User repository on GitHub
Hybrid coding harness: System 2 writes, System 1 (Jev) runs reflexes.
- Forks
- 1
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Tools & CLIs
Bicameral
A hybrid coding harness for Pi where any LLM writes the code while Jev returns typed probabilities that deterministic policy code turns into allow, confirm, block, warn, or steer.
- Stars
- 8
- Language
- TypeScript
- License
- MIT
- Reading time
- 4 min
- Published by
- Community
- Author
- AbdelStark
- Last verified
- Sep 2026
About this project
A hybrid coding harness for Pi. System 2 (any LLM Pi runs) writes the code. System 1 (TypeSafe Jev, trained with Reinforcement Learning for Calibrated Decisions / RLCD) returns typed probabilities. Deterministic policy code turns those scores into allow / confirm / block / warn / steer. System 1 scores only. It never generates text. Isolation is still the job of containers. This is not a sandbox. Architecture · Reflexes · HUD · Install · Policy · Library · Data flow · Fail-safe · Status.
Quoted from the project's README. Read the full README.
Highlights
- Gate scores exfiltration and secret access on tool calls, Honest Finish scores weakened tests on edits, and Stuck Detector watches for repeats at turn end.
- Speculative prefetch starts the Gate judgment while the LLM is still streaming the tool call, under a 900 ms deadline.
- Hints and follow-up turns are templates filled with scores and pre-signals, not text written by a model.
- State is redacted before it reaches Jev: raw file contents, unredacted secrets, full transcripts, policy files, and other API keys are not sent.
- bicameral.decision audit entries do not enter the LLM context, and /why prints the recorded scores, policy rule, and latency with no model-written explanation.
Quickstart
bash
pnpm install
pi install "$(pwd)/packages/pi-bicameral"
cp examples/bicameral.yaml ~/.pi/agent/bicameral.yamlWatch out
MIT-licensed. The packages are not on the public npm registry yet and the GitHub remote is private until launch; running it needs Node 20+, pnpm, and Pi 0.85.1, and it is explicitly not a sandbox.
Questions about this entry
Short answers drawn from the linked project or page.
It is explicitly not a sandbox. System 1 only scores and never generates text, and the README is blunt that isolation is still the job of containers. What you get is a scoring gate in front of tool calls where policy code you own decides allow, confirm, block, warn, or steer.
Not anonymously. The packages are not on the public npm registry yet and the GitHub remote is private until launch, so there is no public clone URL that works for outside visitors. From a checkout you already have, it is pnpm install then pi install on packages/pi-bicameral; the published pi install npm:pi-bicameral path does not exist yet.
With no TYPESAFE_API_KEY the extension uses UnavailableBackend and applies the documented fail-safe: high-risk actions confirm or block, everything else allows. Live Gate judgments run under a 900 ms deadline, and a headless session maps confirm through headless_confirm_as, which defaults to block. Without the key you still get enforcement, just without the scores.
Redacted, truncated JSON state for the current reflex plus the pack's questions, posted to the System One endpoint. Raw file contents, unredacted secrets, full session transcripts, policy files, and other providers' API keys are not sent; redaction replaces private keys, JWTs, AWS key ids, and high-entropy assignments first. Decisions, HUD, /why output, and audit entries stay local and never enter the LLM context.
Gate fires on tool_call and prefetches the judgment while the LLM is still streaming, scoring exfiltration, injected intent, secret access, scope escape and reversibility. Honest Finish runs on tool_result after an edit or write and checks for weakened tests or stubs. Stuck Detector runs at turn_end and watches for repeat failures, oscillation and no progress.
Yes. Set mode: observe and it records decisions and updates the HUD without blocking, confirming or steering; mode: enforce applies the action. Thresholds live in bicameral.yaml merged over DEFAULT_POLICY, and invalid YAML keeps the previous policy.
The footer shows S1 with a p50 latency, or S1 degraded once estimated spend reaches the budget (default $5). /why prints the recorded reflex, pack hash, action, the firing policy rule, latency, and the question answers with the redacted state; it is the record itself, with no model-written explanation.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Vercel's fx safety reviewer, measured at 18× faster
Reddit: Early experiments wiring Jev into Pi as a safety layer and model router.
X: Decision outputs in Grep
Hacker News: jevmem keeps Claude Code project memory by asking Jev which turns are worth saving.
avinashjetwani61 pointsJevmem – automatic project memory for Claude Code, built on JevRead the thread on Hacker NewsMore Jev resources like this
A Pi extension that turns constraints stated in conversation into a scoped, replayable policy and checks every side-effecting tool call against it before it runs, with Jev only classifying how each message changes the policy.
Jev Tools & CLIs#community#typescript#guardrails
Community
Nyarlathoteppppp/pi-heedTypeScript
Nyarlathoteppppp/pi-heedTypeScriptPi extension that supervises a coding agent with Jev judgments: it holds risky tool calls, checks writes against project Markdown rules, and feeds most issues back to the agent as a steer instead of interrupting you. The conscience is beta and off by default.
Jev Tools & CLIs#community#typescript#guardrails
Community
DevMortimer/pi-wardenTypeScript
DevMortimer/pi-wardenTypeScriptA Jev-guided context compaction library for coding agents: instead of summarizing old context, Jev decides which tool calls and results still matter while user and assistant text is kept verbatim.
Jev Tools & CLIs#community#typescript#compaction
A Claude Code Stop hook that reads the turn transcript locally and spends one four-question Jev call only when files changed with no passing check since, blocking unverified done claims.
Jev Tools & CLIs#community#javascript#claude-code
Community
valentynkit/jev-belayJavaScript
valentynkit/jev-belayJavaScriptA Pi extension that keeps conversation text verbatim while pruning stale tool history with Jev keep-or-drop decisions, falling back to Pi's built-in summary compaction when pruning cannot free enough context.
Jev Tools & CLIs#community#typescript#compaction
A Codex plugin and npm library that wraps native session compaction: Jev scores every tool call and result, and the verbatim history it keeps is re-injected after Codex summarizes. The README does not recommend using it.
Jev Tools & CLIs#community#typescript#codex
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.