Overview & Result
Autonomous coding agents can hallucinate or exceed their mandate by deleting files or executing unexpected terminal commands. Pi-Heed acts as an in-line auditor: before any side-effecting tool runs, Jev compares the action against the user's prompt and halts unauthorized operations.
How Jev fits in the loop
- Listen for proposed agent tool calls (e.g. bash execution, file overwriting).
- Feed the user prompt and the tool invocation payload into Jev Score.
- Evaluate the alignment score: does the requested action reasonably fulfill the user goal?
- Auto-approve safe actions; prompt the user or abort if the score falls below threshold.
How to reproduce
- Clone Nyarlathoteppppp/pi-heed and hook it into the Pi agent runtime.
- Trigger an agent workflow that attempts file writes, bash commands, or network requests.
- Inspect how Pi-Heed scores tool parameter alignment and blocks unprompted destructive actions.
Why this build matters
Solves the primary trust barrier in autonomous coding agents by enforcing automated, low-latency oversight over dangerous tools.
Limitations
- Relies on clear natural language descriptions of the tool's intended effect.
- Cannot analyze deeply obfuscated shell scripts or complex binary executions.
- Requires user confirmation for low-confidence boundary decisions.