osborn 0.9.37 → 0.9.39

package/Dockerfile.sandbox

@@ -31,10 +31,23 @@ RUN npm install -g "osborn@${OSBORN_VERSION}" @anthropic-ai/claude-code
 # Persistent workspace + claude config dirs
 RUN mkdir -p /workspace /root/.claude
 
+# Marker so orchestration (machines.ts isManifestAware) can detect this image
+# supports the manifest-driven update flow. Pre-marker machines fall back to
+# the image-swap update path, which brings them onto a marker-aware image;
+# from then on, all updates use the manifest flow defined in the entrypoint.
+RUN touch /etc/osborn-manifest-aware
+
 ENV OSBORN_CWD=/workspace
 ENV OSBORN_API_PORT=8741
 ENV NODE_ENV=production
 
+# HOME points at the volume so user-space config from any tool that respects
+# HOME (gh, git, ssh, aws, etc.) automatically writes to the persistent
+# volume instead of the ephemeral container overlay. The existing /root/.claude
+# symlink machinery below stays in place — it's redundant with HOME=/workspace
+# but harmless.
+ENV HOME=/workspace
+
 WORKDIR /workspace
 
 EXPOSE 8741
@@ -52,6 +65,10 @@ ln -sf /workspace/.claude /root/.claude
 # Suppress Claude Code interactive onboarding prompts
 ONBOARDING_JSON='{"numStartups":10,"installMethod":"npm","autoUpdates":false,"hasCompletedOnboarding":true,"hasTrustDialogAccepted":true,"hasTrustDialogHooksAccepted":true,"hasCompletedProjectOnboarding":true,"hasAcknowledgedCostThreshold":true,"effortCalloutV2Dismissed":true,"theme":"dark","projects":{"/workspace":{"hasTrustDialogAccepted":true,"hasTrustDialogHooksAccepted":true,"hasCompletedProjectOnboarding":true}}}'
 echo "$ONBOARDING_JSON" > /root/.claude.json
+# Additional write at $HOME/.claude.json. With HOME=/workspace this is where
+# Claude Code actually reads its top-level config from; the /root/.claude.json
+# write above becomes dead but is left in place (harmless).
+echo "$ONBOARDING_JSON" > /workspace/.claude.json
 mkdir -p /workspace/.claude
 echo "$ONBOARDING_JSON" > /workspace/.claude/.config.json
 echo "$ONBOARDING_JSON" > /workspace/.claude/claude.json
@@ -79,6 +96,23 @@ if [ -d "$PKG_SKILLS_DIR" ]; then
   done
 fi
 
+# Manifest-driven version check.
+# Orchestration writes /workspace/.osborn-want-version on update (machines.ts
+# updateViaManifest). On every boot we compare to the currently-installed
+# osborn and run `npm install -g osborn@<want>` if they differ. The install
+# lands in the container overlay (default npm prefix) — Fly wipes overlay on
+# stop/start so the install re-runs on every boot until the base image is
+# rebuilt with that version baked in. Update is fast between Fly restarts;
+# only the first boot after a restart pays the npm install cost.
+WANT=$(cat /workspace/.osborn-want-version 2>/dev/null | tr -d '[:space:]')
+if [ -n "$WANT" ]; then
+  CURRENT=$(osborn --version 2>/dev/null | head -1 | tr -d '[:space:]')
+  if [ "$WANT" != "$CURRENT" ]; then
+    echo "[sandbox] osborn ${CURRENT:-none} → ${WANT} (manifest install)"
+    npm install -g "osborn@${WANT}" || echo "[sandbox] install failed — running ${CURRENT:-image-baked} version"
+  fi
+fi
+
 exec osborn
 ENTRYPOINT
 

package/dist/index.js

@@ -1,7 +1,7 @@
 // Load environment variables FIRST before any other imports
 import 'dotenv/config';
 import { voice, initializeLogger } from '@livekit/agents';
-import { Room, RoomEvent, RemoteParticipant } from '@livekit/rtc-node';
+import { Room, RoomEvent } from '@livekit/rtc-node';
 import { AccessToken } from 'livekit-server-sdk';
 // Initialize logger before anything else
 initializeLogger({ pretty: true, level: 'info' });
@@ -2530,51 +2530,16 @@ async function main() {
         console.log('✅ Connected to room:', roomName);
         localParticipant = room.localParticipant;
     });
-    // EARLIEST possible "user is speaking" signal in our setup. Driven by LiveKit's
-    // server-side audio-level VAD on the participant's WebRTC track — fires ~50-100ms
-    // after mic onset, independent of Deepgram STT or any local VAD (we don't run one).
-    //
-    // Flow: user starts talking → ActiveSpeakersChanged includes a RemoteParticipant →
-    // if agent is currently speaking → interrupt the SpeechHandle to flush TTS playback.
-    // The existing handleSpeechDone callback (around line 1320) captures the spoken-text
-    // + JSONL context into lastInterruption; PipelineDirectLLM consumes it on the next
-    // chat() call to enrich the user's message with [INTERRUPTED] context — so the
-    // post-interrupt note flow is preserved even though we're cutting TTS earlier.
-    //
-    // Filter is `instanceof RemoteParticipant`. The agent IS the LocalParticipant in this
-    // room, and when its TTS plays it appears in the active-speakers list too. An earlier
-    // attempt that compared `s.identity !== room.localParticipant?.identity` failed because
-    // localParticipant.identity could be undefined at event-fire time, letting the agent's
-    // own speech trigger a self-interrupt. The type check is bulletproof.
-    //
-    // Realtime mode skipped — the SDK handles interruption internally there, and manual
-    // interrupt for Gemini realtime crashes its state machine (code 1008, memory v0.4.5).
-    let lastActiveSpeakerInterruptAt = 0;
-    room.on(RoomEvent.ActiveSpeakersChanged, (speakers) => {
-        if (!Array.isArray(speakers) || speakers.length === 0)
-            return;
-        const remoteSpeakers = speakers.filter((s) => s instanceof RemoteParticipant);
-        if (remoteSpeakers.length === 0)
-            return;
-        if (currentVoiceMode === 'realtime')
-            return;
-        if (agentState !== 'speaking')
-            return;
-        const now = Date.now();
-        const debounced = now - lastActiveSpeakerInterruptAt < 1000;
-        lastActiveSpeakerInterruptAt = now;
-        try {
-            if (!debounced) {
-                const ids = remoteSpeakers.map((s) => s.identity).join(',');
-                console.log(`🎤 ActiveSpeakersChanged: remote speakers [${ids}] + agent speaking → interrupting TTS`);
-            }
-            currentSession?.interrupt();
-        }
-        catch (err) {
-            if (!debounced)
-                console.warn('⚠️ active-speaker interrupt failed:', err instanceof Error ? err.message : err);
-        }
-    });
+    // NOTE: previously this section also had a RoomEvent.ActiveSpeakersChanged
+    // handler that interrupted TTS on any sustained audio activity (~50ms after
+    // mic onset). That fired too eagerly — coughs, paper rustles, the agent's
+    // own TTS bleeding through the mic, and other non-speech sounds tripped it
+    // ~10-15% of the time, leaving the agent silent with no recovery path
+    // (because no STT transcript would follow). Dropped in favor of the
+    // user_state_changed → 'speaking' handler below, which is fed by Deepgram
+    // Flux STT's speech-vs-noise classification: slower (~100-300ms) but
+    // confidence-aware. The latency tradeoff is worth eliminating the false
+    // interrupts at the root.
     room.on(RoomEvent.Disconnected, () => {
         console.log('👋 Disconnected from room');
         // Clean up active research and voice queue
@@ -2868,19 +2833,20 @@ async function main() {
                 }
             });
             // User state tracking — prevents queue from colliding with server-side VAD.
-            // Also a secondary interrupt trigger: when Deepgram STT classifies speech onset
-            // it propagates here via agent_activity.onStartOfSpeech → _updateUserState('speaking').
-            // Fires later than ActiveSpeakersChanged (Deepgram has ~100-300ms classification
-            // latency vs LiveKit's ~50-100ms audio-level) but acts as a redundant fallback in
-            // case the room-level event drops. interrupt() is idempotent on an already-
-            // interrupted SpeechHandle so calling both paths is harmless.
+            // Also the PRIMARY interrupt trigger now that the over-eager ActiveSpeakersChanged
+            // path is gone. Fires when Deepgram Flux STT classifies frames as speech (not noise)
+            // and propagates via agent_activity.onStartOfSpeech → _updateUserState('speaking').
+            // Latency ~100-300ms after mic onset, which is the cost of confidence-aware
+            // detection — vs the prior ActiveSpeakers handler that fired at ~50ms on any audio
+            // activity and tripped ~10-15% false interrupts on coughs, paper rustle, agent's
+            // own TTS bleeding through the mic, etc.
             sess.on('user_state_changed', (ev) => {
                 const prev = userState;
                 userState = ev.newState;
                 console.log(`👤 User state: ${prev} → ${ev.newState} (agent: ${agentState})`);
                 if (ev.newState === 'speaking' && agentState === 'speaking' && sessionVoiceMode !== 'realtime') {
                     try {
-                        console.log('🎤 user_state_changed=speaking + agent speaking → interrupting TTS (fallback)');
+                        console.log('🎤 user_state_changed=speaking + agent speaking → interrupting TTS');
                         currentSession?.interrupt();
                     }
                     catch (err) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.37",
+  "version": "0.9.39",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {