Bernstein — Multi-Agent Orchestration

Version updated for https://github.com/sipyourdrink-ltd/bernstein to version v3.7.0.

  • 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 orchestrates a crew of CLI coding agents to execute tasks in parallel against a single goal using deterministic scheduling. It provides an HMAC-signed audit chain with bearer-token authentication, signed agent cards, and per-artefact lineage recording.

What’s Changed

v3.7.0

Released 2026-07-17.

Theme

v3.7.0 opens the orchestrator beyond coding agents and turns its determinism and audit chain into standard, independently verifiable receipts. The through-line: any agent, any artifact, and every result provable offline.

Highlights

Verifiable receipts, in standard formats

  • Export the audit chain as standard verifiable receipts (#2617). Projects an HMAC audit-chain range into COSE_Sign1, in-toto attestation, and a transparency-log-style signed receipt, each re-verifiable offline by a standard verifier with zero Bernstein imports. Tampering with a chain entry breaks the receipt exactly as it breaks the chain.
  • Per-goal SLA contracts (#2618). Signed violation receipts, lineage-anchored freshness, and ledger-projected error budgets: a breach is a signed, offline-verifiable receipt, not just a log line.

Universalization: non-coding agents as first-class

  • Browser and computer-use adapter (#2621). A new adapter family fronts autonomous browser and computer-use agents. Every action is a content-addressed lineage anchor; the action sequence replays deterministically and a divergence surfaces as a hash mismatch at the exact action index.
  • Research activity worker (#2630). Non-coding research runs land as artifacts whose every citation is an offline-resolvable lineage anchor, so a report reconstructs from the chain alone.

Durable, resilient execution

  • Durable task suspend and resume (#2616). A task waiting on a human or external event parks with an attested receipt that releases the seat, sandbox, and budget envelope; resume reconstructs byte-identically from the receipt.
  • Cache policy engine (#2625). Composable content-addressed key recipes, repo-drift expiry, claim-based fleet dedup with a signed duplicate-of receipt, and lineage-propagated eviction. Two operators derive byte-identical cache keys and verdicts.
  • Lineage-attested failure receipts in recovery (#2628). An on_fail recovery task now carries a lineage-attested failure receipt, so why-it-failed and what-recovered are a single verifiable edge.

Fleet and deployment

  • Named sandbox pools (#2547). Chain-projected pool manifests, governed overrides with capability and egress ceilings, and signed worker enrolment. Two hosts holding the same manifest compute a byte-identical effective-manifest hash; a widening override is refused with a chained receipt and no sandbox is created.
  • Sovereign deployment profile (#2629). A signed residency posture attestation anchored in the audit chain; posture drift at spawn is a signed, offline-verifiable refusal rather than a silent misconfiguration.
  • Mission timeline and signed daily digests (#2510). A mission screen backed by signed daily progress digests, each a verifiable projection of the day’s chain.

Security

Hardening from coordinated disclosure (all fixed and merged):

  • Memory provenance on the spawned-agent prompt path (#2624). The cross-adapter memory-poisoning invariant is now enforced where the spawned-agent prompt is built, not only in the opt-in query path.
  • Sandbox scope enforcement (#2622). scope_enforcement is no longer auto-relaxed for a sandbox backend that cannot demonstrate a task-scoped mount.
  • Newsletter endpoint (bernstein.run). POST /api/notify now returns a uniform response regardless of prior subscription state, closing a subscriber-membership oracle.

The disclosure policy was also corrected (#2623, #2627, #2634): the prior SECURITY.md advertised a paid bug bounty with no budget behind it. It is replaced with an honest recognition-only coordinated-disclosure policy, and the sovereign profile now attests egress truthfully at spawn time.

Acknowledgments

Thanks to the researchers who reported the issues above through coordinated disclosure:

  • Dmitriy Filatov (Malder) (@Malder-coder) for the memory-provenance and sandbox scope-enforcement reports.
  • Gaurav Popalghat (https://www.linkedin.com/in/noobx/) for the newsletter subscriber-enumeration report.