NewPi Warden
A Pi extension watches tool calls and agent progress for safety and completion problems.
Added to Jevfast
NewHow it uses Jev
Before a guarded tool call, pi-warden builds state from the user task, a redacted action summary, recent conversation, and optionally the task spine, plan, project rules, pattern hits, and other enabled context. Jev returns typed Noul probabilities for questions such as irreversibility and off-task status, plus a Choice for scope; local thresholds and policy combine those answers with offline pattern checks to allow, warn, steer, or hold the call. In steer mode, the result goes back to the agent as tool feedback.
What Jev decides
Illustrative Jev decisionExample answers · not a recorded Jev response · Source ↗
Question 1 · irreversible
YOUR APP
INSTRUCTION
Would this guarded action have a hard-to-undo effect?
STATE
User task, redacted guarded action, recent conversation, and any enabled plan, project rules, and pattern hits.
JEV · NOUL
YesNo
Question 2 · scope
YOUR APP
INSTRUCTION
How does this guarded action relate to the active task and earlier context?
STATE
User task, redacted guarded action, recent conversation, and any enabled plan, project rules, and pattern hits.
JEV · CHOICE
- expected_step
- plausible_side_step
- unrelated
- unclear
App workflow
- Preflight the tool call
Skip read-only calls and run deterministic local pattern checks; retain pattern hits as request evidence in evidence mode.
- Build typed judgment questions
Send the redacted task/action/context state and ask independent Noul questions for irreversibility and other judgments, plus a Choice question for task scope.
- Apply local policy
Convert Jev answers to a verdict using configured thresholds, together with pattern and authorization results.
- Steer or hold
In steer mode, block a held call and return corrective guidance to the agent; otherwise warn, ask for confirmation, or advise according to mode.
Why it is interesting
It combines local pattern detection with contextual Jev judgments and turns most findings into guidance the agent can act on. Its documented real-session example records a held database-reset call and an `irreversible` answer of 0.85.