Bernstein in August 2026: 47 adapters, v3.14.159, one maintainer

[chernistry]

Quick state-of-the-project, written so it's quotable without context. Refreshed 2026-08-10.

The operator problem this addresses

If you run more than one CLI coding agent in earnest, the workflow falls apart in three places: spawning them with the right credentials and sandbox; knowing which one produced which line of which file; deciding what to do when one of them breaks the build at 3am.

Bernstein is a single orchestrator over 47 of those CLIs. It runs them under a shared plan, signs every artefact they write, and gives you one place to inspect, approve, and replay.

Where v3.14.159 lands today

• 47 first-class CLI adapters in src/bernstein/adapters/registry.py: Claude Code, Codex, Gemini, OpenHands, Aider, Cursor, Goose, Qwen, OpenAI Agents SDK, Junie, AWS Q Developer, Devin for Terminal, Cloudflare Agents, plus ~30 more. Drop-in via cli_agent: <name> in a plan.
• ~2,500 test files, multi-OS CI green on Python 3.12 through 3.14. The CLI reference is CI-gated against the live command registry, so a documented flag the parser rejects fails the build.
• MCP server mode (stdio + Streamable HTTP), A2A v1.0 with signed agent cards, Cloudflare Workers tier, S3/GCS/Azure/R2 artifact sinks.
• Apache-2.0. No CLA. No copyright assignment.

What each surface does for the operator

Operator outcome	How it shows up	Where to look
Spawn N CLI agents under one plan, one approval policy	bernstein conduct / bernstein run	src/bernstein/cli/run_cmd.py
Inspect any run from a browser without standing up infra	bernstein gui serve → http://127.0.0.1:8052/ui/ (SPA ships in the wheel, no Node toolchain)	v2.0.0 changelog, tracking issue #1262
Prove which agent + model + prompt produced which byte range of which file	Lineage v2: Ed25519-signed per-artefact records, regulatory_class field, bernstein lineage verify <run_id>	src/bernstein/core/lineage/, src/bernstein/cli/commands/lineage_verify_cmd.py
Run in a sealed environment with no cloud egress	Air-gap wheelhouse + bernstein doctor airgap; cosign or GPG signature path	scripts/build_airgap_wheelhouse.py
Refuse spawns that union private data + untrusted input + external comms	Lethal-trifecta capability matrix; spawn-time, bypass-immune	src/bernstein/core/security/permission_*.py
Hand a stuck CLI session off to chat or back without losing context	`bernstein handoff emit	claim
Replay a deterministic run from the action cache	`bernstein cache action stats	replay <run_id>`
Run a 5-scenario chaos suite against cluster mode before trusting it	bernstein cluster bootstrap-ca, mTLS, 5 Prometheus metrics, 6 audit event types	src/bernstein/core/fleet/, cluster docs
Drive runs from GitHub PR comments or GitLab CI	github_app/ webhook handler + slash commands; adapters/ci/gitlab_ci.py	src/bernstein/github_app/, src/bernstein/core/trigger_sources/gitlab.py
Keep the spawner alive while an external driver runs phased work	Hold/release API: heartbeat-renewed leases, no fixed settle timer	POST /orchestrator/holds, src/bernstein/core/orchestration/holds.py
See every LLM call, tool call, and message a worker made	Per-agent JSONL instrumentation captured in real time via SDK hooks	.sdd/runs/<run_id>/.../agents/, src/bernstein/core/instrumentation.py
Control worker verbosity per role and prove what the terse profile saved	Response-style profiles wired into the spawn path + a ledger-derived, chain-anchored, recomputable savings report	bernstein cost profile-report, src/bernstein/core/cost/profile_attribution.py
Compare two spawn profiles on cost AND quality in one artifact	Three-arm A/B harness (baseline / control / candidate), medians, every cost figure referenced to a ledger line	bernstein eval ab --suite ... --arms 3
Prove two operators are running the same team	Named, sha256-pinned team manifests with drift detection and lineage-recorded digests	`bernstein team list
Compact a long worker's context without silently losing code or errors	Proactive compaction gated by zero-LLM invariant validators, each event an HMAC-chained receipt	bernstein compaction log --task <id>, src/bernstein/core/tokens/

Each row is measurable from a fresh checkout: spawn the surface, observe the output, fail it on purpose, watch the audit trail catch it.

Who built it

One person (me, @chernistry), with merged contributions from @shanemmattner, @0xbboyd, @alexanderxfgl-bit, @gittihub-jpg, @kite-builds, @oldschoola, @puneetdixit200, @sebastian-herrera, @TheCodingDragon0, @vedangit, @zerone0x, and the rest of the contributors graph.

Cited as prior art by Overcode, the NOVA Cognition Framework, and nibzard's agentic-patterns library. Listed in 15+ awesome-lists.

Recent momentum

The pace is real, and every change lands the same way: on a branch, through green multi-OS CI, merged to main, tagged. A representative recent stretch:

• v3.14.159 answered a first-use report (#3514) that described ninety minutes ending in nothing: the plan loader now raises named errors instead of loading malformed fields as defaults, five commands were removed outright, and the reference docs are gated against the registered CLI.

• Signed run receipts that verify fully offline, bernstein audit diagnose naming the first faulty step in a failed run, and Agent Plugins v1.0.0 manifests.

• Hold/release lifecycle API, explicit per-task turn budgets, per-agent run instrumentation, and a council-of-agents redesign with per-member cost attribution.

• Configuration-only model selection: no hardcoded fallbacks, unconfigured paths refuse rather than silently spend on the wrong provider.

• Response-style profiles + per-profile ledger-attributed cost reporting + a three-arm cost-and-quality A/B harness. The savings claim is recomputable from the audit chain, not asserted.

• Proactive context compaction with mechanical validation and signed receipts; a credential-shaped-content gate that refuses to ship secrets into summary requests.

• Schema-enforced worker completion payloads with typed refusal outcomes; sha256-pinned team manifests with drift detection.

Community PRs in this window merged with authorship preserved (thanks especially to @shanemmattner for a deep run across adapters, instrumentation, the council path, and the model-config cleanup).

What's deferred

Tracked on #1262, honest about a single-maintainer surface:

• Web UI polish: a11y audit, theme toggle, mobile-responsive pass, settings-screen wiring, fleet UI (scaffold only), front-end test suite + Playwright e2e.
• ACP server coverage is uneven vs MCP.
• Docs lag the adapter inventory; some adapters have only docstring-level coverage.

Normal scaling for a single-maintainer project. Not a crisis. Listed so anyone evaluating doesn't get surprised.

What's open for contribution

Smallest first:

1. ~30 seconds. Star the repo, or quote one line from this post wherever you naturally post.
2. ~5 minutes. Try pipx install bernstein && bernstein init on a weekend project. If anything is weird, drop a debug bundle on Bug tracking thread — post your debug bundles here #745 (the bug tracking thread). I triage same-day if scope is clear.
3. ~30 minutes. Pick any adapter under src/bernstein/adapters/ whose docs are thin and write a one-page user doc. Merge same-day.
4. ~multi-hour.
   • Web UI contributor entry points are pinned on Web UI development — tracking issue + contributor welcome mat #1262.
   • Orchestration benchmark submissions on community: Submit your orchestration benchmarks to bernstein.run #787.
   • The April adapter-pattern tickets feat: lazy tool search to keep MCP tool descriptions out of context #973-feat: synthesise reusable patterns from task transcripts into prompt updates and skills #987 are still open if you prefer a scoped feature with file paths + LOC estimate.

Not asking for money. GitHub Sponsors exists if anyone insists. The project isn't going closed-source; the licence isn't moving. The ask is attention, in the dose that fits you.

Working with the maintainer

Bernstein gets more of my bandwidth right now than ever, and it shows in the pace above. Alongside it I am open to a senior applied-AI-systems / agent-infrastructure / platform role (remote from Israel, or hybrid in an Israel office) and to serious consulting on multi-agent orchestration. If your team runs coding agents in production and wants this level of rigor in-house: alex [at] alexchernysh [dot] com · linkedin.com/in/alex-chernysh · alexchernysh.com.

• Alex (bernstein.run - @chernistry)