@velocirouter/pi 1.0.1 → 1.1.0

package/README.md

@@ -4,13 +4,10 @@
 It registers VelociRouter as a provider and loads the model catalog from the
 API.
 
-Requires **Pi 0.85.1 or later** from `@earendil-works/pi-coding-agent`.
-The older `@mariozechner/pi-coding-agent` distribution does not support this
-package's catalog refresh. Upgrade Pi before installing the package:
+Requires **Pi 0.86.1 or later** from `@earendil-works/pi-coding-agent`.
+If you already have a supported version, install the package:
 
 ```bash
-npm uninstall -g @mariozechner/pi-coding-agent
-npm install -g @earendil-works/pi-coding-agent@latest
 pi install npm:@velocirouter/pi
 ```
 
@@ -18,6 +15,27 @@ Then run `/login` inside pi, pick **VelociRouter**, and paste your `rt-…` key.
 Pi refreshes the catalog right after login, so the models show up in `/model`
 straight away.
 
+## Install or update Pi (if needed)
+
+If Pi is not installed, or your version is below 0.86.1, install the latest
+version, then run the package installation command above:
+
+```bash
+npm install -g @earendil-works/pi-coding-agent@latest
+```
+
+### Migrating from the legacy npm package
+
+Only if you installed Pi through `@mariozechner/pi-coding-agent`, replace it
+with the current distribution using the commands below. The legacy distribution
+does not support this package's catalog refresh. Then run the package
+installation command above.
+
+```bash
+npm uninstall -g @mariozechner/pi-coding-agent
+npm install -g @earendil-works/pi-coding-agent@latest
+```
+
 ## What it registers
 
 - **Provider** — `velocirouter`, an `openai-completions` endpoint.
@@ -30,8 +48,11 @@ straight away.
   tokens pass it, and those tokens are the next request's input, so the
   combined figure would let a long session overshoot the input ceiling.
 
-Costs are reported as zero: VelociRouter bills your account directly, so pi's
-per-session cost estimate would double-count.
+Session costs use the billed USD total returned by VelociRouter in streaming
+`usage.cost`. Enable **Cost in response** for your user in VelociRouter first.
+This is reporting only; it does not add another charge. If the server omits cost,
+the default model prices remain zero. Per-category cost breakdowns are not
+available, and earlier turns recorded with zero cost are not updated.
 
 The catalog is fetched when an interactive session starts, after `/login`, and
 from the `/model` picker. `pi -p` and `pi --list-models` read the cached copy,
@@ -44,6 +65,19 @@ so open pi interactively once before using them.
 | `VELOCIROUTER_API_KEY` | Used when no key was stored by `/login`. Set this in CI. |
 | `VELOCIROUTER_BASE_URL` | Override the API host (default `https://api.velocirouter.site`). |
 
+## Development checks
+
+Run `npm test` with the Pi AI peer dependency available. To also check host
+module resolution and streaming through Pi's actual extension loader, run:
+
+```bash
+JITI_FS_CACHE=0 PI_TEST_INSTALL_DIR="$(npm root -g)/@earendil-works/pi-coding-agent" \
+  node --test loader.test.js
+```
+
+The loader check uses the host Pi installation and needs no local Pi AI install.
+It is skipped when `PI_TEST_INSTALL_DIR` is unset.
+
 ## Publishing
 
 ```bash

package/cost-stream.js

@@ -0,0 +1,71 @@
+import { createAssistantMessageEventStream, openAICompletionsApi } from "@earendil-works/pi-ai/compat"
+
+const completions = openAICompletionsApi()
+
+// Observe bytes inline: cloning the response would buffer a second copy and
+// could finish reading usage after Pi has already emitted its final message.
+function observeCost(body, onCost) {
+  const decoder = new TextDecoder()
+  let pending = ""
+  function consume(text) {
+    pending += text
+    let boundary
+    while ((boundary = /\r\n\r\n|\n\n|\r\r/.exec(pending))) {
+      const event = pending.slice(0, boundary.index)
+      pending = pending.slice(boundary.index + boundary[0].length)
+      const data = event.split(/\r\n|\r|\n/)
+        .filter((line) => line.startsWith("data:"))
+        .map((line) => line.slice(5).replace(/^ /, ""))
+        .join("\n")
+      try {
+        const cost = JSON.parse(data)?.usage?.cost
+        if (typeof cost === "number" && Number.isFinite(cost) && cost >= 0) onCost(cost)
+      } catch {
+        // Non-JSON events (including [DONE]) remain the SDK's responsibility.
+      }
+    }
+  }
+  return body.pipeThrough(new TransformStream({
+    transform(chunk, controller) {
+      consume(decoder.decode(chunk, { stream: true }))
+      controller.enqueue(chunk)
+    },
+    flush() {
+      consume(decoder.decode() + "\n\n")
+    },
+  }))
+}
+
+export function streamWithCost(model, context, options = {}) {
+  const output = createAssistantMessageEventStream()
+  const upstreamFetch = options.fetch ?? globalThis.fetch
+  let billedCost
+  const fetch = async (...args) => {
+    billedCost = undefined
+    const response = await upstreamFetch(...args)
+    if (!response.ok || !response.body || !response.headers.get("content-type")?.includes("text/event-stream")) {
+      return response
+    }
+    return new Response(observeCost(response.body, (cost) => { billedCost = cost }), {
+      status: response.status,
+      statusText: response.statusText,
+      headers: response.headers,
+    })
+  }
+  // The built-in adapter retains ownership of requests, retries, tools, and errors.
+  const upstream = completions.streamSimple(model, context, { ...options, fetch })
+  void (async () => {
+    for await (const event of upstream) {
+      if (event.type === "done" || event.type === "error") {
+        const message = event.type === "done" ? event.message : event.error
+        if (billedCost !== undefined) {
+          // The server supplies a total, not a per-token-category breakdown.
+          message.usage.cost = { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: billedCost }
+        }
+      }
+      output.push(event)
+    }
+    output.end()
+  })()
+  return output
+}

package/index.js

@@ -1,3 +1,5 @@
+import { streamWithCost } from "./cost-stream.js"
+
 const PROVIDER = "velocirouter"
 const API = "openai-completions"
 const DEFAULT_BASE_URL = "https://api.velocirouter.site"
@@ -87,5 +89,6 @@ export default function velocirouter(pi) {
     // app without these.
     headers: { "HTTP-Referer": "https://pi.dev", "X-Title": "Pi" },
     refreshModels,
+    streamSimple: streamWithCost,
   })
 }

package/package.json

@@ -1,16 +1,23 @@
 {
   "name": "@velocirouter/pi",
-  "version": "1.0.1",
+  "version": "1.1.0",
   "description": "Pi coding agent package for VelociRouter — registers the provider and its model catalog.",
   "type": "module",
   "main": "index.js",
   "license": "MIT",
   "files": [
     "index.js",
+    "cost-stream.js",
     "README.md"
   ],
+  "scripts": {
+    "test": "node --test"
+  },
+  "peerDependencies": {
+    "@earendil-works/pi-ai": ">=0.86.1"
+  },
   "engines": {
-    "node": ">=18"
+    "node": ">=22.19.0"
   },
   "pi": {
     "extensions": [