Bernstein — Multi-Agent Orchestration

Version updated for https://github.com/sipyourdrink-ltd/bernstein to version v3.13.0.

  • 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.

Action Summary

Bernstein is a deterministic orchestrator for CLI coding agents that uses plain Python scheduling to ensure reproducibility. It automates the orchestration of tasks, ensuring checkable results after execution through lineage and replay capabilities. Each task runs in its own git worktree with isolated mutable state, providing an audit chain and receipts for verifying offline.

What’s Changed

A reliability release, and mostly one defect repeated across the release machinery: a step that could not run reported as a step that passed. The npm wrapper had been failing inside a green job since 2.3.0; Homebrew, the SBOMs and the container image never ran at all for a release cut by automation; and the drift detector meant to catch any of that closed its own tickets over registries it had failed to read. All four are fixed, the RPM channel is wired up, and the reconciler now refuses to return a verdict for a channel it could not probe.

Full details for every item: docs/release-notes/v3.13.0.md.

Behaviour changes

  • A delegation chain with no recorded scope is now unproven rather than valid (#3306). Grading is a separate three-state verdict on ChainResult; bernstein delegation verify exits 1 on fail, 3 on unproven, 0 on pass. A script that treated exit 0 as proof of narrowing now sees 3 where it was previously told nothing was wrong. No receipt schema changed.
  • A failed npm wrapper publish now fails its job (#3322). Nothing depends on publish-npm, so the failure surfaces without holding back PyPI or the GitHub Release.
  • reconcile-release fails when it could not read a channel (#3345), and its auto-close of drift issues requires every channel verified, not merely no drift observed.

Release machinery

  • Releases created by automation now dispatch their downstream consumers explicitly — container image, Homebrew tap, SBOM attachment — instead of relying on an event that a GITHUB_TOKEN-created release never raises (#3323, #3345).
  • The SBOM workflow distinguishes a missing release from an API failure; only the former is skippable (#3345).
  • RPM builds publish to Copr from the same tag trigger as every other channel, and a submission failure fails the job (#3325).
  • The coverage-ratchet guard no longer parses an API error body as a baseline after the ratchet branch is auto-deleted (#3347).

Merge gate

  • Every pull-request head now receives a review-bot-ack verdict: the publisher election is deterministic, a cancelled gate run is re-dispatched once instead of standing down silently, and the publisher’s superseded runs no longer pile up in the Actions history (#3313, #3331).

Command-line defects

Six defects found by reading the code rather than waiting for reports:

  • doctor --json lost its entire payload when one tool probe timed out (#3330); undo audit logging resolved through a module alias mypy could not see (#3308); the worker heartbeat busy-looped on re-registration failure (#3309); agents match crashed on any catalog hit (#3310); prompts over 8191 characters could not spawn through a Windows .cmd shim (#3311); and stop --force left the watchdog process alive (#3312).
  • Three typed-interface mismatches in the task lifecycle produced dead-letter paths that lost incident data (#3315, #3316, #3317).

Documentation and changelog surfaces

  • The changelog surfaces stop contradicting each other: CHANGELOG.md and docs/CHANGELOG.md are pointer documents, stranded entries moved to docs/release-notes/unreleased.md, and bernstein changelog serves the page for the installed version (#3343).
  • The release flow is documented as it works; configuration files nothing read are removed (#3340, #3344). Deleted command surfaces are swept out of the docs, and two claims the code could not back are corrected (#3342).

Contributors

  • @aeoess designed and implemented delegation-chain grading (#3306): the closed verifier reason set, fail-dominant composition over the existing narrowing algebra, the scope_ref cross-check, and the third state for a chain that records no scope.