osborn 0.9.129 → 0.9.131

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
@@ -769,16 +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). 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';
+        // (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 }] },

package/dist/index.js

@@ -2214,7 +2214,7 @@ async function main() {
                 const { readSessionHistory } = await import('./session-access.js');
                 const history = readSessionHistory(sessionId, workingDir, {
                     lastN: 10,
-                    types: ['assistant'],
+                    types: ['assistant', 'tool_use'], // tool_use entries also carry Claude's spoken text
                 });
                 recentMessages = history
                     .filter((m) => m.text)
@@ -2230,8 +2230,14 @@ async function main() {
         // BEFORE the previous TTS completed, and we may have already suppressed in-flight
         // tts_say events that arrived during that overlap).
         const carriedSuppressed = lastInterruption?.suppressedText ?? '';
-        lastInterruption = { spokenText: fullText, recentMessages, suppressedText: carriedSuppressed, timestamp: Date.now() };
-        console.log(`📋 Interruption context stored (text: ${fullText.length} chars, JSONL: ${recentMessages.length} chars, suppressed carried: ${carriedSuppressed.length} chars)`);
+        lastInterruption = {
+            spokenText: fullText,
+            recentMessages,
+            suppressedText: carriedSuppressed,
+            fullBlockText: fullBlockText && fullBlockText !== fullText ? fullBlockText : undefined,
+            timestamp: Date.now(),
+        };
+        console.log(`📋 Interruption context stored (text: ${fullText.length} chars, JSONL: ${recentMessages.length} chars, block: ${fullBlockText?.length ?? 0} chars, suppressed carried: ${carriedSuppressed.length} chars)`);
     }
     /**
      * Append text the agent tried to say while the user was speaking, but which we
@@ -2271,6 +2277,7 @@ async function main() {
             spokenText: lastInterruption.spokenText,
             recentMessages: lastInterruption.recentMessages,
             suppressedText: lastInterruption.suppressedText,
+            fullBlockText: lastInterruption.fullBlockText,
         };
         lastInterruption = null;
         return ctx;

package/dist/pipeline-direct-llm.d.ts

@@ -21,6 +21,8 @@ export interface InterruptionContext {
      * and can re-articulate the relevant bits in its next response.
      */
     suppressedText: string;
+    /** Full text of the interrupted TTS block — fallback when JSONL read returns empty */
+    fullBlockText?: string;
 }
 export interface PipelineDirectOptions extends ClaudeLLMOptions {
     onFastBrainResult?: (result: FastBrainPanelResult) => void;

package/dist/pipeline-direct-llm.js

@@ -122,7 +122,10 @@ export class PipelineDirectLLM extends llm.LLM {
                     `Anything in "Your recent messages" below that appears AFTER the quoted heard text is content the user did not hear. The user has no memory of it.`,
                     ``,
                     `Your recent messages (full untruncated — you wrote these):`,
-                    interruptCtx.recentMessages || '(no recent messages found)',
+                    interruptCtx.recentMessages
+                        || (interruptCtx.fullBlockText
+                            ? `(JSONL unavailable — full TTS block text follows)\n"${interruptCtx.fullBlockText}"`
+                            : '(no recent messages found)'),
                     suppressedBlock,
                     ``,
                     `User's message: "${userText}"`,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.129",
+  "version": "0.9.131",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {