change8.27

[sinonkey]

Summary

Type of change

• Bug fix
• New feature / capability
• Refactor (no behavior change)
• Documentation only
• Benchmark / eval change
• Trace schema or replay change
• Tooling / CI

Implementation notes

Verification

Run before requesting review:

python -m compileall dm_agent main.py tests
python -m pytest
python -m dm_agent.evals.cli --variant full --task direct_finish
python -m dm_agent.benchmarks.cli --suite maintenance --list
python -m ruff check .
python -m black --check .

If this change touches benchmarks or live-model behavior:

• Attached relevant bench_reports/*.md or eval_reports/*.md.
• Documented the model and provider used.
• Documented the random seed / sampling parameters.

Safety checklist

• No API keys, .env, config.json, or mcp_config.json in the diff.
• No new default-on dangerous tool replay or shell execution paths.
• Trace defaults still keep full LLM I/O capture opt-in.

Related issues / research log

[hwfengcs]

Make substantial changes