API Quality Check

Use the bundled script to run headless API-quality checks. Treat this skill as script-first: do not recreate LT-lite/B3IT-lite logic inline unless the script is clearly insufficient.

Provider names such as Ark/Volcengine, GLM, DeepSeek, Kimi, SiliconFlow, and similar services are examples only. The primary decision is the endpoint protocol type: OpenAI, OpenAI-Compatible, or Anthropic.

Quick start

Set the path once:

CODEBLOCK0

Run a capability smoke test first:

CODEBLOCK1

For many OpenAI-compatible endpoints, the same command also works if the user pastes the full .../chat/completions URL. The script will normalize it back to the API root automatically.

If you want a ready-to-run provider.json first, generate it with:

CODEBLOCK2

If an endpoint requires client-specific headers, put them in the config JSON as a headers object or pass them with --headers-json. For Kimi coding endpoints, use {"User-Agent":"KimiCLI/2.0.0"} only when the address is under https://api.kimi.com/coding; for the OpenAI-compatible Kimi path, use https://api.kimi.com/coding/v1.

If you already have multiple raw endpoint entries, normalize them into providers.json with:

CODEBLOCK3

Or run the full batch pipeline:

CODEBLOCK4

That command also creates ./api-quality-out/index.html as the landing page for all generated reports.

For one endpoint that you want to check every day and archive by date:

CODEBLOCK5

Workflow

1. Run smoke before any baseline or detect run.
If you have many endpoints, run batch-smoke with a config list before choosing which ones deserve deeper LT/B3IT work.
Read the result:

- b3it_supported=true: the endpoint can return normal first-token text at max_tokens=1 - lt_supported=true: the endpoint also returns logprobs, so LT-lite can run - recommended_detector: the script's direct recommendation for the next step

4. If lt_supported=false, do not force LT-lite; pivot to B3IT-lite or report that LT is unavailable.
Save baselines to explicit JSON files and reuse them for later detection.
Keep outputs file-based for coding CLIs and OpenClaw. Do not depend on GUI state.
For noisy endpoints, prefer the built-in B3IT defaults before tightening or loosening thresholds manually.

Endpoint Types

- OpenAI: use this for official OpenAI-style endpoints.
INLINECODE21: use this for third-party endpoints that follow OpenAI request and response shapes; vendor-specific headers may be required.
INLINECODE22: use this for /v1/messages style endpoints; in this skill it is B3IT-only.

Commands

Capability smoke

CODEBLOCK6

Generate a provider config template

CODEBLOCK7

Generate a batch providers.json template

CODEBLOCK8

Batch capability smoke

CODEBLOCK9

Batch LT-lite baselines

CODEBLOCK10

Batch LT-lite detect

CODEBLOCK11

Batch B3IT-lite baselines

CODEBLOCK12

Batch B3IT-lite detect

CODEBLOCK13

LT-lite baseline

CODEBLOCK14

LT-lite detect

CODEBLOCK15

B3IT-lite baseline

CODEBLOCK16

B3IT-lite detect

CODEBLOCK17

Daily single-endpoint drift run

CODEBLOCK18

Defaults and guardrails

- Default to non-streaming, timeout=60, and temperature values matched to the detector.
Every command can additionally write a human-readable report with --html-output.
OpenAI/OpenAI-compatible requests may include custom JSON headers, either from the config file or --headers-json.
The Kimi-specific {"User-Agent":"KimiCLI/2.0.0"} header is not a general default. Use it only for https://api.kimi.com/coding endpoints; for the OpenAI-compatible Kimi path, use https://api.kimi.com/coding/v1.
The script auto-disables thinking for common reasoning-first providers such as Ark, Doubao, GLM, and Zhipu unless extra_body is explicitly provided in the config JSON.
For OpenAI/OpenAI-compatible endpoints that still return reasoning_content without normal text, the script will retry once with {"thinking":{"type":"disabled"}} before failing.
OpenAI/OpenAI-compatible configs may use either the API root or a full .../chat/completions URL; the script normalizes the base URL internally.
INLINECODE36 writes the normalized config explicitly, including any auto-selected extra_body, so the saved file is portable across Codex, Claude Code, Gemini CLI, and OpenClaw runs.
INLINECODE38 does the same normalization for a whole config list and writes a ready-to-run providers.json.
Anthropic mode is treated as B3IT-only in this skill.
If the endpoint returns reasoning/thinking blocks but no normal text, report that as a capability mismatch instead of fabricating a token result.
If the endpoint does not return logprobs, report that LT-lite is unavailable instead of guessing.
B3IT baseline discovery on OpenAI/OpenAI-compatible endpoints will automatically retry with a stronger candidate-search profile if an initial low-cost run finds no border inputs.
B3IT detection defaults are tuned for lower false positives: detection-repeats=5, min-stable-count=2, min-stable-ratio=0.35, confirm-passes=1.
Daily runs auto-refresh older B3IT baselines when they predate the stable-support filtering fields.

Resources

Open only what you need:

- Workflow examples and CLI usage: INLINECODE45
Protocol-first playbook for OpenAI, OpenAI-Compatible, and Anthropic: INLINECODE49
Kimi coding endpoint quickstart: INLINECODE50
Kimi Anthropic endpoint quickstart: INLINECODE51
JSON config format: INLINECODE52
Example provider list: INLINECODE53
Main executable: INLINECODE54
One-shot batch wrapper: INLINECODE55
Daily single-endpoint wrapper: INLINECODE56

API 质量检查

使用捆绑脚本运行无头API质量检查。将此技能视为脚本优先：除非脚本明显不足，否则不要内联重新创建LT-lite/B3IT-lite逻辑。

供应商名称如Ark/火山引擎、GLM、DeepSeek、Kimi、SiliconFlow及类似服务仅为示例。主要决策在于端点协议类型：OpenAI、OpenAI兼容或Anthropic。

快速开始

一次性设置路径：

bash
export CODEXHOME=${CODEXHOME:-$HOME/.codex}
export APIQ=$CODEXHOME/skills/api-quality-check/scripts/apiquality_check.py
export APIQBATCH=$CODEXHOME/skills/api-quality-check/scripts/runbatchchecks.sh
export APIQDAILY=$CODEXHOME/skills/api-quality-check/scripts/rundailycheck.sh

首先运行能力冒烟测试：

bash
python $APIQ smoke \
--provider OpenAI兼容 \
--base-url https://ark.cn-beijing.volces.com/api/coding/v3 \
--api-key $API_KEY \
--model-id ark-code-latest \
--html-output ./smoke.html

对于许多OpenAI兼容的端点，如果用户粘贴完整的.../chat/completions URL，同样的命令也适用。脚本会自动将其规范化为API根路径。

如果您想要一个立即可用的provider.json，请使用以下命令生成：

bash
python $APIQ init-config \
--provider OpenAI兼容 \
--base-url https://api.siliconflow.cn/v1/chat/completions \
--api-key $API_KEY \
--model-id deepseek-ai/DeepSeek-V3.2 \
--name siliconflow-v3-2 \
--config-output ./provider.json

如果端点需要特定客户端的标头，请将其作为headers对象放入配置JSON中，或使用--headers-json传递。对于Kimi编码端点，仅在地址位于https://api.kimi.com/coding下时使用{User-Agent:KimiCLI/2.0.0}；对于OpenAI兼容的Kimi路径，请使用https://api.kimi.com/coding/v1。

如果您已有多个原始端点条目，请使用以下命令将其规范化为providers.json：

bash
python $APIQ init-batch-config \
--configs ./raw-providers.json \
--config-output ./providers.json

或者运行完整的批处理流水线：

bash
$APIQ_BATCH ./providers.json ./api-quality-out

该命令还会创建./api-quality-out/index.html作为所有生成报告的登录页面。

对于需要每天检查并按日期归档的单个端点：

bash
bash $APIQ_DAILY ./provider.json ./daily-out my-endpoint

工作流程

1. 在任何基线或检测运行之前先运行smoke。
如果您有多个端点，在选择哪些端点值得进行更深入的LT/B3IT工作之前，先使用配置列表运行batch-smoke。
读取结果：

- b3itsupported=true：端点在maxtokens=1时可以返回正常的首个token文本 - lt_supported=true：端点还返回logprobs，因此可以运行LT-lite - recommended_detector：脚本直接推荐的下一步操作

4. 如果lt_supported=false，不要强制使用LT-lite；转而使用B3IT-lite或报告LT不可用。
将基线保存到显式的JSON文件中，并在后续检测中重复使用。
保持输出基于文件，适用于编码CLI和OpenClaw。不要依赖GUI状态。
对于噪声较大的端点，在手动收紧或放宽阈值之前，优先使用内置的B3IT默认值。

端点类型

- OpenAI：用于官方OpenAI风格的端点。
OpenAI兼容：用于遵循OpenAI请求和响应形状的第三方端点；可能需要特定供应商的标头。
Anthropic：用于/v1/messages风格的端点；在此技能中为仅B3IT。

命令

能力冒烟测试

bash
python $APIQ smoke --config ./provider.json --output ./smoke.json

生成供应商配置模板

生成批处理providers.json模板

bash
python $APIQ init-batch-config \
--configs ./raw-providers.json \
--config-output ./providers.json

批处理能力冒烟测试

bash
python $APIQ batch-smoke --configs ./providers.json --output ./batch-smoke.json --html-output ./batch-smoke.html

批处理LT-lite基线

bash
python $APIQ batch-lt-baseline \
--configs ./providers.json \
--output-dir ./lt-baselines \
--output ./batch-lt-baselines.json \
--html-output ./batch-lt-baselines.html

批处理LT-lite检测

bash
python $APIQ batch-lt-detect \
--configs ./providers.json \
--baseline-manifest ./batch-lt-baselines.json \
--output ./batch-lt-report.json \
--html-output ./batch-lt-report.html

批处理B3IT-lite基线

bash
python $APIQ batch-b3it-baseline \
--configs ./providers.json \
--output-dir ./b3it-baselines \
--output ./batch-b3it-baselines.json \
--html-output ./batch-b3it-baselines.html

批处理B3IT-lite检测

bash
python $APIQ batch-b3it-detect \
--configs ./providers.json \
--baseline-manifest ./batch-b3it-baselines.json \
--output ./batch-b3it-report.json \
--html-output ./batch-b3it-report.html \
--detection-repeats 5 \
--min-stable-count 2 \
--min-stable-ratio 0.35 \
--confirm-passes 1

LT-lite基线

bash
python $APIQ lt-baseline --config ./provider.json --output ./lt-baseline.json

LT-lite检测

bash
python $APIQ lt-detect \
--config ./provider.json \
--baseline ./lt-baseline.json \
--output ./lt-report.json

B3IT-lite基线

bash
python $APIQ b3it-baseline --config ./provider.json --output ./b3it-baseline.json

B3IT-lite检测

bash
python $APIQ b3it-detect \
--config ./provider.json \
--baseline ./b3it-baseline.json \
--output ./b3it-report.json \
--detection-repeats 5 \
--min-stable-count 2 \
--min-stable-ratio 0.35 \
--confirm-passes 1

每日单端点漂移运行

bash
bash $APIQ_DAILY ./provider.json ./daily-out my-endpoint

默认值和防护措施

- 默认为非流式、timeout=60，温度值与检测器匹配。
每个命令都可以额外使用--html-output写入人类可读的报告。
OpenAI/OpenAI兼容的请求可能包含自定义JSON headers，来自配置文件或--headers-json。
Kimi特定的{User-Agent:KimiCLI/2.0.0}标头不是通用默认值。仅用于https://api.kimi.com/coding端点；对于OpenAI兼容的Kimi路径，请使用https://api.kimi.com/coding/v1。
脚本会自动禁用常见推理优先供应商（如Ark、豆包、GLM和智谱）的思考功能，除非在配置JSON中显式提供了extrabody。
对于仍返回reasoningcontent但没有正常文本的OpenAI/OpenAI兼容端点，脚本将在失败前使用{thinking:{type:disabled}}重试一次。
OpenAI/OpenAI兼容的配置可以使用API根路径或完整的.../chat/completions URL；脚本会在内部规范化base URL。
init-config会显式写入规范化后的配置，包括任何自动选择的extra_body，因此保存的文件

api-quality-checkAPI质量检测