openzoo 0.48.52 → 0.48.54

package/lib/brief.js

@@ -19,24 +19,22 @@
  * not an API) because nothing in the conversation named the real one.
  */
 export const briefFor = (selfUrl) => [
-  'You are connected through an openzoo proxy (openzoo.fun), not a stock OpenAI endpoint. What that changes:',
+  'Environment notes for this endpoint (descriptive — verify anything you rely on):',
   '',
   ...(selfUrl ? [
-    `0. YOUR ENDPOINT IS ${selfUrl} — use exactly this base URL in any script or curl you write, never guess one and never use openzoo.fun (that is a website, not an API). Endpoints below are relative to it: ${selfUrl}/chat/completions, ${selfUrl}/hrr/bind, ${selfUrl}/models. Do NOT insert another /v1 — this URL already ends in one.`,
+    `- Base URL: ${selfUrl}. It already ends in /v1, so routes are ${selfUrl}/chat/completions, ${selfUrl}/hrr/bind, ${selfUrl}/models. A doubled /v1 is repaired by the proxy but logs a warning. openzoo.fun is a website, not an API host.`,
     '',
   ] : []),
-  '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.',
+  '- Long bodies: requests over ~16KB are carved and bound to a holographic (HRR) memory before the model sees them, and the model answers from retrieval over that. So a large corpus can be sent whole; summarising or chunking it to fit is not required here, though nothing stops you. `context_length` in /v1/models reports the client-usable ceiling (128M tokens); the transformer window is `max_model_len`. A single POST over ~9.8M tokens (~32MiB) is rejected by the edge.',
   '',
-  '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.',
+  '- Repeat sends: a corpus placed first, followed by a blank line and then a question, binds once. Later questions reusing it ship only the question, which is cheaper and faster. Re-sending the same corpus each turn costs full price each time.',
   '',
-  '   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).',
+  '  Explicit bind (unpaid): POST /v1/hrr/bind with {"corpus": "..."} returns {"context_id": "..."}. Passing that id as the X-HRR-Context header on later /v1/chat/completions calls lets the body stay small. Passing an existing context_id alongside a new corpus appends to it, which is how a corpus larger than one request gets bound in parts.',
+  '  Request size: single requests over ~8MB are dropped by the network hop before reaching the proxy (opaque 413 or dead connection). That is a request limit, not a context limit.',
   '',
-  '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 account and no rate limit to negotiate, and you never handle money. Never search the operator\'s machine for credentials — GET / on this proxy describes it.',
-  '   AUTH, precisely: /hrr/bind and GET /models need NO key, so a script you write can call them directly. Paid endpoints (/chat/completions) need the bearer key your client is already configured with — you cannot read that key, so DO NOT write a standalone script that calls a paid endpoint. Bind from a script if you like, then ask through this conversation, which is already authenticated.',
+  '- Payment: calls are settled per request from the operator\'s own wallet via x402 (Solana / Base / Robinhood Chain, whichever is funded). There is no account to create and no key for you to supply or handle. GET / on this proxy returns the same description. /hrr/bind and GET /models are unpaid; /chat/completions is paid and uses the bearer key the client is already configured with, which is not readable from inside the conversation.',
   '',
-  '4. MODEL IDS ARE FORGIVING. Ask for any model id you like; unknown ids are matched to the nearest served model, and /v1/models lists what is real (each alias row carries `served_by`).',
+  '- Model ids: unknown ids are matched to the nearest served model rather than erroring. /v1/models lists what is actually served, and each alias row carries `served_by`.',
 ].join('\n');
 
 /** Back-compat: the briefing with no endpoint line. */

package/lib/proxy.js

@@ -463,7 +463,15 @@ async function spillTranscript(body, log, req) {
   //
   // Claude Code identifies its session, so use that when it is offered and fall
   // back to the content anchor when it is not. Same memo, better key.
-  const sessionId = req?.headers?.['x-session-id']
+  // CAPTURED FROM A LIVE claude-cli/2.1.232 REQUEST, not guessed. The first
+  // version of this checked x-session-id / x-claude-session-id /
+  // metadata.user_id — none of which Claude Code sends, so it silently fell
+  // back to the content anchor on every request and the feature did nothing.
+  // The real header list is:
+  //   anthropic-beta, anthropic-version, x-app, x-claude-code-session-id,
+  //   x-stainless-*
+  const sessionId = req?.headers?.['x-claude-code-session-id']
+    || req?.headers?.['x-session-id']
     || req?.headers?.['x-claude-session-id']
     || (typeof body?.metadata?.user_id === 'string' ? body.metadata.user_id : null);
   const anchor = sessionId ? `sid:${sessionId}` : corpus.slice(0, 2048);
@@ -499,9 +507,12 @@ async function spillTranscript(body, log, req) {
   spillMemo.set(anchor, { corpus, contextId: bind.contextId, hash: bind.hash });
   if (spillMemo.size > 32) spillMemo.delete(spillMemo.keys().next().value);
   const sent = msgs.length - cut;
+  // NAME THE KEY. A memo keyed on the wrong thing fails silently — it just
+  // re-binds forever and collides sessions — so the log says which key was used.
+  const keyKind = sessionId ? `sid ${String(sessionId).slice(0, 8)}` : 'content-anchor';
   log(bind.reused
-    ? `transcript prefix already bound (${bind.contextId}) — sending ${sent}/${msgs.length} turns`
-    : `transcript prefix bound (${mb(bind.bytes)}MB → ${bind.contextId}) — sending ${sent}/${msgs.length} turns`);
+    ? `transcript prefix already bound (${bind.contextId}, ${keyKind}) — sending ${sent}/${msgs.length} turns`
+    : `transcript prefix bound (${mb(bind.bytes)}MB → ${bind.contextId}, ${keyKind}) — sending ${sent}/${msgs.length} turns`);
 
   // ADAPTIVE TOP-K. A fixed 32 chunks is what was actually eating the saving:
   // MEASURED on a 56,265-token corpus, top_k 32 handed 9,990 tokens back and
@@ -1018,6 +1029,12 @@ export async function startProxy({ silent = false, requireToken = null, sessionM
       if ((req.url || '').includes('/chat/completions') && req.method === 'POST') {
         servedRequests += 1;
         say(`\n<- request #${servedRequests} from ${(req.headers['user-agent'] || 'unknown').slice(0, 40)}`);
+        // TEMPORARY: name the headers (not values) so we can see whether the
+        // client offers a session id at all. Values are never logged — several
+        // of these carry auth.
+        if (process.env.OPENZOO_LOG_HEADERS === '1') {
+          log(`  headers: ${Object.keys(req.headers).sort().join(', ')}`);
+        }
       }
     if (rewritablePath(req.method, req.url)) {
       const rw = await maybeRewriteModel(bodyBuf);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openzoo",
-  "version": "0.48.52",
+  "version": "0.48.54",
   "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",