Question 1

Where is my API key stored, and is it actually protected?

Accepted Answer

In plaintext JSON at the agent directory's typesafe/config.json, defaulting to ~/.pi/agent and following PI_CODING_AGENT_DIR. The directory is set to mode 0700 and the file to 0600, writes go through a same-directory temporary file and rename, and reads, writes and deletes all refuse a symlink. If your threat model requires OS keychain protection, this does not provide it.

Question 2

Will the key leak into my session transcript?

Accepted Answer

Not through the extension, but it is visible while you type. /typesafe setup cannot mask input because pi's ctx.ui.input() has no masking support, so characters appear on screen. The key never enters the transcript, session file, tool output or error messages, and /typesafe status shows only the file path. Setup does not verify the key against the API, so a wrong one surfaces as a 401 on the first tool call.

Question 3

Can the model rewrite my rubric every time it calls the tool?

Accepted Answer

Not with typesafe_ask. The question lives in a YAML or JSON file: type, instructions and criteria come from the file only and the call cannot override them, while bind supplies the facts that change. A bind name the template does not use is an error, every slot must be supplied, and both directions fail loudly so a typo cannot silently drop a fact.

Question 4

Does loading the extension phone home or need an environment variable?

Accepted Answer

No. Loading performs no network access and reads no credentials; the credential file is only read when a tool or command actually runs, so a /reload or restart right after installing is safe offline. The five tools call POST /v1/systemone directly with native fetch, and there is no SDK runtime dependency.

Question 5

What are the five tools it exposes to the model?

Accepted Answer

typesafe_noul for a yes/no probability, typesafe_choice to pick one of a fixed set, typesafe_score to rate along ordered levels, typesafe_evaluate for several independent questions over one state in one request, and typesafe_ask for a question pinned in a local YAML or JSON file.

Question 6

Does setup verify the API key before storing it?

Accepted Answer

No. /typesafe setup does not call the API to verify the key, so a wrong key surfaces on the first typesafe_* call as a 401 and the tool then tells you to run setup again. Setup trims the value, rejects whitespace or non-ASCII, and leaves the old key untouched if you decline.

Question 7

How do I batch several questions in one call?

Accepted Answer

Use typesafe_evaluate with one state and multiple independent questions returned in a single request; the questions cannot see each other's answers, so make a second request only when a later question needs an earlier answer. Output includes answers, full distributions, confidence, and token usage.

Pi TypeSafe Jev

About this project

Highlights

Quickstart

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community