Bernstein — Multi-Agent Orchestration

Version updated for https://github.com/sipyourdrink-ltd/bernstein to version v3.17.0.

  • 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 multi-agent CLI orchestration tool designed to help users manage complex agent-based systems. It automates the process of orchestrating multiple agents, ensuring consistent and reliable execution across different environments and configurations.

What’s Changed

This release began as an arXiv preprint. The evaluation section kept turning into a bug report, so we shipped the bug report first.

The task graph dropped work

  • A failed or cancelled task now reaches dependents outside its subtree. (#4243, #4258)
  • A node whose every inbound edge skipped is materialised. (#4253)
  • Decomposition recurses under a bound; completion has a typed record. (#4179, #4185, @Chirag6722)

Verification says what it checked

  • audit verify is read-only; append, rewrite and identity are separate verdicts. (#4244, #4245)
  • A run’s read-path set is derived from its journal. (#4180, @Louis20060723)
  • Absence claims carry a coverage record; compaction binds a pre-compaction hash. (#4188, #4189)
  • verify_result_bundle refuses wrong-typed fields instead of raising. (#4057, #4072, @timsurrealedu; #4252)
  • A malformed chain link is a verdict, not an exception. (#4055, #4051, @sujeito-operator)
  • A resume that clears its authority check records that it did. (#3649, @Maqbool61)
  • Run attestation reaches the command line. (#4139, @Silentpartnercoding)

Agents that died quietly

  • A spawned-then-dead agent is diagnosed, not silently detached. (#4246, #4255)
  • A custom qwen base URL picks an auth type; agents stop dying at zero tokens. (#4265)
  • system_addendum now rides resume, container and sandbox spawns. (#4254)
  • Agent and skill catalogs are opt-in, lockfiled and tamper-checked. (#3973, #3975, @Chirag6722)
  • OpenCode declares its capability strategy. (#4099, @KrzysiekSko)

Volunteer workers

  • A claimed task ends in a diff or a structured refusal. (#4061)
  • Untrusted issue text is normalised before it becomes a prompt. (#4059)
  • Claims carry duplicate detection over a durable lease store. (#3873, #4036, @madhavkbinoy)
  • An unreadable manifest refuses; “absent” and “unstattable” differ. (#4058, #4066, @sujeito-operator)

The same commit, the same verdict

  • Hypothesis and Schemathesis smoke profiles are derandomised. (#4105, #4108, #4125, @vaibhav8a; #4116, @Chirag6722)
  • Memory-heavy test files run serially instead of OOMing. (#4121, @Phoenix1504e)
  • Bundle and typecheck-ts gates report on a queued ref and can be required. (#4069, #4081, @sujeito-operator)

Also

  • Task reads are tenant-scoped, and sub-router routes are scope-checked again. (#4161, #4163, #4023; #4171, @Phoenix1504e)
  • More path checks route through the containment helper. (#4095, #4101, @vaibhav8a; #4136, @Louis20060723)
  • eval list emits run paths unwrapped, so they pipe. (#4114, @Fezaru)
  • TUI pane columns scroll. (#4090, @sahilmathur254)
  • READMEs in 23 languages. (#4094, #4172)

Upgrading

Upgrade in place. PostgresTaskStore.list_tasks() now returns a bounded page when given no limit; callers reading whole tables must page through offset. (#4169, #4240)

Contributors

@Chirag6722 · @sujeito-operator · @vaibhav8a · @Phoenix1504e · @Louis20060723 · @madhavkbinoy · @timsurrealedu · @Silentpartnercoding · @KrzysiekSko · @Maqbool61 · @sahilmathur254 · @Fezaru — and Renovate and Dependabot, still awake.

What’s Changed

New Contributors

Full Changelog: https://github.com/sipyourdrink-ltd/bernstein/compare/v3.16.0...v3.17.0