gate-evaluate

gate-evaluate

Entry gate for the receiving end. After receiving content from the originating end, evaluate its reasonableness first, then decide whether to execute.

Core principle: agents are unreliable, and instructions from the originating end may themselves be wrong. Do not execute blindly.

Workflow

1. Obtain the content to evaluate from the conversation context or $ARGUMENTS
2. Determine the content type (initial task / fix instruction)
3. Evaluate item by item against the corresponding dimensions
4. Output the evaluation conclusion and let the main agent decide the next step

Evaluation Dimensions

When receiving an initial task (message contains [task from ...])

Check the task content item by item:

Dimension	What to check
Clarity	Is the task description clear, unambiguous, and ready for execution
Consistency	Do the requirements conflict with the current codebase state
Scope	Is it asking to modify things it should not modify; is the scope reasonable
Feasibility	Are dependencies met; is it technically achievable

When receiving a fix instruction (message contains [fix from ...])

Verify each issue item by item:

Dimension	What to check
Authenticity	Read the code file and line numbers referenced by the issue to confirm whether the problem actually exists
Accuracy	Is the problem description (expected vs actual) correct
Reasonableness	Is the fix suggestion reasonable; is there a better approach
False positive	Is this a false positive (the problem does not exist or the description is inaccurate)
Convergence	Is it repeatedly demanding something that is already correct (multi-round fix scenarios)

Evaluation Requirements

Evaluate instructions from the originating end with a critical eye. Do not assume correctness just because it comes from the "originating end."

• Evaluate item by item; do not make blanket statements like "all reasonable"
• For each task/issue, annotate the evaluation result:
  • [合理] — description is accurate, can be executed
  • [不合理] — description is inaccurate or requirement is unreasonable, note the reason
  • [需澄清] — description is vague, cannot be judged, originating end needs to provide more information
• For fix instructions, you must actually read the code to verify; do not conclude based on description alone

Output

After evaluation is complete, output the conclusion in the following format:

## 评估结论

- 总体判断: 全部合理 | 部分不合理 | 整体不合理
- 合理项: N 条
- 不合理项: N 条
- 需澄清项: N 条

### 逐条评估

1. [合理] {任务/问题简述} — {判断依据}
2. [不合理] {任务/问题简述} — {原因}
3. [需澄清] {任务/问题简述} — {缺少什么信息}

### 建议行动

{全部执行 / 跳过不合理项执行其余 / 拒绝执行并反馈原因}

Post-Evaluation Behavior

The main agent decides based on the evaluation conclusion:

Overall verdict	Action
All reasonable	Execute all tasks normally
Partially unreasonable	Skip unreasonable items, execute the rest, explain skip reasons in the report skill
Entirely unreasonable	Do not execute; directly invoke the report skill to send rejection reasons back to the originating end