Version updated for https://github.com/sipyourdrink-ltd/bernstein to version v2.14.0. This action is used across all versions by 5 repositories. Action Type This is a Composite action. Go to the GitHub Marketplace to find the latest changes. What’s Changed v2.14.0 Released 2026-07-04. Multi-provider orchestration hardening. A large batch of correctness and run-safety fixes for real-codebase runs across mixed providers (Claude, OpenAI, OpenRouter, DeepSeek, MiniMax, Qwen, Gemini), contributed by @shanemmattner and finished to green. Fixes Stalled-manager watchdog no longer misdiagnoses a healthy manager as an auth failure: a manager doing real root-cause investigation before it POSTs its first child task is no longer killed at a hardcoded deadline, the config override is honored, and the failure record names the real cause. (#2179) Failure classifier stops false-positive-killing healthy workers: bare-substring patterns (413, 429, 401, max_tokens, context window) no longer match structured tool-call log data; detection is anchored to real error shapes. (#2183) Janitor acceptance checks tolerate idiomatic worker paths: path_exists honors explicit globs and an opt-in fuzzy basename fallback, so a run where workers placed correct output at repo-idiomatic paths is not cascaded to a false sev1. (#2186) Injected .claude/skills/bernstein-*.md files are excluded from work-branch commits, so they stop causing a merge conflict on every worker merge. (#2187) Per-call token usage is priced and surfaced on the openai_agents provider path, so budget guards are no longer inert on non-Claude runs. Model pricing matches the most specific key first, so mini and flash variants price at their own rate instead of the parent model rate. strict_json_schema is disabled for non-OpenAI models that reject it, with diagnostic logging. Plus the rest of the 22-fix batch across adapters, tasks, cost, routing, quality, and observability. Features Tunable agent run-length limits: max_turns, an error-budget floor, and max_agent_runtime_s, configurable per run. Internal Pricing table extracted into a dependency-free cost.model_prices leaf so adapters can price a call without reaching scheduler internals; public create_pr API preserved; a diagnostic pre-call log that dumped request headers/body verbatim is now redacted.

Version updated for https://github.com/sipyourdrink-ltd/bernstein to version v2.13.0. This action is used across all versions by 5 repositories. Action Type This is a Composite action. Go to the GitHub Marketplace to find the latest changes. What’s Changed 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) Features (per-role model configuration) 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) Quality Resolved refurb FURB123 findings in the OWASP control-map builders.

Version updated for https://github.com/sipyourdrink-ltd/bernstein to version v2.11.0. This action is used across all versions by 5 repositories. Action Type This is a Composite action. Go to the GitHub Marketplace to find the latest changes. What’s Changed 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).