osborn 0.9.128 → 0.9.129

package/dist/claude-llm.js

@@ -483,7 +483,12 @@ export class ClaudeLLM extends llm.LLM {
         return 'claude.agent-sdk';
     }
     get model() {
-        return this.#opts.model || 'claude-opus-4-8'; // Opus 4.8 orchestrator with named sub-agents
+        // The [1m] suffix opts into Opus 4.8's 1M context window (Claude Code's
+        // context-1m-2025-08-07 beta). WITHOUT it the SDK runs opus at its 200k
+        // base, so auto-compaction fired at ~153k every ~10 min (confirmed in the
+        // live agent log: compact_boundary pre_tokens≈153k). With [1m] the window
+        // is 1M and compaction happens far later. Overridable via opts.model.
+        return this.#opts.model || 'claude-opus-4-8[1m]';
     }
     get sessionId() {
         return this.#sessionId;
@@ -766,11 +771,14 @@ export class ClaudeLLM extends llm.LLM {
     pushMessage(userText, sdkOptions, callbacks) {
         // Auto-compact threshold — fires PreCompact when context fills to this %.
         // Higher = uses more of the window before compacting (more context retained
-        // per turn, but less headroom for the next reply). Raised 85→92 because
-        // users reported compaction firing too often; 92% pushes it as late as is
-        // safe before Claude starts hard-capping replies near the 1M limit. Pairs
-        // with autoCompactWindow=1_000_000 (settings.json, read via settingSources).
+        // per turn, but less headroom for the next reply). 92% pushes it as late as
+        // is safe before Claude starts hard-capping replies near the 1M limit.
         process.env.CLAUDE_AUTOCOMPACT_PCT_OVERRIDE = '92';
+        // Compaction WINDOW in thousands of tokens (min(actual_context, this)). Now
+        // that the model runs with [1m] (1M context), set this to 1000 (=1,000,000)
+        // so the compaction budget matches the real window instead of being clamped
+        // to opus' 200k base — the actual cause of the ~153k early compaction.
+        process.env.CLAUDE_CODE_AUTO_COMPACT_WINDOW = '1000';
         const userMessage = {
             type: 'user',
             message: { role: 'user', content: [{ type: 'text', text: userText }] },
@@ -1004,7 +1012,7 @@ class ClaudeLLMStream extends llm.LLMStream {
                 permissionMode: this.#opts.permissionMode,
                 allowedTools,
                 // model: this.#opts.model || 'haiku', // haiku for speed with limited tools, sonnet for full research capabilities (including tool use trace in response)
-                model: this.#opts.model || 'claude-opus-4-8', // Opus 4.8 orchestrator with named sub-agents (Haiku tested but ignored delegation rules)
+                model: this.#opts.model || 'claude-opus-4-8[1m]', // Opus 4.8 + [1m] → 1M context (see get model() note); prevents ~153k early compaction
                 enableFileCheckpointing: true,
                 settingSources: ['project', 'user'],
                 extraArgs: { 'replay-user-messages': null },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.128",
+  "version": "0.9.129",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {