v2.11.0¶

Released 2026-07-02.

Features¶

• The openai_agents runner accepts optional sampling and endpoint parameters: temperature, top_p, top_k, base_url, and api_key_env on the runner manifest, flowing into the SDK client and model settings. Absent fields keep the previous behavior byte-identical. When base_url is set the runner switches to the chat-completions API and excludes the custom client from tracing, so a third-party key is never sent to the default tracing endpoint. Every effective parameter is logged in the runner start event, so runs stay self-describing. Design validated in daily runs by @shanemmattner (#2159). (#2173)
• api_key_env is fail-closed: it must name a known LLM provider key from the built-in allowlist; anything else requires the operator to allow it via BERNSTEIN_ALLOWED_API_KEY_ENVS on the host, which a repository cannot set. Requesting sampling parameters on an adapter without the new SUPPORTS_SAMPLING_PARAMS capability fails loudly instead of silently dropping them. (#2173)
• SDK runners now write heartbeats, so they are visible to the stall watchdog between spawn and exit. (#2173)

Fixes¶

• The Docker sandbox path from v2.10.0 is hardened: each spawned agent gets its own sandbox session (one exec timeout no longer tears down every agent's container), committed work is bundled out of the container and fetched into the host repo under sandbox/<session_id> refs, sandbox lifecycle events land in the HMAC-chained audit log with emissions serialized so concurrent lifecycles cannot fork the chain, and provisioning probes task-server reachability and warns on daemons without host networking. (#2162, #2172)
• The bernstein worker loop can spawn agents again: it constructed the spawner with arguments that never existed and raised TypeError on the first claimed task. The server URL now also reaches spawned agents through the environment allowlist. (#2163, #2171)

Dependencies¶

• Routine CI action digest updates (github/codeql-action, docker/setup-buildx-action).