openzoo 0.48.41 → 0.48.42

package/lib/launch.js

@@ -143,6 +143,25 @@ 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';
+  // STOP CLAUDE CODE COMPACTING ON A WINDOW THAT IS NOT REAL.
+  //
+  // It counts the WHOLE local transcript and compacts against a window it looks
+  // up from the MODEL NAME — hardcoded 200k for a recognised Claude id, which is
+  // why `/autocompact 1m` answered "capped to 200k by model" and why raising
+  // CLAUDE_CODE_MAX_CONTEXT_TOKENS alone did nothing. It never sees that the
+  // proxy binds the prefix and forwards ~8 of 166 turns.
+  //
+  // An `openzoo-` alias is the same model upstream (models.js resolves by
+  // SUFFIX, so openzoo-claude-sonnet-5 -> anthropic/claude-sonnet-5) but is not
+  // in that table, so the client falls back to the advertised context_length —
+  // which this gateway publishes as 128,000,000.
+  //
+  // OPENZOO_NO_ALIAS=1 opts out. Worth knowing: model-specific client behaviour
+  // (thinking budget, tool formats) also keys off the name, so if a session
+  // behaves oddly this is the first thing to turn off.
+  if (process.env.OPENZOO_NO_ALIAS !== '1' && !process.env.ANTHROPIC_MODEL) {
+    env.ANTHROPIC_MODEL = process.env.OPENZOO_CLAUDE_ALIAS || 'openzoo-claude-sonnet-5';
+  }
   // RAISE THE CEILING, because the proxy already removed it.
   //
   // Claude Code counts the WHOLE local transcript against the model's window and
@@ -360,6 +379,25 @@ 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';
+  // STOP CLAUDE CODE COMPACTING ON A WINDOW THAT IS NOT REAL.
+  //
+  // It counts the WHOLE local transcript and compacts against a window it looks
+  // up from the MODEL NAME — hardcoded 200k for a recognised Claude id, which is
+  // why `/autocompact 1m` answered "capped to 200k by model" and why raising
+  // CLAUDE_CODE_MAX_CONTEXT_TOKENS alone did nothing. It never sees that the
+  // proxy binds the prefix and forwards ~8 of 166 turns.
+  //
+  // An `openzoo-` alias is the same model upstream (models.js resolves by
+  // SUFFIX, so openzoo-claude-sonnet-5 -> anthropic/claude-sonnet-5) but is not
+  // in that table, so the client falls back to the advertised context_length —
+  // which this gateway publishes as 128,000,000.
+  //
+  // OPENZOO_NO_ALIAS=1 opts out. Worth knowing: model-specific client behaviour
+  // (thinking budget, tool formats) also keys off the name, so if a session
+  // behaves oddly this is the first thing to turn off.
+  if (process.env.OPENZOO_NO_ALIAS !== '1' && !process.env.ANTHROPIC_MODEL) {
+    env.ANTHROPIC_MODEL = process.env.OPENZOO_CLAUDE_ALIAS || 'openzoo-claude-sonnet-5';
+  }
   // RAISE THE CEILING, because the proxy already removed it.
   //
   // Claude Code counts the WHOLE local transcript against the model's window and

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openzoo",
-  "version": "0.48.41",
+  "version": "0.48.42",
   "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",