pi-system-one

System One decisions for Pi. Plug in TypeSafe Jev, Reflex, or your own provider.

pi install npm:pi-system-one

Pi built-in classifier vs pi-system-one (as of Pi v1.0.3)

Pi already supports Jev through codemode. It can classify one question or combine classification with other tool calls in a script. This extension exposes decisions as a dedicated system_one tool, with a /judge prompt and a skill that guides when to use it. Choose based on your workflow, not on a claim that either route is more accurate.

Pi's v1.0.3 Codemode guide documents scripts, model calls, and orchestration. Its CLI guide explains how to enable codemode. Pi extensions can also call classifiers without Codemode; the table compares the agent-facing Codemode route with this extension's agent-facing tool, not every way to write a Pi extension.

Pi v1.0.3 defines its Score answer with only score and confidence. Its response parser copies only those two Score fields. TypeSafe's HTTP API returns probabilities and legend too. Pi's classifier input types and codemode docs document the narrower inputs shown above. These references describe Pi v1.0.3, not future releases.

If you already use codemode to combine tools and only need Pi's classifier response, use Pi's built-in route. For standalone decisions, choose this extension when you want a direct tool without a script or Codemode setup. It also provides prompt guidance, compatible custom endpoints, and the full Score distribution when the backend supplies it. The extension can run on a host without codemode if that host supports its extension APIs. Its OMP package entry makes it installable in OMP, but it does not guarantee support for every Pi or OMP version. /so config can save non-secret settings or select Pi's native classifier. Native mode rejects Score questions because Pi v1.0.3 does not expose their probability distribution. Optional auto mode routes native-compatible Choice/Noul requests to Pi's classifier when registered and other requests to TypeSafe HTTP; it does not retry a failed call through another backend. See Pi classifier docs for v1.0.3.

For Oh My Pi (OMP), install the published package with:

omp install npm:pi-system-one

For a local checkout, install or load the package directory directly:

omp install /path/to/pi-system-one
# or
omp --extension /path/to/pi-system-one

The system_one tool registers at startup. Without settings, it uses TypeSafe Jev through this extension's HTTP provider and asks Pi for your TypeSafe credential at call time. Run /login typesafe first, or set TYPESAFE_API_KEY.

Custom endpoint

For Reflex or another backend that supports POST /v1/systemone, run /so config and choose custom, or retain your existing environment setup:

export SYSTEM_ONE_BASE_URL=http://localhost:8008
export SYSTEM_ONE_MODEL=reflex
# Set SYSTEM_ONE_API_KEY only when your endpoint requires a key.

SYSTEM_ONE_BASE_URL selects a custom endpoint when no mode is saved. A saved mode and custom endpoint win on restart. TypeSafe defaults to jev-latest; other endpoints may require SYSTEM_ONE_MODEL.

Configuration

Run /so or /so config to select typesafe (default HTTP path), custom (a configured System One endpoint), native (Pi classifier), or auto (per-request routing). /so config auto, /so config native, and /so config typesafe switch directly. Non-secret settings are saved to <agent-dir>/pi-system-one.json, where <agent-dir> is Pi's agent directory (~/.pi/agent by default or PI_CODING_AGENT_DIR). Terminal Pi and Paseo must use the same agent directory to share that file and Pi's saved login. Restart or reload an already-running client after the other client changes settings. Use /so status to inspect current mode and credential availability.

For example, a saved custom endpoint looks like this:

{
  "mode": "custom",
  "baseUrl": "http://localhost:8008",
  "model": "reflex",
  "timeoutMs": 10000
}

Never store a key in JSON. A key entered into /so config stays in this session only. Leave key input blank to keep it; enter - to forget it. For custom endpoints, set SYSTEM_ONE_BASE_URL with SYSTEM_ONE_API_KEY if using environment credentials. An environment key without an environment endpoint never follows a saved custom URL. Custom URLs saved by /so config must be HTTP(S), without URL credentials, query, or fragment; put authentication in the key input or environment instead. Native mode supports Choice and Noul but rejects Score questions rather than hide missing probabilities. For full Score results, select typesafe, custom, or auto. Auto uses native only when the complete request is compatible and Pi has a registered classifier. Otherwise, it uses canonical TypeSafe HTTP, not a custom endpoint. A failed native call does not trigger an HTTP retry. Registered models may still fail without valid credentials.

Use /so debug to inspect up to 20 recent system_one calls in a terminal overlay. It shows selected provider (pi-native, typesafe, or configured), route reason, sanitized HTTP POST target, outcome, timing, and question-type counts. Failed calls distinguish provider selection from an attempted request, and show fixed native failure categories (unsupported Score, classifier exception/error, invalid response, or abort) without provider-controlled error text. Diagnostics stay in memory for the current Pi session; they never store request state, criteria, credentials, raw URLs, or raw provider error text. Target display retains origin and common API path segments but redacts URL userinfo, queries, fragments, and other path segments. In RPC mode, /so debug sends a brief notification instead; JSON and print modes have no UI. /so status remains focused on settings.

A current-session selection takes effect immediately. Saved mode and saved custom endpoint win on restart. SYSTEM_ONE_BASE_URL selects a custom endpoint only when no mode is saved; without either, TypeSafe HTTP remains the default. Environment keys and models are used only with their matching endpoint, while SYSTEM_ONE_TIMEOUT_MS can override a saved timeout. /so status shows the mode source and notes when a saved choice ignores an environment endpoint. To return to environment-only routing, remove the saved non-secret settings file. When /so config changes a custom URL, blank model input clears the old endpoint's model; enter a replacement when needed.

/judge

The fastest way to use System One. A prompt shortcut that expands into instructions calling the system_one tool explicitly, so routing doesn't depend on inferring intent. For short, unstructured asks:

/judge should I ship this? <paste diff plus test summary>

/judge which team should take this ticket? <paste ticket>

system_one skill

For everything else, the bundled skill teaches the agent when to call system_one, which question type to use (choice / noul / score), and how to read answers. It loads automatically — just ask for quantified answers over material already in context. Probabilities and confidence are the words that route to the tool. Retrieve facts first; the judge weighs supplied state, it cannot recall facts.

See SKILL.md for the full reference and use cases for copy-pasteable recipes.

Examples — same decisions, quantified wording:

Pick one: fix forward with a minimal patch, or revert and re-land
later. Give me the probability of each and your confidence.
<paste context>

This just landed in our support queue — which team is the most likely
owner: billing, technical, or sales? Give me the probability for each
and your confidence for the pick. Also, how urgent is it on a low /
medium / critical scale? And should I escalate (yes/no probability,
no confidence needed)?
<paste ticket>

Should I ship this or revert it? Give me the probability that shipping
causes a follow-up bug within a week (yes/no probability only — noul
has no confidence field). Also rate the test coverage — none, partial,
or full — with probabilities and confidence for the pick.
<paste diff plus test summary>

Independent judgments over the same context should be batched into one call, and answers should come back as calibrated, state-conditional probabilities — with confidence for choice/score only (noul has no confidence) — not prose guesses, and never as recalled facts.

Running your own backend? See the system one core package