model-tester（模型测试员）

Use scripts/model_tester.py to run repeatable test prompts and compare requested vs actual model usage from OpenClaw logs.

Run

From the skill directory (or pass absolute paths):

python3 scripts/model_tester.py --agent menial --case extract-emails
python3 scripts/model_tester.py --model openai/gpt-4.1 --case math-reasoning
python3 scripts/model_tester.py --agent chat --model openai/gpt-4.1 --case all --out /tmp/model-test.json

Inputs

• --agent <name>: Target agent (chat, menial, coder, etc.)
• --model <name>: Requested model alias/name to test
• --case <id|all>: Case from references/test-cases.json or all
• --timeout <sec>: Per-case timeout (default 120)
• --out <file>: Optional JSON output file

Require at least one of --agent or --model.

What the runner does

1. Load test cases from references/test-cases.json.
2. Start openclaw logs --follow --json in parallel.
3. Run openclaw agent --json with a bounded test prompt (asks agent to use a subagent for the task).
4. Parse response + tailed logs.
5. Emit machine-readable JSON and a short human summary.

Output format

Top-level JSON:

• tool
• timestamp
• agent
• requested_model
• results[]

Each result entry returns:

• test_case
• agent
• requested_model
• actual_model (parsed from logs when available)
• status (ok/error)
• result_summary
• runtime_seconds
• tokens (when discoverable)
• errors[]

Privacy & Safety

The tester spawns isolated subagent tasks with predefined test prompts only — no user data is passed to models. It tails OpenClaw logs to extract:

• which model was actually selected (routing validation)
• token usage statistics
• runtime metrics

Log extraction uses regex patterns to find model/token fields. No personally identifiable information or arbitrary log content is captured — only structured fields related to the test execution.

Notes

• Model extraction and token extraction are best-effort because log fields may vary by OpenClaw/provider version.
• If openclaw config is invalid or gateway is unavailable, the script returns status=error with stderr details.
• Edit references/test-cases.json to add custom prompts for your benchmark set.
• All test cases are generic; no workspace or user data is baked in.