@techs/pi-decision0.2.0
Jev-backed tool guardrail for pi
Untested on omp
This is a pi package. No one has reported whether it works on omp yet. If you try it, add your result below.
Reports 0
No reports yet.
Readme
@techs/pi-decision
English | 简体中文
A Pi tool_call guardrail extension using Jev as its current judgment provider. Before each tool call, it assesses six risk dimensions: destructive, secretExposure, privacyExposure, externalSideEffect, privilegeEscalation, and scopeViolation. You can disable or customize built-in dimensions and add your own. The risk policy comes from the provider-neutral core, but this Pi extension currently creates the Jev provider directly; using another judgment model requires a new adapter and Pi integration. The agent's main model still generates responses, and Pi's own permission rules continue to apply.
export AI_GATEWAY_API_KEY=your_gateway_key
pi install npm:@techs/pi-decision
The default shadow mode observes asynchronously without blocking tools. Configure credentials for Pi's main model separately. Judgments are written to $XDG_STATE_HOME/dsh-decision/pi-audit.jsonl, or ~/.local/state/dsh-decision/pi-audit.jsonl when XDG state is unset.
| Environment variable | Default | Purpose |
|---|---|---|
AI_GATEWAY_API_KEY |
None | Vercel AI Gateway key; takes priority over JEV_API_KEY |
JEV_API_KEY |
None | Fallback key; required for the direct backend |
PI_DECISION_JEV_BACKEND |
vercel |
vercel for Gateway; direct for api.typesafe.ai |
PI_DECISION_ENFORCEMENT |
shadow |
shadow observes; enforce applies judgments (experimental) |
PI_DECISION_TOOLS |
All tools | Comma-separated exact tool names, such as bash,write |
PI_DECISION_ON_FAILURE |
allow |
Stance when a judgment request fails: allow, ask, or deny |
PI_DECISION_OUTBOUND |
redact |
Mask recognizable secret patterns; raw sends unchanged input |
PI_DECISION_RISKS |
None | JSON custom risk dimensions, like dsh guardrail.customRisks |
PI_DECISION_AUDIT |
on |
Set to off to disable audit records |
PI_DECISION_AUDIT_PATH |
XDG state path | Audit JSONL path |
- Integration and verification (Chinese)
- Threshold evaluation (Chinese)
- Repository README