Fix Toolathlon Daytona direct runtime

[bingran-you]

Summary

• route all Daytona ACP sessions through PTY transport, including direct non-compose sandboxes
• snapshot task-local Toolathlon tokens before private groundtruth/evaluation paths are locked
• wrap google-cloud-mcp launches so Toolathlon wildcard allowlists like promo-assets-for-b* work with the MCP server exact-match checks

Validation

• uv run python -m pytest tests/test_source_adapters.py -q
• uv run python -m pytest tests/test_process.py::TestDaytonaPtyProcessSecretTransport -q
• uv run python -m pytest tests/test_acp.py -k 'daytona_dind_uses_pty_transport or daytona_direct_uses_pty_transport' -q\n- uv run ruff check src/benchflow/acp/runtime.py src/benchflow/sandbox/process.py src/benchflow/adapters/_toolathlon.py src/benchflow/adapters/_toolathlon_container.py tests/test_process.py tests/test_acp.py tests/test_source_adapters.py\n- uv run ruff format --check src/benchflow/acp/runtime.py src/benchflow/sandbox/process.py src/benchflow/adapters/_toolathlon.py src/benchflow/adapters/_toolathlon_container.py tests/test_process.py tests/test_acp.py tests/test_source_adapters.py\n- git diff --check\n- Live Daytona Toolathlon proof: ab-testing, run .local/toolathlon-daytona-stability/proof-ab-testing-google-prefix-tmpwrap-20260705T0925Z, score 1/1 (100.0%), errors 0, reward 1.0\n

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 2c836209db

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Add the guarded commit to this regression docstring

This new regression test for the direct-Daytona PTY fix does not name the PR or commit it guards; /workspace/benchflow/AGENTS.md requires that “Regression tests must name the PR/commit they guard.” Please include f96a1958112604a459c103d521d930467045d26f or the eventual PR number here and in the companion new regression docstring so future reviewers can preserve the test’s provenance.

Useful? React with 👍 / 👎.