@mohan-cao/pi-jev-response-router

pi extension that calls Jev before each agent run, selects a response policy, and verifies the final answer.

• bounded_verification — at most three concrete claims/questions; answer with correct / partially correct / incorrect plus concise corrections.
• decomposition_required — important assumptions or tradeoffs materially affect the answer; decompose first, then synthesize.
• normal — leave Pi's normal response behavior unchanged.

Jev is called through its native HTTP API and is not registered as a chat model for obvious reasons. The extension registers an auth-only pi provider so /login can store and resolve the Jev API key using Pi's normal auth flow.

How routing works

Rather than one mutually exclusive 3-way Choice (where normal competes in the argmax), the classifier asks narrow, independently evaluated questions:

1. requires_decomposition — does a reliable answer require decomposition?
2. bounded_verification — is this at most three independently checkable claims?
3. premise_defect (Score 0–3) — does the request presuppose something false, or true only under a narrower framing than it implies?

The mode is composed in code with decomposition taking precedence and normal as the residual, which keeps decomposition-shaped requests from being misrouted as normal. requires_decomposition judges the current request only, so an earlier answer's verbosity cannot inflate the score.

The selected policy is injected as a named system-prompt section (jev-response-policy), so Pi emits a minimal prompt patch and the provider cache prefix survives. (Returning systemPrompt would replace the whole prompt on every mode change, i.e. a full cache miss.)

Premise correction

premise_defect is a modifier, not a mode, so it composes with any response shape. Above PI_JEV_ROUTER_PREMISE_THRESHOLD (default 2.8 of 3) a second section (jev-premise-policy) asks the model to state the correction in one sentence before answering. It is deliberately not applied to bounded_verification, which already emits a verdict and corrections.

Level 2 of the scale ("worth naming, but the request is still answerable") is logged but never acted on by default — that is the band for minor slips in wording rather than wrong beliefs.

Verification (signal only)

After a run settles, the final answer is sent back to Jev: does it address the request, does it avoid committing where a position was called for, and did it meet its obligation — correcting a false premise or surfacing the inputs the answer actually depends on. The result is surfaced as a footer status:

• 🤷 evasive — the answer dodged or did not commit
• 🚩 unmet — technically true but it did not do the required work

The footer preference selects the rendering: compact (icon + label, default), icons (glyph only), or off.

This is verify-only: it never retries or rewrites the answer. Retry is deliberately out of scope for now.

Two independent footer signals

Both run after generation, in parallel with verification. They are separate features with separate toggles, because they are orthogonal — you can be stuck while implementing, or framing a problem badly during general conversation.

Model nudge (/jev-router phase on|off) — what kind of work is next, and whether your running model suits it:

• ↪ build · <model> — the work moved; switch when you want to

Fires only when the running model is known to serve a different phase; an unmapped model yields no nudge. It never switches models and never blocks a prompt — switch with /model, and the nudge clears itself. Inert until routes are configured (see below).

Coaching hint (/jev-router coaching on|off) — is this conversation progressing?

• ♾️ paralysis — analysis paralysis; consider zooming out
• 🖼️ framing — the framing is the blocker, not the effort

Display-only. It never gates the model nudge and never changes routing.

| Environment variable | Default | Meaning | | --- | --- | --- | | PI_JEV_PHASE_BUILD_MODEL | unset | Fallback model for the build phase | | PI_JEV_PHASE_DESIGN_MODEL | unset | Fallback model for the design phase | | PI_JEV_PHASE_GENERAL_MODEL | unset | Fallback model for the general phase | | PI_JEV_PHASE_CONFIDENCE_THRESHOLD | 0.7 | Minimum confidence before a model nudge | | PI_JEV_PHASE_HISTORY_TURNS | 8 | Conversation turns supplied to the phase judge | | PI_JEV_TRAJECTORY_THRESHOLD | 0.7 | Minimum confidence before a coaching hint | | PI_JEV_TRAJECTORY_HISTORY_TURNS | 8 | Conversation turns supplied to the trajectory judge |

Phase routes

Routing needs to know which model serves which phase. Set it from Pi rather than from the environment:

/jev-router route             # list the routes and where each came from
/jev-router route build       # pick from the models you have available
/jev-router route build gpt-5-codex
/jev-router route build clear

route <phase> with no model opens Pi's model picker over modelRegistry.getAvailable(), so choosing a route does not mean knowing model ids by heart. The choice is written to ~/.pi/agent/pi-jev-response-router.json and survives /reload and new sessions.

Precedence: a saved route beats the matching environment variable, and unset phases fall through to the variable. clear stores an explicit empty value, so it also overrides a variable that would otherwise set that route. /jev-router status marks saved routes with *.

Routes are matched on the bare model id — the value ctx.model.id returns — so two providers offering the same id are indistinguishable here.

Decision log

On by default: one JSON line per settled turn at ~/.pi/agent/jev-decisions.jsonl (/jev-router log on|off). It records judgments, not conversation text — the classification mode and signals, the verification result, the phase and trajectory judgments, and the model that answered.

Acceptance of a model nudge is derived by diffing consecutive records: if recommendedModel at turn N equals modelRunning at turn N+1, the nudge was acted on. That is the only correctness proxy available without hand-labelling — and it only produces data when routes are configured, since without them no nudge ever fires.

Install from npm

pi install npm:@mohan-cao/pi-jev-response-router

then start Pi and run /login, select TypeSafe Jev (response router) and paste the TypeSafe API key.

You can also provide TYPESAFE_API_KEY in the environment. An explicitly stored /login key takes precedence.

Commands

/jev-router help
/jev-router status
/jev-router on
/jev-router off
/jev-router verify on
/jev-router verify off
/jev-router phase on
/jev-router phase off
/jev-router coaching on
/jev-router coaching off
/jev-router log on
/jev-router log off
/jev-router route
/jev-router route <phase> [model]
/jev-router debug on
/jev-router debug off
/jev-router footer compact
/jev-router footer icons
/jev-router footer off
/jev-router clear-cache
/jev-router classify UDP preserves datagram boundaries, right?

/jev-router classify ... lets you demo the Jev route without invoking the main model.

Configuration

All configuration is optional:

| Environment variable | Default | Meaning | | --- | --- | --- | | TYPESAFE_API_KEY | unset | Fallback Jev credential when no /login key is stored | | PI_JEV_ENDPOINT | https://api.typesafe.ai/v1/systemone | Native System One endpoint | | PI_JEV_MODEL | jev-latest | Jev model selector | | PI_JEV_ROUTER_TIMEOUT_MS | 5000 | Per-request HTTP timeout | | PI_JEV_ROUTER_RETRIES | 2 | Retries for HTTP 429/529 | | PI_JEV_ROUTER_DECOMPOSITION_THRESHOLD | 0.6 | P(decomposition) at or above which decomposition wins | | PI_JEV_ROUTER_BOUNDED_THRESHOLD | 0.6 | P(bounded verification) at or above which bounded verification wins | | PI_JEV_ROUTER_PREMISE_THRESHOLD | 2.8 | Expected premise defect (0-3) at or above which the premise is corrected | | PI_JEV_ROUTER_HISTORY_TURNS | 4 | Prior turns included in Jev state (0 disables) | | PI_JEV_ROUTER_CACHE_TTL_MS | 300000 | Classification cache TTL (0 disables) | | PI_JEV_ROUTER_CACHE_MAX | 64 | Classification cache entry cap | | PI_JEV_ROUTER_VERIFY | true | Post-generation verification on by default | | PI_JEV_ROUTER_VERIFY_EVASIVE_THRESHOLD | 0.6 | P(evasive) at or above which to flag the answer as evasive | | PI_JEV_ROUTER_VERIFY_ANSWERS_THRESHOLD | 0.35 | P(answers the request) at or below which to flag the answer as evasive | | PI_JEV_ROUTER_VERIFY_OBLIGATION_THRESHOLD | 1.5 | Expected obligation failure (0-3) at or above which to flag the answer as unmet |

PI_JEV_ROUTER_MIN_CONFIDENCE is deprecated. In 0.1.x it was effectively a no-op (confidence is in [0, 1] and the default was 0). It now serves as a shared fallback default for the two real thresholds when they are unset.

Persisted preferences

/jev-router on|off, debug, verify, and footer write to a small JSON file under the agent config directory, so they survive /reload and new sessions:

~/.pi/agent/pi-jev-response-router.json
{
  "enabled": true,
  "debug": false,
  "verify": true,
  "footer": "compact"
}

The path honors PI_CODING_AGENT_DIR. A missing or corrupt file falls back to the environment defaults. Only runtime toggles persist; endpoint, model, thresholds, and routes stay in the environment. Precedence is file > environment > default, so a command is durable while an env var still works as a one-shot override when no file exists.

footer selects how footer statuses render — compact (icon + label, default), icons (glyph only), or off. It is reserved for the work-phase and coaching statuses and has no effect on the current quality status yet.

The router fails silently: if Jev is unavailable, unauthenticated, or returns an invalid schema, Pi answers normally. Turn on debug notifications to surface failures and route decisions in the TUI.