osborn 0.9.128 → 0.9.130

package/dist/claude-llm.js

@@ -149,6 +149,16 @@ function ensureCompactionSettings() {
     }
 }
 ensureCompactionSettings();
+// Enable Opus 4.8's native 1M context window + set the auto-compact threshold
+// at MODULE LOAD — BEFORE any query() spawns — so they actually apply to the
+// persistent session. Setting these per-message in pushMessage() is too late:
+// the SDK subprocess reads env once at cold-start, so later writes never reach
+// the running query — the reason auto-compaction kept firing at ~150k every
+// turn. ENABLE_1M_CONTEXT is Claude Code's documented switch for the context-1m
+// (1M) window; without it opus runs at its 200K default. NOTE: 1M activation
+// also depends on account entitlement (auto on Team seats; else usage credits).
+process.env.ENABLE_1M_CONTEXT = '1';
+process.env.CLAUDE_AUTOCOMPACT_PCT_OVERRIDE = '92';
 // Research mode tools — full research capabilities
 // Named sub-agents — the orchestrator delegates to these specialists. Each has
 // a specific role, model, and tool set. Module-level + exported so the HTTP
@@ -483,7 +493,12 @@ 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;
@@ -764,13 +779,9 @@ export class ClaudeLLM extends llm.LLM {
      * @param callbacks - Event callbacks for the background consumer
      */
     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).
-        process.env.CLAUDE_AUTOCOMPACT_PCT_OVERRIDE = '92';
+        // (Compaction/1M env is set once at module load — see ENABLE_1M_CONTEXT /
+        // CLAUDE_AUTOCOMPACT_PCT_OVERRIDE near ensureCompactionSettings(). Setting it
+        // here per-message was too late for the persistent query.)
         const userMessage = {
             type: 'user',
             message: { role: 'user', content: [{ type: 'text', text: userText }] },
@@ -1004,7 +1015,7 @@ 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": "0.9.130",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {