@zazencodes/pi-jev-model-router2.0.0

pi-jev-model-router

A Pi extension that adds a /jev slash command. Put /jev at the end of a prompt, or at the start, and Jev, TypeSafe's System 1 model, picks the model and the thinking level for it in about 200 ms. Then the prompt goes to that model.

design a plugin system for our CLI so third parties can add commands /jev

/jev design a plugin system for our CLI so third parties can add commands

Jev: complex (p=1.00) → gpt-6-astra, thinking high (p=0.91) in 198 ms

Both forms do the same thing. The trailing form lets you decide to route after you've written the prompt, and it is the one that works in print mode (pi -p). /jev anywhere else in a prompt is plain text. Prompts without /jev behave exactly like stock Pi.

Install

You need Pi (tested with 1.0.2) and a TypeSafe API key from https://console.typesafe.ai/keys.

pi install npm:@zazencodes/pi-jev-model-router

The package has no dependencies of its own: it calls Jev through Pi's classifier models. Add -l to install it for the current project only. To pull the latest version later:

pi update npm:@zazencodes/pi-jev-model-router

By default the extension calls Jev directly at TypeSafe, which Pi authenticates with TYPESAFE_API_KEY, so export it before you start Pi (to use another Jev provider, see Customize):

export TYPESAFE_API_KEY=your_typesafe_key
pi

To try it for one run without installing:

pi -e npm:@zazencodes/pi-jev-model-router

If /jev is sent to the model as plain text instead of showing a Jev: … notification, the extension isn't loaded. Check pi list.

How it works

<prompt> /jev   or   /jev <prompt>
   │
   ▼
Jev: two Choices in one call                     (~200 ms)
   tier:     simple │ standard │ complex
   thinking: low │ medium │ high │ xhigh │ max
   │
   ▼
cache guard ── mid-session switch? ── ask first
   │
   ▼
pi.setModel(Luna │ Sol │ Astra) + pi.setThinkingLevel(...)
   │
   ▼
the prompt, without /jev  ──►  the routed model answers

The routed model and thinking level stay selected after the turn. It's an ordinary model switch, shown in Pi's footer, and the next /jev routes again. If the model doesn't support the thinking level Jev picked, Pi clamps it and the notification shows both, e.g. thinking max → high (model limit).

By default, tiers map to these models (see Customize to change them):

Tier	Jev's criterion	Model
simple	A question, a lookup, or a small mechanical edit	openai-codex/gpt-6-luna
standard	An ordinary feature, bug fix, test, or code review	openai-codex/gpt-6-sol
complex	Subtle design, a cross-cutting change, or hard debugging	openai-codex/gpt-6-astra

The defaults use the openai-codex provider (a ChatGPT subscription, /login in Pi).

Why it pays

Standard uncached API prices per 1M tokens (USD, checked 2026-10-06):

Tier	Model	Input	Output
simple	GPT-6 Luna	$0.10	$0.50
standard	GPT-6 Sol	$2.00	$10.00
complex	GPT-6 Astra	$10.00	$50.00
router	Jev	$0.042	$0

Astra's token rates are 100x Luna's, so sending easy prompts to Luna saves money. Across 22 calls to Jev at TypeSafe inside Pi we measured 108–306 ms per routing decision (median 198 ms), including the network round trip. Other Jev providers add a hop: through OpenCode we measured a median of about 390 ms.

An LLM router would add another generation call before every routed prompt. Jev returns a typed label from a fixed set, with a calibrated probability, and bills input tokens only.

Why /jev asks before switching

Providers cache the conversation prefix and discount later reads of it. A different model can't reuse that cache, so after a mid-session switch it may re-read the whole history at full input price. When /jev would switch models in a session that already has messages, it asks first and says how many messages would be re-sent. Answer No and nothing changes; re-run the prompt after /new.

Route at the start of a task, not in the middle of one.

Try it

Start Pi on your usual model and try these in order:

#	You type	Routed to	Cache warning?
1	what does git rebase --onto do? Answer in one sentence. /jev	gpt-6-luna	No: first prompt of the session
2	write unit tests for the parseDuration function in utils/time.ts /jev	gpt-6-sol	Yes: Luna → Sol mid-session
3	Thanks. What does git bisect do? One sentence.	not routed	No: plain prompt
4	/jev we get an intermittent deadlock in the job queue under load, only in production. find the root cause	gpt-6-astra	Yes: answer No to keep the model unchanged

Then run /new and repeat prompt 4. On the first prompt of a fresh session it routes to Astra with no warning.

Customize

Create ~/.pi/agent/jev-model-router.json (in $PI_CODING_AGENT_DIR if you set it). This file is the defaults:

{
  "classifier": "typesafe/jev-latest",
  "tiers": {
    "simple": "openai-codex/gpt-6-luna",
    "standard": "openai-codex/gpt-6-sol",
    "complex": "openai-codex/gpt-6-astra"
  },
  "thinkingLevels": ["low", "medium", "high", "xhigh", "max"]
}

The file replaces the defaults as a whole, so give all three keys. Run /reload in Pi after you edit it. Models are provider/id, as in /model.

• tiers: the model for each tier. Any chat model listed in /model works. Leave out a tier to disable it: without complex, Jev chooses between simple and standard only, so nothing goes to Astra.
• thinkingLevels: the levels Jev may pick from. Leave out xhigh and max to cap reasoning at high.
• classifier: the Jev model that routes. Pick the one you have credentials for:

classifier	Authentication
typesafe/jev-latest	TYPESAFE_API_KEY (fastest: direct to TypeSafe)
openrouter/typesafe/jev-1.13	OPENROUTER_API_KEY or /login
opencode/jev-1.13, opencode/jev-1.13-free	OPENCODE_API_KEY
vercel-ai-gateway/typesafe-ai/jev	AI_GATEWAY_API_KEY
cloudflare-workers-ai/typesafe/jev	CLOUDFLARE_API_KEY and CLOUDFLARE_ACCOUNT_ID

Example: disable Astra and the highest thinking levels

This config never routes to Astra and never uses xhigh or max:

{
  "classifier": "typesafe/jev-latest",
  "tiers": {
    "simple": "openai-codex/gpt-6-luna",
    "standard": "openai-codex/gpt-6-sol"
  },
  "thinkingLevels": ["low", "medium", "high"]
}

Jev now chooses between two tiers and three thinking levels. The criteria for the remaining options are unchanged, so a hard prompt goes to the strongest tier and level you left on. With this config, prompt 4 from Try it (the production deadlock) routes to gpt-6-sol with thinking high instead of to Astra.

Keep all three keys: the file replaces the defaults as a whole, so a config that lists only tiers is an error. Run /reload in Pi after saving it.

Rules

Each of tiers and thinkingLevels needs at least two entries for Jev to choose between. The tier names and Jev's criteria for them are fixed in pi-jev-model-router.ts. Criteria wording moves results a lot, so if you change them in a clone of the repository, test on your own prompts. A clone installs with npm install && pi install . and loads in place, so /reload picks up edits.

Errors

The extension fails loudly and has no fallback model. You get an error, and the prompt is not sent, when:

• jev-model-router.json is invalid. Pi reports this when it loads the extension.
• The classifier is not configured (e.g. TYPESAFE_API_KEY is missing) or the Jev call fails.
• The routed model isn't in Pi's model registry. Fix tiers.
• pi.setModel fails, usually because the provider isn't logged in. Run /login.
• You use /jev while the agent is working. Wait for the run to finish.
• A mid-session switch needs confirmation and Pi has no UI (print or JSON mode).

In print mode (pi -p), put /jev at the end of the prompt. pi -p "/jev <prompt>" routes and then exits before the prompt runs, because Pi doesn't wait for a run that a command starts.

Changelog

See CHANGELOG.md. Releases are tagged vX.Y.Z and follow semver: a patch release fixes bugs or docs, a minor release adds features, and a major release contains breaking changes.

License

GPLv3. You can use, modify and share this code, and any version you distribute must also be released under GPLv3.

Created by ZazenCodes