Bernstein — Multi-Agent Orchestration

Version updated for https://github.com/sipyourdrink-ltd/bernstein to version v3.16.0.

  • 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 an open-source tool designed to orchestrate deterministic multi-agent CLI operations. It automates the execution of various AI language models like Claude Code, Codex, and Gemini CLI in parallel, ensuring consistent and predictable results by gating output and recording detailed logs. The action supports a wide range of models and provides clear documentation and support for integration into development workflows.

What’s Changed

Someone pulled the plug mid-write. The journal ended on half a line, and the task that owned it was gone. That used to be the end of the story. Here it is a repair.

Replay stopped losing the ending

  • A crash-torn journal tail is repaired and the suspended task resumes. (#3955, @Louis20060723)
  • Journal consistency and sealed identity are separate verdicts, so a clean prefix stops passing for a whole journal. (#3908, @Silentpartnercoding)
  • event_count counts usable events through the sealed prefix instead of trusting the file. (#4025, @sujeito-operator)
  • A journal line that is not valid UTF-8 is discarded, not raised. (#4008, @sujeito-operator)
  • replay --as-json emits JSON on every exit path, including the ones that fail. (#4017, @Chirag6722; #3991)

Proof instead of assurance

  • A result bundle verifies offline against DSSE. No network, no registry, nobody to ask. (#3903, @Nickgonzales76017)
  • allowed_files is enforced at the merge acceptance gate, not merely declared. (#4002, @Aeirx)
  • A merge whose incoming change cannot be verified is refused rather than waved through. (#4022)
  • The credential token_type allowlist derives from its own Literal and cannot drift from it. (#4030, @Aeirx)

Volunteer workers

Bernstein opted into its own volunteer program, which is either good engineering or a conflict of interest. The project manifest is content-addressed, the sandbox profile is derived rather than declared, and bernstein volunteer verify says which. (#3907, #3909, #3912, #3916, #3920)

Boundaries

  • Sixteen hand-rolled path checks became one containment helper. (#3855, #3857, #3858, @vaibhav8a)
  • Card matches are Luhn-validated before redaction: real numbers go, coincidences stay readable. (#3818, @vaibhav8a)
  • Budget forecasts are scoped to the caller’s tenant. (#3850, @vaibhav8a)
  • Upgrade proposals are gated on the proposer’s measured history. Confidence is not history. (#3840, #3843, @Nickgonzales76017)

The CI said it was fine

It was not. A gate that always succeeded, a mutation lane that could not fail, a workflow whose header described another workflow. Now they say what they do.

  • Silently-inert gates and misleading green, across the workflow surface. (#3945, #3960)
  • Workflow headers and names match behaviour. (#3956, @ThinkerDesigns)
  • Merges land through a queue, and release bumps go through it too. (#3964)

Also

  • READMEs in Hindi and Bengali. (#3992)
  • First-run paths for evolve and listen are fenced. (#3848, @jvsilva12600009)
  • A dashboard vocabulary reference. (#3790, @atirna)
  • The TUI renders a toast again, and its task list stops lying about stale rows. (#3938; #3932, @ShlokShar)

Upgrading

Upgrade in place. JournalVerifyResult reports chain consistency, reader coverage and sealed identity as three verdicts instead of one ok bit; anything reading the old bit needs updating.

Contributors

@Chirag6722 · @sujeito-operator · @vaibhav8a · @Louis20060723 · @Nickgonzales76017 · @Aeirx · @ThinkerDesigns · @Silentpartnercoding · @ShlokShar · @atirna · @jvsilva12600009 — and Renovate and Dependabot, who never sleep.

Soundtrack

Two footwork takes on the same lyric sheet.

Lyrics
[Intro — dry vocal, chopped]
Version three... one... six
Tag it — tag it — tag it
(queue empty, main green)

[Drop — chop loop]
Drain the queue, drain the queue
Merge — merge — merge —
Green on main, green on main
Fake green? Found it. Gated.

[Break — replay]
Torn tail? Stitched.
Crash mid-write? Resume.
Replay it byte for byte
Same hash — same hash — every time

[Drop — receipts]
Show me the receipt (-ceipt -ceipt -ceipt)
Verify it offline
No trust, just proof
No trust — just proof

[Break — volunteer]
Lend your worker to the cause
Sandbox locked, scope declared
Zero creds, clean room
We run our own program

[Bridge — tenant]
Your data — your lane
Tenant-scoped, stay in your lane
Card match? Redacted.
No leak — no leak

[Outro]
Readme speaks four new tongues
Mutation gate bites now
The ratchet only turns one way
Three... one... six. (tag it)

What’s Changed

New Contributors

Full Changelog: https://github.com/sipyourdrink-ltd/bernstein/compare/v3.15.1...v3.16.0