AgentSkillsCN

17th-task-inspector

利用GPT-5.2-Codex对实现过程进行深度检验，对照TASK.md文档。 提供详尽的跨模型验证，并附带细致的通过/未通过分析。 适用场景：最终验证、关键功能验收，以及在宣称“已完成”之前进行的严格把关。 不适用于：快速检查（建议使用task-validator）、样式审查、测试执行。

SKILL.md
--- frontmatter
name: 17th-task-inspector
description: |
  Use this skill for deep inspection of implementation against TASK.md using GPT-5.2-Codex.
  Provides thorough cross-model verification with detailed pass/fail analysis.
  Appropriate for: final verification, critical features, before claiming "done".
  Not appropriate for: quick checks (use task-validator), style review, test execution.

Task Inspector

Deep inspection of implementation against task specification using GPT-5.2 / GPT-5.2-Codex.

Usage

code
/task-inspector                    → gpt-5.2-codex, high (default)
/task-inspector high               → gpt-5.2-codex, high
/task-inspector xhigh              → gpt-5.2-codex, xhigh
/task-inspector high CODEX         → gpt-5.2-codex, high
/task-inspector xhigh CODEX        → gpt-5.2-codex, xhigh
/task-inspector high NONE          → gpt-5.2, high
/task-inspector xhigh NONE         → gpt-5.2, xhigh

Argument Parsing

Format: /task-inspector {effort} {model}

ArgumentValuesDefault
efforthigh, xhighhigh
modelCODEX, NONECODEX

Model mapping:

  • •CODEX → gpt-5.2-codex (optimized for coding)
  • •NONE → gpt-5.2 (general model)

When to Use

  • •After implementation complete, before claiming "done"
  • •Critical features requiring high confidence
  • •Cross-model verification needed

Implementation

Parse arguments and invoke the 17th-task-inspector subagent:

code
Task(
    description="Inspect implementation ({model}, {effort})",
    subagent_type="glm-worker:17th-task-inspector",
    prompt="""
Inspect implementation against {task_file_path}.

Configuration:
- Model: {model}  (gpt-5.2-codex or gpt-5.2)
- Effort: {effort}  (high or xhigh)
- Project root: {project_root}
""",
    run_in_background=true
)

Variables

VariableDefaultDescription
efforthighhigh or xhigh
modelgpt-5.2-codexgpt-5.2-codex or gpt-5.2
task_fileTASK.mdTask specification file
task_file_path./TASK.mdFull path to task file
project_rootCurrent directoryProject root

Output

  • •Inspection report: .kkachi/{branch_name}/completed/task-inspector-{epoch_timestamp}.md
  • •Verdict: PASS / PARTIAL / FAIL
  • •Summary: X/Y requirements met, issues found

Difference from task-validator

task-validatortask-inspector
ModelGLM-4.7GPT-5.2(-Codex)
PurposePre-PR quick checkDeep verification
DepthSurface matchingSemantic analysis
CostLowMedium-High

Notes

  • •Subagent uses Codex in read-only sandbox
  • •Results are verified before reporting
  • •Ensure .kkachi/ is in .gitignore