Bernstein — Multi-Agent Orchestration

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.