Skip to content

chernistry/bernstein

v3.3.0 Feature

This release adds 3 notable features for engineering teams evaluating rollout.

✓ No known CVEs patched
Read the diff → Tool health → What is this tool? →

✓ No known CVEs patched in this version

Topics

agent-orchestrator agentic-ai ai-agents aider air-gap audit-trail
+14 more
claude-code cli-tool codex-cli coding-agent deterministic-replay deterministic-scheduler hmac-audit mcp-server model-context-protocol multi-agent parallel-worktrees provenance python reproducibility

Affected surfaces

auth breaking_upgrade

Summary

AI summary

Dashboard gains scoped token authentication with governance receipts, agy adapter adds nightly conformance canary, run review board projects runs into a web view.

Changes in this release

Feature Medium

Dashboard now supports scoped tokens with read‑only or operator scope.

Dashboard now supports scoped tokens with read‑only or operator scope.

Source: llm_adapter@2026-07-15

Confidence: high

Feature Medium

CLI command `bernstein dashboard-token` issues, lists, and revokes scoped tokens; revocation is a signed journal row.

CLI command `bernstein dashboard-token` issues, lists, and revokes scoped tokens; revocation is a signed journal row.

Source: llm_adapter@2026-07-15

Confidence: high

Feature Medium

agy adapter added to the matrix with golden transcripts and contract fixture.

agy adapter added to the matrix with golden transcripts and contract fixture.

Source: llm_adapter@2026-07-15

Confidence: high

Feature Medium

Nightly conformance canary probes adapters and stores sealed receipts in the audit chain.

Nightly conformance canary probes adapters and stores sealed receipts in the audit chain.

Source: llm_adapter@2026-07-15

Confidence: high

Feature Medium

Run review board provides deterministic web projection of run journal with sealed evidence bundles.

Run review board provides deterministic web projection of run journal with sealed evidence bundles.

Source: llm_adapter@2026-07-15

Confidence: high

Feature Medium

Windows platform layer gains process‑cleanup receipts, Job Object grouping, and junction‑aware worktree isolation groundwork.

Windows platform layer gains process‑cleanup receipts, Job Object grouping, and junction‑aware worktree isolation groundwork.

Source: llm_adapter@2026-07-15

Confidence: high

Feature Medium

`bernstein skills package` builds agent skill/plugin bundles with versioned manifest and lineage receipt on install.

`bernstein skills package` builds agent skill/plugin bundles with versioned manifest and lineage receipt on install.

Source: llm_adapter@2026-07-15

Confidence: high

Feature Medium

Every dashboard authentication decision records a signed governance receipt in the journal.

Every dashboard authentication decision records a signed governance receipt in the journal.

Source: llm_adapter@2026-07-15

Confidence: low

Dependency Low

Dependency updates performed.

Dependency updates performed.

Source: llm_adapter@2026-07-15

Confidence: high

Bugfix Low

All open scanner alerts resolved: applied idiom cleanups, moved dashboard password comparison to a costly key derivation, and marked two false‑positive log lines.

All open scanner alerts resolved: applied idiom cleanups, moved dashboard password comparison to a costly key derivation, and marked two false‑positive log lines.

Source: llm_adapter@2026-07-15

Confidence: low

Full changelog

v3.3.0

Released 2026-07-11.

An operator-capability release. The dashboard gains authentication with
scoped tokens whose every decision lands a governance receipt, a new agy
adapter joins the matrix with a nightly conformance canary sealing its
probe results, and the run review board projects the run journal into a
web view backed by sealed evidence bundles. Groundwork lands for Windows
parity and for installing bernstein as an agent skill or plugin.

Dashboard authentication and scoped tokens (#2366)

The dashboard is no longer an all-or-nothing surface. Operators issue
scoped credentials, and every authentication decision is a receipt:

  • Password login and scoped bearer tokens coexist; tokens carry either a
    read-only or an operator scope, and route access is enforced per scope.
  • Token grants and revocations are signed journal rows; a tampered grant
    fails signature verification and is treated as no grant at all.
  • Every allow or deny decision is recorded as a governance journal
    receipt, so an access review can reconstruct exactly who reached which
    surface and under which credential.
  • bernstein dashboard-token issues, lists, and revokes tokens from the
    CLI; revocation is itself a signed journal row, never a silent delete.

agy adapter and adapter conformance canary (#2368)

agy joins the adapter matrix, and adapter contracts are now exercised
nightly instead of waiting for a user's broken run:

  • The agy adapter maps goals onto the agy CLI with the same contract
    surface as every other adapter, including golden transcripts and a
    contract fixture.
  • A nightly canary probes every primary adapter against whatever upstream
    version is installed; each probe result is a sealed, content-addressed
    receipt mirrored into the audit chain.
  • A last-green table names the newest upstream version each adapter
    passed with, and every row carries the hash of the receipt that
    attested it. bernstein doctor reads the same projection and advises
    when the installed version is newer than the last attested one.
  • Issue automation is threshold-gated and deduplicated: one upstream
    flake never opens an issue, and the same failure fingerprint never
    opens two.

Run review board core (#2365)

Finished runs can now be reviewed on a board instead of by reading raw
journals:

  • The board is a deterministic projection of the run journal: the same
    journal always produces the same board state, byte for byte, with no
    hand-maintained status anywhere.
  • A web view lists runs with their outcomes, decisions, and artefacts,
    consuming sealed evidence bundles so what the reviewer sees is what the
    journal attests.
  • Further review-board phases (annotations, sign-off flows) are tracked
    separately.

Windows parity groundwork (#2367)

The platform layer gains the pieces Windows needs, validated so far on
the existing test matrix:

  • Process cleanup emits reap receipts into the audit chain, so a killed
    worker tree is reconstructable after the fact on every platform.
  • Job Object process management groups worker processes on Windows the
    way process groups do on POSIX, closing the orphaned-child gap.
  • Worktree isolation is junction-aware, so per-task isolation survives
    filesystems where symlinks require elevation.
  • Full validation on Windows runners is tracked separately.

Agent skill and plugin packaging groundwork (#2369)

Installing bernstein as an agent skill or plugin gets its packaging
foundation:

  • bernstein skills package builds a skill/plugin bundle from the
    packaged templates, with a manifest generated from the package version
    so the bundle can never drift from the release.
  • Each install produces a lineage receipt, so an installed bundle can be
    traced back to the exact release artefact it came from.
  • Remaining distribution surfaces are tracked separately.

Housekeeping

  • Resolved all open scanner alerts: applied suggested idiom cleanups,
    moved dashboard password comparison onto a computationally expensive
    key derivation, and annotated two exception-name-only log lines as
    scanner false positives.
  • Dependency updates.

Weekly OSS security release digest.

The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.

No spam, unsubscribe anytime.

Share this release

Track chernistry/bernstein

Get notified when new releases ship.

Sign up free

About chernistry/bernstein

Deterministic multi-agent orchestrator for 18 CLI coding agents (Claude Code, Codex, Cursor, Aider, Gemini CLI, OpenAI Agents SDK, and more). MCP server mode (stdio + HTTP/SSE) exposes the orchestrator to any MCP client. Git worktree isolation per agent, HMAC-chained audit trail, cost-aware model routing via contextual bandit. ~11K monthly PyPI downloads, Apache 2.0.

All releases →

Related context

Earlier breaking changes

  • v3.7.1 `bernstein approve` and `bernstein reject` now enforce identifier regex `[A-Za-z0-9._-]{1,64}`.
  • v3.7.1 Tampered mission ledger reports as unverified rather than not-found.
  • v3.7.1 `mission define` now refuses phases without gate tasks.
  • v3.5.0 MCP client, transport, and gateway become stateless; calls carry content‑derived trace IDs in _meta.

Beta — feedback welcome: [email protected]