v2.13.0¶

Released 2026-07-02.

Run-safety guardrails and per-role endpoint configuration.

Fixes (run safety)¶

GitHub backlog auto-sync is now opt-in and off by default. Previously a run in a repository with open GitHub issues would pull every open issue into the backlog before task scoping, which could silently discard a seeded goal and spawn work against the entire issue list. Enable it explicitly with the github.sync_backlog seed config key (or the BERNSTEIN_SYNC_GITHUB_BACKLOG env override). (#2178)
A seeded goal is no longer silently dropped when the backlog is non-empty: the run now prints a loud warning naming the precedence and how to force the goal, instead of quietly planning from the backlog. (#2178)
Agent worktree merges refuse to land on the repository default branch. The merge and push path resolves the protected default (origin/HEAD, then init.defaultBranch, then the conventional names, treating both main and master as protected when the remote head is ambiguous) and refuses to merge or push agent work onto it, recording the refusal, so a run started from a default-branch checkout can no longer push unreviewed commits straight to the trunk. (#2178)

role_model_policy entries gain optional base_url and api_key_env next to model/provider, so different roles can target different OpenAI-compatible endpoints in one workflow (for example a fast manager endpoint and cheaper worker endpoints). api_key_env names an environment variable and is validated against the same fail-closed provider allowlist as the runner. YAML anchors give reuse across roles with no new file format. Absent fields keep today's behavior. (#2159)
ModeProfile gains top_p, top_k, and max_tokens beside its existing temperature, and the previously-deferred apply_mode_to_spawn wiring is completed so a mode profile's sampling parameters actually reach the spawn and the runner manifest. (#2159)
Opt-in builtin tools for the openai_agents runner, for runs without an MCP gateway, selected by tool_source: builtin (the gateway remains the default). read_file, write_file, and list_dir are workdir-confined (absolute and parent-escape paths are rejected). run_command is a restricted process-exec primitive: bare-name commands only, shell interpreters blocked, resolved against PATH, available only under a configured OS sandbox provider or an explicit opt-in; its filesystem confinement is the OS sandbox, not the builtin. Every builtin call is recorded to the run event log so a gateway-free run stays auditable. (#2159)