@lokeraar/pi-enclave-bridge 0.1.0 → 0.1.2

package/README.md

@@ -1,10 +1,70 @@
 # @lokeraar/pi-enclave-bridge
 
+[![Version: 0.1.2](https://img.shields.io/badge/version-0.1.2-blue.svg)](https://www.npmjs.com/package/@lokeraar/pi-enclave-bridge)
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
+
 EnClave provider for [Pi](https://pi.dev). The live router catalog shows up in
 `/model` with **real context windows, real per-token prices, live membership and
 the router's task aliases** — resolved from the model catalog Pi already ships,
 with no extra account required.
 
+<a href="https://github.com/Gentleman-Programming/gentle-ai">
+  <img width="220" src="https://raw.githubusercontent.com/Gentleman-Programming/gentle-ai/main/docs/assets/brand/built-with-gentle-ai.png" alt="Built with Gentle-AI" />
+</a>
+
+## 🔗 Where to find me
+
+| | |
+|---|---|
+| 📦 **npm** | [`@lokeraar/pi-enclave-bridge`](https://www.npmjs.com/package/@lokeraar/pi-enclave-bridge) |
+| 🌐 **Pi catalog** | [pi.dev/packages/@lokeraar/pi-enclave-bridge](https://pi.dev/packages/@lokeraar/pi-enclave-bridge) |
+| ⭐ **Source** | [github.com/Lokeraar/pi-EnClave-bridge](https://github.com/Lokeraar/pi-EnClave-bridge) |
+
+Install it in Pi:
+
+```bash
+pi install npm:@lokeraar/pi-enclave-bridge
+```
+
+> 💛 If this bridge ever saved you from guessing a model's limits, a ⭐ on the
+> repo goes a long way — it is the only thing that helps somebody else find it.
+> Every star is read by a human, and so is every bug report.
+
+## 📋 Releases
+
+### 0.1.2 — documentation
+
+Links in both directions, so the npm page and the source repo point at each
+other, and a note asking for a star. No behaviour change.
+
+### 0.1.1 — a clamp that left room for an actual prompt
+
+The ceiling clamp kept back 2,048 tokens for the prompt. That router counts
+`messages + tools + max_tokens` against the window, not the output alone, and a
+real Pi call carries the system prompt plus every tool schema — order 20k
+tokens. The reserve only covered a toy request.
+
+`gpt-oss-120b` exposed it: a 131,072 window with a 117,964 ceiling left 13,108
+tokens for input, below what Pi sends, so **every** call failed with `400` as
+soon as the conversation carried anything. Measured with the same prompt:
+
+```
+input ~16k   ceiling 117964   HTTP 400   "needs about 133,972 tokens"
+input ~16k   ceiling  98304   HTTP 200
+input ~28k   ceiling  98304   HTTP 200
+```
+
+The published value is a ceiling, not a fixed request: Pi reduces it per turn to
+`min(published, contextWindow − prompt − reserve)`. The reserve just had to be
+realistic.
+
+### 0.1.0 — first release
+
+Values resolved from the model catalog Pi ships, with the vendor's model card
+outranking every catalog. `/login` once and the live router catalog appears in
+`/model` with real context windows, real per-token prices, live membership and
+the router's task aliases.
+
 ## ⚡ Quick Start
 
 ```bash
@@ -116,6 +176,44 @@ They are deliberately **never** resolved from a catalog. OpenRouter has a model
 called `auto` too, advertising a 2,000,000 window — a different thing that shares
 the name, and a lie here.
 
+## 🐞 Fixes
+
+**A ceiling that the window cannot hold is impossible, not large.** OpenRouter
+listed `inkling` at 471,859 against a 262,144 window here, and the endpoint
+refused it with *"This request needs about N tokens (messages + tools +
+max_tokens)"*. Clamped to the window minus a prompt reserve.
+
+**The reserve has to survive a real conversation.** 2,048 tokens covered a toy
+request. This router counts `messages + tools + max_tokens` against the window,
+not the output alone, and a real Pi call carries the system prompt plus every
+tool schema — order 20k tokens. `gpt-oss-120b` had 13,108 tokens of input room
+and **every** call returned `400` as soon as the conversation carried anything.
+
+**`maxTokens` must never be `null`.** Pi's model list calls `.toString()` on it
+and crashes with *"Cannot read properties of undefined"*, taking the whole list
+with it.
+
+**`compat` is never inherited from a catalog.** OpenRouter ships
+`thinkingFormat: "openrouter"` plus seven other flags describing how *it* wants
+reasoning framed. This endpoint speaks the OpenAI shape — verified by sending
+`reasoning_effort` and watching what came back. Copying those flags would change
+the request format on an endpoint they were never tested against.
+
+**An accepted value is not an implemented one.** The endpoint accepts all six
+effort levels for `glm-5.3`; the vendor card says the model only implements
+low, high and max. The extras are accepted and then ignored, which is worse than
+not offering them: Pi would show a thinking level that silently does nothing.
+
+**A catalog that omits a key has not claimed anything.** An explicit `null` is a
+claim and is applied; an absent key is silence and does not erase a known value.
+Thinking maps are merged key by key, so a one-key catalog entry cannot delete a
+seven-key one.
+
+**A scoped package publishes private by default.** `npm publish` failed with
+`E402 "You must sign up for private packages"`, which reads like a billing
+problem and is not one: restricted packages need a paid plan. Declared in the
+manifest as `"publishConfig": { "access": "public" }`.
+
 ## 🔑 Authentication
 
 ```

package/enclave-live.ts

@@ -181,8 +181,14 @@ const num = (v: unknown, fallback: number) => (typeof v === "number" && v > 0 ?
 /**
  * Tokens held back from an output ceiling so the prompt has room. See the clamp
  * in buildBlock.
+ *
+ * This must cover a real Pi request, not a toy one. The router counts
+ * `messages + tools + max_tokens` against the window, so a reserve smaller than
+ * the system prompt plus every tool schema makes the model unusable in practice:
+ * gpt-oss-120b (131,072 window) shipped a 117,964 ceiling that left ~13k for
+ * input, below what Pi sends, and every call 400'd once the conversation grew.
  */
-const PROMPT_RESERVE_TOKENS = 2_048;
+export const PROMPT_RESERVE_TOKENS = 32_768;
 
 /** Aliases the router exposes: never resolved from a donor. See donors.ts. */
 function isAliasId(id: string, aliases: readonly CatalogAlias[]): boolean {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lokeraar/pi-enclave-bridge",
-  "version": "0.1.0",
+  "version": "0.1.2",
   "description": "EnClave provider bridge for Pi: /login once and the live router catalog appears in /model with real context windows, real per-token prices, live membership and the router's task aliases \u2014 resolved from the model catalog Pi already ships.",
   "type": "module",
   "main": "./index.ts",

package/scripts/test-sync.mjs

@@ -20,7 +20,7 @@ import { join } from "node:path";
 const ROOT = join(new URL(".", import.meta.url).pathname, "..");
 const { bareName, resolveModel, readBundledCatalog, readPiCatalogs, findBundledCatalogDir, findBundledCatalogs, ROUNDING_TOLERANCE } =
   await import(join(ROOT, "donors.ts"));
-const { buildBlock } = await import(join(ROOT, "enclave-live.ts"));
+const { buildBlock, PROMPT_RESERVE_TOKENS } = await import(join(ROOT, "enclave-live.ts"));
 
 let passed = 0;
 const failed = [];
@@ -214,7 +214,8 @@ console.log("\nthe block: matching, near misses, endpoint-owned fields");
   check("alias price is the catalog ceiling", by["cyberouter/auto"].cost.input === 1.4, String(by["cyberouter/auto"].cost.input));
   check("aliases carry no donor record", by["cyberouter/auto"].donor === undefined);
 
-  check("an impossible ceiling is clamped below the window", by["cyberouter/inkling"].maxTokens === 262144 - 2048, String(by["cyberouter/inkling"].maxTokens));
+  check("an impossible ceiling is clamped below the window", by["cyberouter/inkling"].maxTokens === 262144 - PROMPT_RESERVE_TOKENS, String(by["cyberouter/inkling"].maxTokens));
+  check("the clamp leaves room for a real prompt", PROMPT_RESERVE_TOKENS >= 32_768, String(PROMPT_RESERVE_TOKENS));
   check("maxTokens is never null (Pi crashes formatting it)", out.models.every((m) => typeof m.maxTokens === "number"), "hay un null");
 }