openzoo 0.9.9 → 0.10.1

package/lib/brief.js

@@ -16,7 +16,11 @@ export const BRIEF = [
   '',
   '1. CONTEXT IS EFFECTIVELY UNBOUNDED. Any model here accepts corpora far past its own attention window: bodies over ~16KB are automatically carved and bound to a holographic (HRR) memory before the model sees them, and you then query against that. `context_length` in /v1/models is the client-usable ceiling (128M tokens), not the transformer window (that is `max_model_len`). DO NOT summarise, truncate, or "chunk to fit" a corpus to preserve context — send it whole and ask your question. One POST should stay under ~9.8M tokens (~32MiB) or the edge rejects it; for more than that, bind in several calls.',
   '',
-  '2. THE BODY NEVER SHIPS TWICE. Put the corpus first, then a blank line, then your question. The corpus binds ONCE and every later question that reuses it ships only the question — near-free, and much faster. Re-pasting the same corpus each turn wastes real money. You can also bind explicitly: POST /v1/hrr/bind {"corpus": "..."} (free) returns a context_id; send it as the X-HRR-Context header on later calls.',
+  '2. THE BODY NEVER SHIPS TWICE. Put the corpus first, then a blank line, then your question. The corpus binds ONCE and every later question that reuses it ships only the question — near-free, and much faster. Re-pasting the same corpus each turn wastes real money.',
+  '',
+  '   EXPLICIT BIND (free, no payment): POST /v1/hrr/bind with {"corpus": "..."} returns {"context_id": "..."}. Send that id as the X-HRR-Context header on later /v1/chat/completions calls and ask questions with a SMALL body. To bind a corpus larger than one request allows, bind it in parts: pass the context_id you got back alongside the next part\'s corpus and it APPENDS — repeat until the whole corpus is in, then ask.',
+  '   Keep any single request under ~8MB. Bigger bodies are dropped by the network hop before they reach the proxy (an opaque 413 or a dead connection). This is a REQUEST size limit, not a context limit — the bound context can be far larger, which is what parts are for.',
+  '   Paths: your base_url already ends in /v1, so post to {base_url}/hrr/bind — NOT {base_url}/v1/hrr/bind (that double /v1 404s; the proxy repairs it, but do not rely on that).',
   '',
   '3. PAYMENT IS HANDLED. Every call is paid per-request from the operator\'s wallet via x402 (Solana / Base / Robinhood Chain, whichever is funded). There is no API key to find, no account, no rate limit to negotiate. Never search the operator\'s machine for credentials or endpoints — GET / on this proxy returns everything discoverable about it.',
   '',

package/lib/proxy.js

@@ -15,11 +15,31 @@ import { injectBrief } from './brief.js';
 const HOP_BY_HOP = new Set([
   'host', 'connection', 'keep-alive', 'transfer-encoding', 'upgrade',
   'proxy-authorization', 'proxy-connection', 'te', 'trailer', 'content-length',
+  // `Expect: 100-continue` is sent by curl and most HTTP libraries once a body
+  // passes ~1KB. undici REFUSES it outright ("expect header not supported"),
+  // so forwarding it made every LARGE-body request fail while small ones
+  // worked — i.e. it broke exactly the corpus calls this proxy exists for.
+  'expect',
   // The harness's api key (sk-openzoo or anything) is accepted and dropped:
   // the zoo takes payment, not keys.
   'authorization',
 ]);
 
+/**
+ * Be forgiving about the path, the same way we are about model ids.
+ * Harnesses are configured with a base_url that ALREADY ends in /v1, so an
+ * agent building "{base}/v1/hrr/bind" sends /v1/v1/hrr/bind and gets a 404 it
+ * cannot diagnose (observed: agent concluded the bind endpoint "is not
+ * functioning as advertised" and fell back to stuffing the corpus inline).
+ * Collapse repeated /v1 and add a missing one.
+ */
+function normalizePath(url) {
+  const [path, query] = (url || '/').split(/(?=\?)/);
+  let p = path.replace(/^(?:\/v1)+(?=\/v1\/)/, '');       // /v1/v1/x -> /v1/x
+  if (!/^\/v1(\/|$)/.test(p) && /^\/(hrr|chat|models|completions|embeddings|usage)/.test(p)) p = `/v1${p}`;
+  return p === path ? url : `${p}${query || ''}`;
+}
+
 function upstreamHeaders(req) {
   const out = {};
   for (const [k, v] of Object.entries(req.headers)) {
@@ -251,6 +271,11 @@ export async function startProxy({ silent = false, requireToken = null, sessionM
   let tunnelGate = null;
 
   const server = http.createServer(async (req, res) => {
+    const normalized = normalizePath(req.url);
+    if (normalized !== req.url) {
+      log(`path ${req.url} -> ${normalized} (base_url already ends in /v1)`);
+      req.url = normalized;
+    }
     const url = `${config.apiBase}${req.url}`;
     // Requests that arrived over the public quick-tunnel URL carry cloudflared's
     // headers; nothing dialing 127.0.0.1 directly does. That distinction is what
@@ -454,7 +479,14 @@ export async function startProxy({ silent = false, requireToken = null, sessionM
         log(err.message);
         jsonErr(res, 402, err.message);
       } else {
-        jsonErr(res, 502, `openzoo proxy error: ${err.message}`);
+        // "fetch failed" alone is undiagnosable — undici hides the real
+        // network error in `cause`. Surface it (and log the stack) or every
+        // transport hiccup looks identical to a payment bug.
+        const cause = err.cause?.message || err.cause?.code || err.cause;
+        const detail = cause ? `${err.message} (${cause})` : err.message;
+        log(`proxy error: ${detail}`);
+        if (process.env.OPENZOO_DEBUG) console.error(err.stack);
+        jsonErr(res, 502, `openzoo proxy error: ${detail}`);
       }
     }
   });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openzoo",
-  "version": "0.9.9",
+  "version": "0.10.1",
   "description": "Local x402-paying proxy + MCP server for openzoo.fun — point any OpenAI-compatible harness (Cursor, Claude Code, aider, SDKs) at localhost and it pays per call from a local burner wallet. Solana and Base rails live; Robinhood experimental.",
   "license": "MIT",
   "type": "module",