openzoo 0.48.39 → 0.48.41

package/lib/launch.js

@@ -143,6 +143,20 @@ export async function launchClaude(argv) {
   delete env.ANTHROPIC_API_KEY;
   env.ANTHROPIC_BASE_URL = base;
   env.ANTHROPIC_AUTH_TOKEN = process.env.ANTHROPIC_AUTH_TOKEN || 'sk-openzoo';
+  // RAISE THE CEILING, because the proxy already removed it.
+  //
+  // Claude Code counts the WHOLE local transcript against the model's window and
+  // hard-stops with "Context limit reached - /compact or /clear to continue". It
+  // cannot know the proxy binds the old prefix into leCore and forwards only the
+  // system block plus a bounded tail - MEASURED live, 8 of 166 turns actually
+  // sent. Disabling auto-compact WITHOUT raising this made it worse: it used to
+  // compact and carry on, and instead it just stopped.
+  //
+  // Safe only because spillTranscript is real: OPENZOO_TAIL_MAX_CHARS bounds what
+  // leaves this machine however long the conversation gets.
+  if (process.env.OPENZOO_KEEP_COMPACT !== '1' && !process.env.CLAUDE_CODE_MAX_CONTEXT_TOKENS) {
+    env.CLAUDE_CODE_MAX_CONTEXT_TOKENS = process.env.OPENZOO_CLAUDE_CONTEXT_TOKENS || '1000000';
+  }
 
   if (terminal) {
     const cli = resolveClaudeCli();
@@ -209,7 +223,15 @@ export async function launchClaude(argv) {
       // because OPENZOO_KEEP_TAIL_MSGS bounds what is actually sent.
       // OPENZOO_KEEP_COMPACT=1 restores stock behaviour.
       const prevCompact = settings.autoCompactEnabled;
-      if (process.env.OPENZOO_KEEP_COMPACT !== '1') settings.autoCompactEnabled = false;
+      // DO NOT touch autoCompactEnabled. Disabling it here mutated the user's
+      // GLOBAL ~/.claude/settings.json and applied to every session, openzoo or
+      // not — and because Claude Code's HARD ceiling is separate from its
+      // AUTOMATIC compaction, it removed the escape hatch without removing the
+      // wall. A live session ended up unable to compact AND unable to continue:
+      // every keystroke returned "Context limit reached". Strictly worse than
+      // stock. The correct lever is CLAUDE_CODE_MAX_CONTEXT_TOKENS on the
+      // spawned process (set below), which is scoped to this launch and leaves
+      // compaction available as a fallback.
       fs.mkdirSync(path.dirname(settingsPath), { recursive: true });
       fs.writeFileSync(settingsPath, `${JSON.stringify(settings, null, 2)}\n`);
       restoreStatus = () => {
@@ -217,6 +239,7 @@ export async function launchClaude(argv) {
           const cur = JSON.parse(fs.readFileSync(settingsPath, 'utf8'));
           if (prev === undefined) delete cur.statusLine; else cur.statusLine = prev;
           // Never leave a global setting mutated after we exit.
+          // Clean up the flag older builds left behind in global settings.
           if (prevCompact === undefined) delete cur.autoCompactEnabled;
           else cur.autoCompactEnabled = prevCompact;
           fs.writeFileSync(settingsPath, `${JSON.stringify(cur, null, 2)}\n`);
@@ -337,6 +360,20 @@ export async function launchHarness(cmd, args) {
   delete env.ANTHROPIC_API_KEY;   // conflicts with the gateway auth-token path
   env.ANTHROPIC_BASE_URL = base;
   env.ANTHROPIC_AUTH_TOKEN = process.env.ANTHROPIC_AUTH_TOKEN || 'sk-openzoo';
+  // RAISE THE CEILING, because the proxy already removed it.
+  //
+  // Claude Code counts the WHOLE local transcript against the model's window and
+  // hard-stops with "Context limit reached - /compact or /clear to continue". It
+  // cannot know the proxy binds the old prefix into leCore and forwards only the
+  // system block plus a bounded tail - MEASURED live, 8 of 166 turns actually
+  // sent. Disabling auto-compact WITHOUT raising this made it worse: it used to
+  // compact and carry on, and instead it just stopped.
+  //
+  // Safe only because spillTranscript is real: OPENZOO_TAIL_MAX_CHARS bounds what
+  // leaves this machine however long the conversation gets.
+  if (process.env.OPENZOO_KEEP_COMPACT !== '1' && !process.env.CLAUDE_CODE_MAX_CONTEXT_TOKENS) {
+    env.CLAUDE_CODE_MAX_CONTEXT_TOKENS = process.env.OPENZOO_CLAUDE_CONTEXT_TOKENS || '1000000';
+  }
   console.error(`openzoo: launching \`${cmd}\` on the zoo (ANTHROPIC_BASE_URL=${base}) — every turn pays x402`);
   const child = spawn(cmd, args, { stdio: 'inherit', env });
   child.on('exit', (code) => process.exit(code ?? 0));

package/lib/proxy.js

@@ -1111,7 +1111,12 @@ export async function startProxy({ silent = false, requireToken = null, sessionM
             // promptTokens, so a tail that rivals the corpus silently falls
             // back to markup and direct collapses onto billed.
             const lc = x.lecore || {};
-            log(`spill priced: ${x.pricing} · corpus ${lc.corpusTokens ?? '?'} tok vs sent ${lc.tokensBefore ?? '?'} -> ${lc.tokensAfter ?? '?'} · billed ${(x.billedUsd ?? 0).toFixed(5)} direct ${(x.directUsd ?? 0).toFixed(5)}`);
+            // counterfactualTokensUsed is the basis the gateway ACTUALLY priced
+            // on. lecore.corpusTokens is often absent and reading it printed
+            // 'corpus ?' on calls that were pricing fine — which sent a whole
+            // night's debugging after a number that was never the input.
+            const basis = x.counterfactualTokensUsed ?? lc.corpusTokens;
+            log(`spill priced: ${x.pricing} · basis ${basis ?? '?'} tok vs sent ${lc.tokensBefore ?? '?'} -> ${lc.tokensAfter ?? '?'} · billed ${(x.billedUsd ?? 0).toFixed(5)} direct ${(x.directUsd ?? 0).toFixed(5)}`);
             spillSpend += x.billedUsd;
             spillDirect += typeof x.directUsd === 'number' ? x.directUsd : x.billedUsd;
           }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openzoo",
-  "version": "0.48.39",
+  "version": "0.48.41",
   "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",