through: issuing a large refund, deleting production data, sending money.

Hooks give you 100% deterministic enforcement. They are code that runs around tool execution, outside the model's discretion. Use hooks when a failure has financial, legal, or safety consequences.

Two Hook Points to Know

This lesson covers two enforcement points:

PostToolUse — fires after a tool runs and intercepts the result before the model sees it. You can trim, validate, redact, or reshape tool output.
Outgoing-call hooks — block a policy-violating action before it leaves your system (e.g. a refund above a threshold).

Together they bracket the dangerous part of the agentic loop: the moment a tool result enters the conversation, and the moment a side effect would escape it.

Where Hooks Sit in the Loop

Recall the agentic loop: send request, inspect stop_reason, and when it is tool_use, run the tool and append the result to history, then repeat until end_turn.

A PostToolUse hook wraps that 'run the tool' step. The model requests a tool call; your code executes it; before appending the result to the message history, the hook inspects and may modify it.

def run_tool_with_hook(tool_name, tool_input, tool_use_id):
    raw_result = execute_tool(tool_name, tool_input)
    # PostToolUse: inspect/transform BEFORE the model sees it
    safe_result = post_tool_use_hook(tool_name, raw_result)
    return {
        "type": "tool_result",
        "tool_use_id": tool_use_id,
        "content": safe_result,
    }

PostToolUse: Trim Verbose Output

A common, practical PostToolUse job is trimming tool output to the relevant fields. Raw API responses are often huge, and bloated context triggers the lost-in-the-middle problem — models attend to the start and end of context more than the middle.

The hook reshapes the result deterministically, so the model only ever sees the fields that matter.

def post_tool_use_hook(tool_name, raw_result):
    if tool_name == "lookup_order":
        # Keep only fields the model needs; drop the rest
        return {
            "order_id": raw_result["id"],
            "status": raw_result["status"],
            "total": raw_result["total"],
        }
    return raw_result

PostToolUse: Structured Errors, Not Generic Ones

PostToolUse is also where you normalize errors. A generic status like "Operation failed" blocks recovery — the model cannot tell a transient fault from a permission problem.

Reshape failures into structured errors so the model can route intelligently: a flag, a category, retryability, and context.

def post_tool_use_hook(tool_name, raw_result):
    if raw_result.get("error"):
        return {
            "isError": True,
            "errorCategory": "transient",   # transient/validation/business/permission
            "isRetryable": True,
            "message": raw_result["error"],
            "attempted_query": raw_result.get("query"),
            "partial_results": raw_result.get("partial", []),
        }
    return raw_result

Distinguish Failure from Empty

One subtle rule PostToolUse helps enforce: an access failure is not the same as a valid empty result.

A failed query (timeout, auth) is an error the model may retry.
An empty result (zero matching orders) is a legitimate answer — retrying changes nothing.

Encode the difference so the model never burns iterations retrying a query that simply has no matches.

def post_tool_use_hook(tool_name, raw_result):
    if tool_name == "lookup_order":
        if raw_result.get("connection_error"):
            return {"isError": True, "errorCategory": "transient",
                    "isRetryable": True}
        # Empty is a VALID answer, not an error
        return {"orders": raw_result.get("orders", []),
                "isError": False}
    return raw_result

Outgoing-Call Hooks: The Hard Stop

Now the headline use case. The exam's Customer Support scenario has a tool process_refund. Policy: refunds over $500 require a human.

You could write that rule in the prompt — but a prompt is ~90% reliable, and a single missed refund is a real financial loss. So you wrap the outgoing call in a hook that blocks deterministically when the threshold is exceeded. The model's judgment never gets a vote here.

REFUND_LIMIT = 500

def outgoing_call_hook(tool_name, tool_input):
    if tool_name == "process_refund" and tool_input["amount"] > REFUND_LIMIT:
        # Block the side effect; hand control back to the model
        return {
            "blocked": True,
            "reason": "Refund over $500 requires human approval.",
            "next_action": "escalate_to_human",
        }
    return execute_tool(tool_name, tool_input)

Programmatic Preconditions

Outgoing-call hooks also enforce preconditions — ordering guarantees a prompt cannot reliably provide. Example: never process a refund until get_customer has returned a verified identity.

The hook checks a fact in your own state, not the model's claim that it 'already verified'. That is the difference between a deterministic guarantee and a hopeful instruction.

def outgoing_call_hook(tool_name, tool_input, session_state):
    if tool_name == "process_refund" and not session_state.get("verified_customer_id"):
        return {
            "blocked": True,
            "reason": "Identity not verified. Call get_customer first.",
        }
    return execute_tool(tool_name, tool_input)

Feed the Block Back to the Model

Blocking is only half the job. A hook that silently swallows the action leaves the model confused and the loop stalled — that is silent suppression, an anti-pattern.

Instead, return the block as a tool result the model can read and act on. A good block message names the reason and the correct next step (here, escalate_to_human), so the agent recovers cleanly inside the same loop.

def handle_tool_call(tool_name, tool_input, tool_use_id, state):
    outcome = outgoing_call_hook(tool_name, tool_input, state)
    if isinstance(outcome, dict) and outcome.get("blocked"):
        return {"type": "tool_result", "tool_use_id": tool_use_id,
                "is_error": True, "content": outcome["reason"]}
    return {"type": "tool_result", "tool_use_id": tool_use_id,
            "content": outcome}

Hooks vs Iteration Caps

Do not confuse enforcement hooks with the loop's safety net. An iteration cap stops a runaway loop, but it is a backstop — never the primary control mechanism, and never a substitute for policy enforcement.

Hooks are the opposite: a precise, intentional guarantee on a specific action. The model still drives decisions; hooks reserve hard code only for the guarantees that genuinely matter.

When to Reach for a Hook

Decision rule for the exam and for production:

Use a hook when failure is financial, legal, or safety-critical, or when you need a deterministic precondition or threshold (refund > $500, verified ID before payout).
Use a prompt for soft guidance, tone, and the ~90% of behavior where an occasional miss is acceptable.

Hooks = deterministic. Prompts = probabilistic. Match the tool to the cost of being wrong.

Quick Check

A support agent has a process_refund tool. Company policy: refunds above $500 must be approved by a human. Most refunds are small and automated. What is the architect-grade way to enforce this?

Recap

Key takeaways:

PostToolUse intercepts tool results before the model sees them — trim verbose output, normalize generic failures into structured errors, and distinguish access failure from a valid empty result.
Outgoing-call hooks block policy-violating actions (refund > $500) and enforce preconditions (verified ID before payout).
Hooks = 100% deterministic; prompts = ~90% probabilistic. Use hooks when failure is financial, legal, or safety-critical.
Always feed a block back as a structured result — never suppress silently. Iteration caps are a safety net, not enforcement.

常见问题解答

「PostToolUse 与外发调用钩子」课时是免费的吗？

是的 — 「PostToolUse 与外发调用钩子」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 Claude Architect 课程的其余内容，请升级到 CoddyKit PRO。 Claude Architect 课程共包含 4 节课。

「PostToolUse 与外发调用钩子」这节课中我会学到什么？

拦截工具结果，并阻止违反策略的操作。你通过在浏览器中直接运行的动手代码来练习 Claude Architect，全天候 AI 导师会在你学习这节课的过程中回答你的问题。

学习 Claude Architect 需要有经验吗？

无需任何先前经验。CoddyKit 上的 Claude Architect 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。这是第 1 节课，共 4 节。

「PostToolUse 与外发调用钩子」课时需要多长时间？

大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。

我能在这节 Claude Architect 课中编写并运行代码吗？

能。每节 Claude Architect 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。

此课程中的所有课时

← 返回 Claude Architect