🎯 场景：tool call 通过不了 schema 校验时（类型错、缺必填字段等），按错误信息修复参数。保留原意图，不许伪造缺失字段——必填字段不能从原 call 推断时直接 escalate。

Quick Use

Use when: A tool call failed schema validation and you want to repair it before escalating to a strategy reflection. Fill in: {{tool_schema}} = JSON schema for the tool's parameters; {{attempted_call}} = the malformed call object; {{error_message}} = the validator's error message. You'll get: A repaired call (or repairable: false if a required field cannot be inferred), plus a list of the changes made. Output is JSON.

Purpose

Take a malformed tool call (schema validation failure, type mismatch, missing required field) plus the error message, and emit a corrected call that satisfies the tool's parameter schema. Used inside agent loops to recover from a single bad call without escalating to the full reflection step or restarting the whole trajectory. Output is the repaired call as structured JSON, ready to re-submit to the tool layer.

Prompt

You repair a malformed tool call. The call failed validation; produce
a corrected version that satisfies the schema.

Tool schema (JSON Schema for the tool's parameters):
{{tool_schema}}

Attempted call:
{{attempted_call}}

Error message:
{{error_message}}

Rules:
1. Preserve the user-intent of the attempted call. Do NOT change tool
   choice, do NOT invent new arguments unrelated to the original
   intent.
2. Fix only what the error message indicates is wrong, plus any other
   schema violation you can see.
3. If a required field is missing AND its value cannot be inferred
   from the attempted call, set `repairable: false` and explain in
   `escalation_reason` rather than fabricating a value.
4. If the attempted call's tool name is not consistent with the
   schema (wrong tool entirely), set `repairable: false` — that is a
   strategy issue, not a repair issue.

Return ONLY this JSON object:
{
  "repairable": true | false,
  "repaired_call": {
    "name": "<tool name from schema>",
    "arguments": { ... }
  },
  "changes": ["<short description of each fix applied>"],
  "escalation_reason": "<if repairable=false: one-sentence reason; else empty string>"
}

If repairable=false, set "repaired_call" to the attempted_call
unchanged (so the caller has a consistent shape to handle).

Example

Input:

tool_schema: {
  "name": "search",
  "parameters": {
    "type": "object",
    "required": ["query", "max_results"],
    "properties": {
      "query": {"type": "string"},
      "max_results": {"type": "integer", "minimum": 1, "maximum": 20}
    }
  }
}
attempted_call: {
  "name": "search",
  "arguments": {"query": "Reykjavik population 2024", "max_results": "ten"}
}
error_message: "validation error at /arguments/max_results: expected integer, got string 'ten'"

Expected output:

{
  "repairable": true,
  "repaired_call": {
    "name": "search",
    "arguments": {"query": "Reykjavik population 2024", "max_results": 10}
  },
  "changes": ["Converted max_results from string 'ten' to integer 10."],
  "escalation_reason": ""
}

Failure Modes

• Over-repair / drifting intent — model also changes the query string ("Reykjavik population 2024" → "Iceland demographic statistics 2024") under the guise of "fixing" the call. Detect by comparing string fields between attempted and repaired calls; flag any change to fields not mentioned in the error message.
• Fabricating required fields — when a required field is missing, model invents a plausible value instead of escalating. The rubric forbids this; verify by spot-checking outputs where repairable=true and a required field appears in changes — its value should be derivable from the attempted call's other fields, not arbitrary.
• Tool swapping — model "fixes" the call by switching to a different tool that would have been more appropriate. The rubric forbids this; if it happens, return repairable: false so the reflection step can decide whether to switch tools.
• Schema misreading — model reads the schema's properties as required (or vice versa) and adds fields that should be optional. Mitigation: use a strict JSON-schema validator on the repaired call before submitting; on validation failure, escalate.
• Silent format drift — model returns valid JSON but with extra prose around it. Use JSON-mode if available, or a strict parser that rejects extras.

Tuning Notes

• 模型差异：本卡对 schema understanding 比对 reasoning 要求更高；中档 模型在简单类型修复（string→int）上稳定，在嵌套 schema / 联合类型 上失败率上升。复杂 schema 建议用 frontier 模型。
• 温度：0.0，repair 必须可重现。
• 调用预算：单次 tool call 失败立刻 repair 一次；连续 2 次 repair 失败 应触发反思（见 agent/self-critique-reflection），不要无限循环 repair。
• 与 agent/react-planner-with-tool-schema 的关系：planner 的失败处理 规则要求"不要立刻重试同样的 call"。本卡产生的 repaired_call 不算 "同样的 call"，可以直接重试。
• 与 agent/self-critique-reflection 的关系：本卡处理语法层失败； reflection 处理策略层失败。repairable: false 是一个明确的"升级到 reflection"信号。
• 安全考虑：repair 必须保留原 intent；这是为了防止恶意构造的 error message 诱导 agent 改写 query 去搜索其他内容。生产中要监控 intent-drift 率。

Changelog

• 0.1.0 — Initial card.