Bernstein — Multi-Agent Orchestration

Version updated for https://github.com/sipyourdrink-ltd/bernstein to version v3.18.2.

  • 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 multi-agent CLI orchestration tool that automates and orchestrates complex workflows, ensuring consistency across different environments. It simplifies the management of microservices, pipelines, and tasks by providing a comprehensive framework for defining and executing multi-step processes efficiently.

What’s Changed

Sixty-four commits since the last patch, most of them the machine finding its own mistakes. That is the arrangement working, not failing.

Receipts that outlive what they describe

A finished run’s journal now maps onto a signed TRACE 0.2 Trust Record through the install Ed25519 identity, behind the optional bernstein[trace] extra (#4666). The record names its own signing key — the subject is a self-certifying did:key URI — states the runtime it was produced under, and binds a delegated execution to its parent by the hash of the parent’s own signed record; three vectors minted from real runs verify offline (#4692). Sealed audit segments carry content-addressed hash tiles beside them (#4629). A refused read-set emits a receipt that verifies with no network and no repository (#4650). Journal verification is bound to plan rendering, so an approval binds to the rendering the reviewer actually saw (#4655). Task context packs are byte-deterministic, content-addressed containers (#4646), and sequential evaluation statistics gain time-uniform bounds computed in exact arithmetic (#4651).

The swarm stops stepping on itself

Migration checkpoints record chunk completion, so a restarted migration no longer re-spawns finished work and the reduce step can run at all (#4541). A re-run reuses a chunk whose task is still in flight instead of spawning a second owner for the same files (#4624). A run whose planning failed closes instead of idling to its wall-clock timeout (#4529). Fair scheduling caps the age boost, so an old task can no longer starve a fresh urgent one (#4675). Drain stops re-admitting a dead agent from a previous run’s agents.json (#4479), and any unmerged agent branch is rescued to refs/rescue/<run-id>/ before a forced delete (#4677). A spawn supervisor that exhausted its respawns records the failure on park, where a pruned store used to lose it (#4637). Issue-intake runs have no coordinator, so they now share a claim vocabulary two of them can read off each other’s comments (#4521).

bernstein pr says what happened

A description built from a git command that failed used to report that the branch changed nothing. It now reports the failure, and provenance is computed only from a diff that exists (#4669). The anchoring step no longer exits non-zero after the pull request is already open (#4686), and the description is composed against the base the request merges onto (#4693).

Adapters, plugins, volunteers

The goose adapter passes through the model it was told to use — a role bound to a specific model had been running on whatever the CLI defaulted to — parses the stream, and reports usage like the other adapters (#3679). Aider writes a per-spawn analytics log (#3674). Declared third-party trigger and reporter entry points load at runtime, and a malformed entry is named in a warning rather than becoming a silent no-op (#4531). Volunteer runs gain a second independent reviewer (#4685) and issue comments filtered so the model receives what the reviewer read (#4652).

Housekeeping

Modules nothing reached are gone, with the documentation rows that advertised them (#4683, #4643). The team-hub page documented a loader that had been deleted as unreachable; it now documents the manifest parser that ships, and deleting a file the drift playbook names as a doc’s source of truth fails on the pull request that deletes it rather than on the push afterwards (#4696). Committed coverage shards and scratch files are out of the tree, ignored by shape rather than by name (#4695). Context-file staleness on a pull request is attributed to that request’s own commits (#4634), and CONTRIBUTING documents how milestones are targeted, which had been folklore (#4285). Three gates were failing for reasons no code change could fix: the lint gate called ruff outside uv run on a project that never activates its virtualenv (#4547); a required status check filtered its trigger on paths, so every pull request outside the filter waited forever on a run that never started (#4556); and a completion signal named its test by a path written when the task was planned rather than one that exists (#4554).

Contributors

  • Chirag Honnyal — branch rescue on drain (#4681), the priority age-boost cap (#4680), spawn-supervisor failure records (#4662), staleness attribution (#4664).
  • Amir Fathi — swarm checkpoint completion (#4618), the drain liveness check (#4621), the reaper patch-target fix (#4688).
  • Sujeito Operator — a producer for the parked-session store (#4636), in-flight chunk reuse (#4625).
  • Jesus Esquer — agent timeout extension reaching the spawned process (#4608), reaper caller coverage (#4613).
  • Abishek N — the deterministic TaskContextPack container (#4694), dead-module removal (#4696).
  • Atirna — plugin entry-point registries (#4614).

What’s Changed

Full Changelog: https://github.com/sipyourdrink-ltd/bernstein/compare/v3.18.1...v3.18.2