Bernstein — Multi-Agent Orchestration

Version updated for https://github.com/sipyourdrink-ltd/bernstein to version v3.9.0.

  • 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 that automates CLI coding tasks using various adapters. It provides reproducible runs by scheduling them in plain Python and ensures checkable results through an always-on lineage spine and replay journal. The action supports multiple agents, including Claude Code, Codex, Gemini CLI, and 40+ more, with options for local and air-gap installations and auditing receipts offline.

What’s Changed

A feature release. This wave widens the verifiability surface: the run journal, lineage spine, and HMAC audit chain gain new first-class receipts across identity and delegation, outbound payments, non-coding artifacts, evaluation, and cross-agent interop, and a hardening pass runs through the event, admission, MCP-audit, cost, and process-reap surfaces. No existing public API changes; upgrade in place, but read the behaviour-change note below first. The exhaustive index is docs/reference/FEATURE_MATRIX.md.

Behaviour changes (read before upgrading)

  • A halted mission phase no longer halts the whole mission, and is no longer reported as the active phase (#2683). Phases already run under isolated budget envelopes at dispatch, so exhausting one phase’s envelope never gated a sibling’s spend. The status projection had not caught up: it reported a halted phase as the active phase and folded any halt into an overall halted mission, so one exhausted envelope masked a runnable sibling. The projection now skips halted (and passed) phases when selecting the active phase, and reports the mission halted only when no phase remains runnable. A halted phase beside a runnable one now leaves the mission active with the runnable phase reported active.

    The active phase and the overall mission state are both hashed into mission_status_hash, so this correction deliberately moves that hash for any ledger carrying a halt, and for every ledger it moves the schema-version field alone. MISSION_STATUS_SCHEMA_VERSION is therefore bumped from 1 to 2. That field is how a verifier tells “folded under newer projection rules” apart from “tampered with”: an archived ledger stays reproducible because project_mission (and project_mission_from_ledger) accept an explicit schema_version, and projecting an archived ledger under schema_version=1 reproduces its original status hash byte for byte. The pinned v1 goldens assert exactly that reproduction.

    Halt semantics are unchanged and stay chain-anchored: a halted phase still seals its halt receipt onto the work ledger and mirrors it onto the HMAC audit chain. Isolation means a sibling phase continues, not that a halt is weakened. The mission daily-progress digest embeds mission_status_hash, so its digest_hash and receipt_id move with the projection even though the digest document schema itself is unchanged.

Identity and delegation

  • Attenuated delegation capability tokens: each hop is a detached-JWS token over a JCS-canonical body whose caveats can only narrow (permissions, task-id allowlist, path prefixes, expiry, uses, depth), verifiable offline from the token bytes with the registry unavailable, and minted as a chain-anchored event (#2901).
  • The install-identity signature on outbound agent-facing requests now binds the request body (an RFC 9530 Content-Digest folded into the RFC 9421 covered set), and the per-hop delegation ledger serialises its read-modify-write under a cross-process lock so concurrent writers can no longer fork the chain (#2935).
  • A2A server surface: a signed capability card served at both well-known paths, message/send + tasks/get over JSON-RPC with SSE, auth declared in the card, and every inbound task returning a chain-anchored lineage receipt an external caller can verify offline; clients without streaming or artifacts still work through polling (#2911).
  • AGNTCY ADS publication: bernstein a2a publish --surface agntcy-ads emits an OASF capability descriptor that is a deterministic projection of the signed card, carrying a provenance attestation signed by a trust root distinct from the card key; republishing an unchanged node is byte-identical (#2943).

Payments and cost governance

  • Signed payment mandates with chain-anchored receipts: a SpendMandate is an Ed25519-signed, content-addressed authorization bounding amount, currency, recipient, and expiry in a human_present or delegated presence mode; every attempt is a TransactionReceipt (a refusal is a first-class receipt with a closed reason), amounts are string-encoded integer nano-units with no float in a signed payload, and cumulative spend is aggregated under a lock so concurrent workers sharing a mandate cannot exceed the cap (#2902). The mandate and consent-receipt surfaces were then hardened (#2936).
  • x402 settlement hook: metered MCP gateway calls settle through the mandate framework, binding the signed intent, the authorized calls, and the settlement reference into one chain-anchored consent receipt (#2937).

Non-coding artifacts and evaluation

  • Typed artifact contract (slice 1): an ArtifactKind / ArtifactSpec threaded onto Task (defaulting to code_diff), widened lineage artefact kinds, per-kind canonical serialisers over one shared core, three new completion signals (schema_valid / criteria_match / hash_stable), and bernstein artifact verify — so a report, dataset, or action log is a signed, content-addressed lineage record rather than a blob (#2903).
  • Figure grounding for report artifacts: every material number in a report must trace to an anchored source (attachment, artifact, or query receipt) or the task does not complete; the figures sidecar is inside the artifact’s own content hash, and bernstein artifact verify renders per-figure provenance (#2912).
  • Content-addressed query receipts: a read-only SQL result set an agent consumed is canonicalised and its hash bound into a signed lineage entry; bernstein datasource verify proves it offline and --re-execute reports MATCH or DRIFT, and DML/DDL is refused (#2909).
  • Deterministic replay-debugging surface: bernstein replay debug walks the Merkle step chain to name the exact diverging step and field, emits a byte-identical two-run path-diff, forks from a step, and produces an offline-verifiable debug receipt (#2900).

Cluster and lifecycle

  • Leaderless MESH claim journal (phases 1-2): a signed, append-only, hash-chained claim journal with a pure project_claims fold that two nodes reduce to byte-identical state, and a deterministic lowest-entry-hash supersede rule; the SQLite ledger becomes a projection of the journal on the opt-in path while STAR behaviour is unchanged (#2907).
  • Mission phase isolation is described in the behaviour-change note above (#2683).

Integrations

  • Generic OData v4 system-of-record integration: a watermark/delta trigger source with a deterministic, chain-anchored poll cursor, and receipt-anchored write-backs (GET-before-PATCH with If-Match, typed 412/428 conflicts, draft-activate) where every write-back is a signed audit event verifiable by bernstein audit verify (#2908).
  • Self-hosted OpenAI-compatible endpoint certification: bernstein endpoints certify --base-url produces the existing signed certification record for an arbitrary endpoint, with a docs page covering the certify → verify loop and the exercised endpoint families (#2913).

Hardening and security

  • Bounded hardening passes through the event feed, triggers, webhooks, and receipts (#2890); the pool/lease admission engine, ledger, tags, and verifier (#2891); MCP stateless/gateway/client audit ordering (#2892); provider-state mutation capture (#2893); cost-scheduling receipts and the price table (#2894); and process reap, platform compat, and worktree isolation (#2895).
  • bernstein audit verify is now total over an undecodable chain segment: a non-UTF-8 segment is reported as a named verification failure at the exact byte offset instead of raising out of verify() (#2880).
  • bernstein lineage verify fails closed with a distinct exit code when the audit key is missing, instead of misreporting a setup problem as tamper (#2882).
  • The SLA-receipt log sink escapes caller-supplied paths so a CR/LF path can no longer forge log lines (#2881); adapter capability-profile selection is recorded at spawn so replay detects drift (#2883).
  • The CI weekly-digest and skills-publish workflows were hardened, along with skills packaging (#2945, #2946).

Observability

  • The OpenTelemetry GenAI-semconv gaps in the signed OTLP span projection were closed (#2896).

CLI and documentation

  • Operator-surface follow-ups from the 3.8.2 sweep: an honest dashboard capacity denominator, stop --force process-group reaping, a live-scan ps cross-check, and a bernstein cluster status node view (#2897).
  • The README now front-loads the four differentiators as a scannable at-a-glance ahead of the full capability list (#2938); the feature matrix was refreshed for currency across the verifiability, identity, payments, and provenance surfaces shipped since the last pass (#2941); the agent-doc mirrors gained currency guards so the generated views cannot silently drift (#2906).
  • A decision record measures the cost of enabling the HMAC audit chain by default and defines the migration path, without flipping the default (#2905).

Thanks

Thanks to @Maqbool61 for documenting and helping certify the self-hosted OpenAI-compatible endpoint path (#2913).

pip install --upgrade bernstein==3.9.0