Pi-Jev-Guide

Redacted current and recent requests, applicable AGENTS.md text, tool parameters, and optionally redacted tool output. The version pins requests to https://api.typesafe.ai with request logging off, and without a key only local rules run.

jevguard login validates with a fixed test request then saves to jev-guard/auth.json under the Pi user directory with 0600 permissions, never in project config or session records. Logout deletes it and falls back to TYPESAFE_API_KEY if still set.

tool_call supports warn, confirm, or block while tool_result can only warn or hide because the tool already ran. On Jev errors pending actions become one confirmation, or a block with no interface, and confirms cover only the current call.

No. Completion tracking watches known writes via write and edit and treats observed commands as records, not proof of correctness. Shell-hidden modifications, pipelined or masked exit codes, and parallel checks may not count, and it never rolls back files.

Run pi install with the repository URL, then /reload in an open Pi session, /jevguard login to paste a TypeSafe API key into a hidden field, and /jevguard add to pick check timing, then local match or a Jev judgment, then the action on a hit. It is tested against pi-coding-agent 0.85.1 and Node 22.19 or newer, and is an independent experimental plugin rather than a TypeSafe or Pi product. A DSH adapter is not implemented.

task-frame, local-risk, semantic-action, redact-output, semantic-output, repeated-failure and completion-check, all on by default and toggleable by ID from /jevguard rules. They handle goal framing across turns, local risk rules, one Jev action check that sends four questions plus an output check, credential redaction, repeated identical failures on the third occurrence, and a completion reminder when known writes lack a later successful verification record.

Off by default; /jevguard recovery on enables it. It waits for Pi's agent_settled, then counts only clear network, timeout, temporary service or output-length errors, waiting 3 seconds first and then backing off at 20 and 40 seconds, with at most 3 automatic continuations per user task. Jev must judge an explicit next step within existing authorization at 0.9 or above and no need to wait for the user at 0.2 or below before a continuation is sent; on doubt it stays stopped.