sipyourdrink-ltd/bernstein

Audit-grade multi-agent orchestration for CLI coding agents (Claude Code, Codex, Gemini CLI, +40 more). HMAC-chained audit log, signed agent cards, per-artefact lineage, air-gap deploy. The orchestrator your compliance team will sign off on. https://bernstein.run

Python541 ★AI Agent FrameworksGitHub ↗
Quality: excellent 22/24
PAI: integrate 0.75
Appraised: 2026-06-01 outdated

Perfect health and docs; Python loses typed-language point, dep manifest unconfirmed from provided data.

HMAC-chained audit log, Ed25519 signed agent cards, 44 CLI adapters — compliance-grade coding agent orchestration.

Per-artefact lineage traces every file write to producer, prompt SHA, model, and cost.

Contents

Overview

Verdict

Rating Summary
Quality excellent (22/24) Perfect health and docs with 44 CLI adapters, HMAC audit chain, and active v2.7.0 release; docked for Python (not typed) and unconfirmed dep count.
PAI Relevance integrate (0.75) Novel compliance orchestration patterns directly applicable to PAI's Delegation subsystem; subprocess-wrappable as a PAI skill despite Python runtime.

Quality Assessment

22/24 — actively-maintained / well-documented / solid

Health: 8/8 (actively-maintained)

Failed:

Passed:

Documentation: 8/8 (well-documented)

Failed:

Passed:

Engineering Signals: 6/8 (solid)

Failed:

Passed:

PAI Relevance

Dimension Score Assessment
Harvest Value 2 The HMAC-SHA256 audit chain per RFC 2104, Ed25519-signed agent cards via JWS/JCS canonicalization, and per-artefact lineage model are novel architectural patterns directly applicable to PAI's Delegation and Agents orchestration subsystem, neither of which currently implements tamper-evident audit trails or per-output provenance.
Integration Readiness 1 Bernstein is Python, not TypeScript/Bun, so it cannot be bun add-ed; however it ships a clean CLI (bernstein run, bernstein lineage verify) installable via pipx, making subprocess invocation from a PAI skill straightforward with moderate adapter code.
Overlap Risk 1 Partial overlap with PAI's Delegation skill (parallel work routing) and Agents skill (agent composition); Bernstein's coding-agent-specific scope — git worktree isolation, merge queue serialization, and compliance audit chain — extends well beyond what those skills cover.
Gap Fill 2 PAI has no capability for compliance-grade multi-agent coding orchestration: no HMAC audit chain, no signed agent cards, no per-artefact lineage, no parallel git worktree management — all clear gaps in the current Capability Manifest.

Composite: 0.75

What Next

Landscape Position

Category: AI Agent Frameworks

In this category: lobehub--lobehub, VoltAgent--voltagent, bmad-code-org--BMAD-METHOD, tkellogg--open-strix, zeenie-ai--MachinaOS, gastownhall--gastown, gordonbrander--busytown, humanlayer--12-factor-agents, NorthwoodsSentinel--meridian-protocol

Standing: Bernstein is the first compliance-grade coding agent orchestrator appraised in this category — uniquely positioned by its HMAC audit chain and signed agent cards versus the general-purpose agent frameworks already on file; nearest structural peers are gastownhall--gastown (multi-agent workspace) and VoltAgent--voltagent (agent engineering platform), but neither ships tamper-evident audit infrastructure or coding-agent-specific worktree isolation.

Evidence Base

Density: 9/10 — Available: full README (8KB, substantive), repo metadata (stars, forks, issues, license, dates, topics), latest release tag, CI badge signals, language, archived status, PyPI publication evidence. Missing: dependency manifest (not retrieved), test file listing, contributor count, downstream dependents.

Notes

Bernstein is solo-maintained (chernistry) and explicitly says so — the "no vendor SLA, GitHub issues only" caveat is real risk for production adoption. The repo is 71 days old at appraisal with 541 stars and v2.7.0 already released, indicating unusually fast iteration velocity. The air-gap deploy profile (wheelhouse path documented, no SaaS hop, file-based .sdd/ state) makes it credible for enterprise-sensitive environments where other orchestrators route through a cloud data plane. The deterministic scheduler's hard wall against emergent agent collaboration is an intentional design choice, not a limitation — worth flagging for any PAI use case that requires adaptive coordination rather than reproducible replay.