Judger

An evaluation workflow for measuring how well generated code, SQL, or general text answers perform. TDD here is not involved; the workflow runs code tests, executes SQL against SQLite, or uses a general text evaluator.

In plain words
What is it for?
Use it to calculate pass@k for coding tasks such as HumanEval or MBPP, check generated SQL by running it, and evaluate general text responses.
Why use it?
It turns model outputs into repeatable scores instead of relying on manual inspection. It also supports separate settings for evaluation models, batches, samples, and GPUs.

Skill for Claude CodeCodex

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add skills/opendcai/dataflow-loopai/judger
Any agent
npx skills add OpenDCAI/Dataflow-LoopAI --skill judger
Clone the repo
git clone --depth 1 https://github.com/OpenDCAI/Dataflow-LoopAI

Made for: Claude Code, Codex.

Per session 0 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 1,752 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00000 $0.01752
Opus 5 $0.00000 $0.00876
Sonnet 5 $0.00000 $0.00350
Haiku 4.5 $0.00000 $0.00175

Measured 3d ago against content hash b5c4d624ecae, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

Judger scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

skills/Judger/SKILL.md · 186 lines

How it starts

The opening of the file, as written. The whole thing — 186 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Judger Skill

Purpose

无 LangGraph 的独立评测流水线。支持三种任务类型:

  • code — 代码生成评测(human-eval / mbpp),计算 pass@k
  • text2sql — SQL 生成评测,SQLite 执行校验
  • general_text — 通用文本评测(One-Eval DataFlowEvalTool)

How to Invoke

唯一入口:loopai.skills.Judger.run()

DB_PATHTASK_ID 从环境变量自动获取:

DB_PATH=api/db/db.sqlite3 TASK_ID=<task_id> \
python -c "from loopai.skills.Judger import run; run()"

或通过 CLI:

DB_PATH=api/db/db.sqlite3 TASK_ID=<task_id> loopai-judger

Configuration

配置通过 Configer skill 写入 TaskModel.state,分两部分:

全局字段(state["judger"] 顶层,所有 bench 共享)

字段 默认值 说明
eval_model_path 模型路径(必填)
eval_temperature 0 采样温度
eval_top_p 0.95 Top-P 采样
eval_batch_size 10 批处理大小,bench 可覆盖
eval_case_num 10 每问题样本数,bench 可覆盖
eval_vllm_tensor_parallel_size 1 vLLM 张量并行数
eval_vllm_gpu_memory_utilization 0.9 vLLM GPU 显存利用率
cuda_visible_devices "0" 指定 GPU
output_dir "./outputs" 输出根目录

Bench 配置(state["judger"])

所有评测集通过 benchlistextra_benchlist 列表配置。格式必须是 JSON 数组[{...},{...}]),不是 JSONL(每行一个对象):

[{"name":"gsm8k","task_type":"general_text","problem_path":"/data/gsm8k/test.jsonl","eval_type":"key2_qa"},{"name":"human_eval","task_type":"code","problem_path":"/data/humaneval.jsonl","case_num":10}]
{
  "benchlist": [
    {
      "name": "gsm8k",
      "task_type": "general_text",
      "problem_path": "/data/gsm8k/test.jsonl",
      "eval_type": "key2_qa",
      "key_mapping": {}
    },
    {
      "name": "human_eval",
      "task_type": "code",
      "problem_path": "/data/humaneval.jsonl",
      "case_num": 10,
      "batch_size": 10,
      "format_type": ""
    },
    {
      "name": "bird_dev",
      "task_type": "text2sql",
      "problem_path": "/data/bird/dev.jsonl",
      "text2sql_dir": "/data/bird/dev_databases",
      "case_num": 10,
      "batch_size": 10
    }
  ],
  "extra_benchlist": []
}

Read the full file on GitHub · 186 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 3d ago First seen · 186 lines · 0 tokens per session scan A b5c4d624ecae

Subscribe to this mod's changes

Judger is a skill published in the GitHub repository OpenDCAI/Dataflow-LoopAI (22 stars, last pushed 3d ago), licensed Apache-2.0. It costs nothing until one of its globs matches a file; then it loads 1,752 tokens. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.

Related

Other skills, from other repositories

loopx-auto-research

Use when a LoopX worker is operating an auto-research lane, demo pane, frontier item, evidence packet, promotion/retirement decision, or visible tmux/Codex auto-research rehearsal. Identity must come from the LoopX role profile and quota/frontier packet; this skill only provides role-specific execution checklists…

huangruiteng/loopx · 79 tokens

loopx-change-quality

Qualify the exact final diff for a LoopX-managed goal. Use when goal policy enables changequalityqualification, before a non-trivial delivery or merge, and when producing or repairing an exact-scope quality receipt. The workflow is language-neutral, permits at most one policy-authorized safe-fix pass, and never grants…

huangruiteng/loopx · 75 tokens

loopx-material

Operate an explicitly activated LoopX Material Lifecycle for a connected project. Use for material-store inventory, lossless migration, candidate/archive transitions, exact-read-backed ranking, ranked-entry rebuilds, bounded Explore intake, owner-gated apply, rollback, and audit. Do not use for ordinary one-off…

huangruiteng/loopx · 74 tokens

loopx-pr-review

Use for /loopx-pr-review or evidence-backed PR queue review. Run loopx pr-review first, execute the capability-owned review plan for each selected exact head, then publish full bilingual PR reviews (complete Chinese five-block review plus one concise English verdict) that match the verified findings. Use…

huangruiteng/loopx · 79 tokens

loopx-self-repair

Diagnose and repair LoopX control-plane drift or agent behavior drift. Use when a LoopX task makes unexpectedly small progress, follows a stale or contradictory recommendedaction, ignores a higher-priority blocked item while doing fallback work, reports vague owner/user gates, loses todo projection, misaligns…

huangruiteng/loopx · 102 tokens

loopx-pr-program

Use when LoopX must manage a multi-PR or multi-MR delivery program across one or more repositories: inventory current change requests, reconcile new/merged/closed or retargeted work, preserve requirement and dependency priorities, maintain a roadmap document, or monitor material lifecycle/check/review changes over…

huangruiteng/loopx · 95 tokens