pi-fast-all

pi-fast-all

One fast-mode toggle for pi that works across providers.

Providers all sell "pay more, go faster," but each spells it differently. pi-fast-all keeps the differences in a rule table instead of in code, so one /fast covers all of them — and adding a provider is a config edit, not a release.

provider	mechanism
anthropic	body speed: "fast" + anthropic-beta: fast-mode-2026-02-01
openai	body service_tier: "priority"
openai-codex	body service_tier: "priority"
anything else	write a rule

Fast tiers bill at roughly 2x the standard token rate (2.5x on gpt-5.5), carry their own rate limits, and — on Anthropic — do not share prompt-cache prefixes with standard-speed requests. So this ships off by default and is a session toggle rather than a persistent setting.

Install

pi install npm:pi-fast-all
pi -e npm:pi-fast-all      # or try it for one run only

Then /reload (or restart pi).

Usage

/fast          # toggle for this session
/fast on
/fast off
/fast once     # fast for the next turn only, then back to normal
/fast auto     # fall back to the config default
/fast status   # state, model, active mechanism, and what you have spent
/fast rules    # list every mechanism, flagging models that do not resolve

/fast once is the one to reach for. The usual case is "this next question is hard, spend the money" — leaving the toggle on afterwards is how a 2x bill happens by accident. It stays armed across the whole turn, tool calls included, and disarms when pi settles.

Start a session with it already on:

pi --fast

While fast mode is on and the current model matches a rule, a fast marker appears in the footer. /fast status explains itself when it isn't active:

Fast is on (session override) but inactive for anthropic/claude-sonnet-4-6:
anthropic has fast mode, but not for claude-sonnet-4-6.

Rules

A rule declares how one provider does fast:

{
  "provider": "anthropic",              // required
  "api": "anthropic-messages",          // optional narrowing
  "models": ["claude-opus-5"],          // ids, "provider/id", or ["*"] for all
  "body": { "speed": "fast" },          // merged into the request body
  "headers": { "x-example": "1" },      // set on the model before requests
  "betaHeader": "fast-mode-2026-02-01", // appended to anthropic-beta, deduped
  "costMultiplier": 2,                  // what fast multiplies the bill by
  "costCountedByHost": false,           // true if pi already prices this tier
  "note": "~2x token cost"              // shown by /fast status
}

A rule applies only when provider, api, and model all match. If any body key is already present on the payload, the extension backs off entirely — another extension owns that knob.

Config

Optional file at <agent-dir>/extensions/pi-fast-all.json (stock pi: ~/.pi/agent/extensions/pi-fast-all.json):

{
  "enabled": false,
  "indicator": "widget",
  "indicatorText": "⚡fast",
  "indicatorColor": "#ffa657",
  "warnAboveUsd": 5,
  "extraRules": []
}

• enabled — state at session start. /fast overrides it for the session and is never written back.
• indicator — widget (above the editor), status (footer status line), both, or off. widget is the default because a host may replace pi's footer — arc did, and its replacement dropped every extension status until that was fixed.
• indicatorText / indicatorColor — the marker. A #rrggbb colour is emitted as a truecolor escape; anything else (warning, accent, …) goes through the theme.
• warnAboveUsd — recolour the marker once the fast premium passes this many dollars. 0 disables.
• rules — replaces the built-in rules.
• extraRules — appends to whichever rule set is in effect. Use this to add a provider without losing the defaults.

See pi-fast-all.example.json.

How it works

Step	Mechanism
body patch	before_provider_request — pi lets extensions replace the serialized provider payload right before it's sent, and provider SDKs pass unknown body fields through.
headers	writing ctx.model.headers on session_start / model_select.

before_provider_headers looks like the obvious hook for headers, but it only sees the auth/options headers — not the anthropic-beta value the Anthropic client builds internally — so setting anthropic-beta there would silently drop the Claude Code betas on OAuth accounts. Writing model.headers is the layer that wins, and because it overrides rather than appends, the extension re-adds claude-code-20250219 and oauth-2025-04-20 when ctx.modelRegistry.isUsingOAuth(model) is true. Switching to a model without a rule clears the header again.

What it costs you

/fast status tracks only the requests this extension actually patched:

Fast is on (session override) and active for anthropic/claude-opus-5: speed="fast"
— ~2x token cost. 4 fast requests: $2.10 billed, ~$1.05 of that the fast premium.
pi's own cost display is short by ~$1.05 — it does not price Anthropic's speed tier.

That last sentence is not a rounding note. pi-ai reads service_tier back off OpenAI responses and applies the multiplier, so OpenAI numbers are right. It has never heard of Anthropic's speed, so it prices fast Anthropic turns at the standard rate and pi's session cost is short by roughly the premium. This extension corrects for that using each rule's costMultiplier / costCountedByHost. Your invoice remains the source of truth.

Past warnAboveUsd (default $5 of premium) the marker turns red.

Development

npm install
npm run typecheck
npm test              # unit tests, no network
npm run test:wire     # end-to-end against a stub server (needs a pi on PATH)

npm test drives the extension through a fake of the pi API: rule matching, header juggling, the OAuth beta case, one-shot mode, and spend accounting, all offline.

npm run test:wire is the one that catches integration lies. It points the real openai provider at a local stub with pi.registerProvider("openai", { baseUrl }), runs a real pi against it, and asserts service_tier: "priority" arrives in the HTTP body — exercising the shipped default rule with no API key and no spend. Use PI_BIN=arc npm run test:wire for a differently branded pi.

Prior art

This exists because each of these solves one provider, and none of them generalize:

• @diegopetrucci/pi-claude-fast — Anthropic speed, allowlist stops at Opus 4.8.
• pi-fast — OpenAI Codex priority, hardcoded model set.
• @pi-plugins/fast-mode — OpenAI priority, configurable model list.
• pi-fast-mode — service_tier per configured provider/model pair.

License

MIT