Bicameral
An experimental Pi coding harness where Jev scores reflexes and deterministic policy decides how the agent proceeds.
What Jev decides
Whether a tool call risks exfiltration or scope escape, whether an edit weakens work, and whether the agent is stuck or making progress.
Bicameral evaluates proposed tool calls, edits, and recent progress with Jev judgment packs, then maps typed probabilities to allow, confirm, block, warn, follow up, or steer actions.
- Jev role
- Score safety, edit quality, and progress reflexes around a separate coding model.
- Input
- Redacted, capped state containing the user goal, proposed tool call, compact edit signals, recent action summaries, and one reflex question pack.
- Output
- Noul and Score answers consumed by local policy.
- Decision frequency
- At tool-call, edit-result, and turn-end hooks, subject to deadlines and budget policy.
- Fallback
- Use deterministic high-risk command patterns when Jev is unavailable; require confirmation or block in headless mode while allowing lower-risk work.
How it works
Overview
Bicameral is an early hybrid extension for the Pi coding agent. A normal language model still plans, explains, edits, and calls tools. Jev receives redacted and truncated state for three reflexes: Gate checks exfiltration, injected intent, secret access, scope escape and reversibility; Honest Finish checks weakened tests and introduced stubs; Stuck Detector checks repeated failure, oscillation and lack of progress.
Local TypeScript and YAML policy converts those probabilities into bounded actions. A HUD and audit record show the policy rule and answers without inserting the record into the coding model context. Offline tests use a FakeBackend and do not call the TypeSafe API. At review time the project was version 0.1.0 with three commits, unpublished packages, and no live Jev test claim from this directory.
- Language
- TypeScript
- License
- MIT
- Author
- AbdelStark
Sources
Related projects
Jev Review
A code-review pipeline assembled from small typed judgments instead of one long reviewer prompt.
What Jev decides
What is risky, which files and evidence matter, what could fail, how severe it is, and who should review it.
Distill
Route coding-agent work, reasoning effort, and context retention through bounded Jev decisions.
What Jev decides
Which configured model and effort should handle a call, what evidence deserves attention, whether another attempt is useful, and what context should be retained.
jev-model-router
Route Claude Code model tiers and reasoning effort with one parallel Jev classification.
What Jev decides
Which fast, balanced, or deep tier fits the work, how much reasoning it needs, and whether production risk requires escalation.