osborn 0.8.7 → 0.8.8

package/dist/claude-auth.js

@@ -150,6 +150,13 @@ export async function checkClaudeAuthStatus() {
  * Strips ALL whitespace first (like vutran1710/claudebox) to handle
  * Ink UI wrapping the URL across multiple lines.
  * Also cleans trailing "Pastecodehereifprompted" that Ink appends.
+ *
+ * IMPORTANT: strips the `redirect_uri` query parameter (which points to a
+ * localhost callback server on the *sprite*, not the user's machine). With
+ * no redirect_uri, claude.com falls back to showing the auth code in-page,
+ * which the user pastes back into the modal. This is the only flow that
+ * works for cloud sandboxes — the localhost redirect breaks both on phones
+ * (no listener) AND on desktops (sprite's localhost is unreachable).
  */
 function extractOAuthUrl(text) {
     // Strip ANSI codes
@@ -168,6 +175,39 @@ function extractOAuthUrl(text) {
         if (idx > 0)
             url = url.substring(0, idx);
     }
+    // Strip the localhost redirect_uri so claude.com shows a pasteable code
+    // instead of trying to redirect. URL() can't be used here because it
+    // re-encodes the path, so we surgically delete the redirect_uri param.
+    url = stripRedirectUri(url);
+    return url;
+}
+/**
+ * Strip the `redirect_uri` query param from an OAuth URL.
+ *
+ * Background: `claude setup-token` spawns a one-shot localhost HTTP server on
+ * a random port and registers it as the redirect_uri. That works fine when the
+ * user is on the same machine as the CLI, but on a sprite the URL points to
+ * the *sprite's* localhost — unreachable from the user's browser regardless
+ * of whether they open the auth link on their PC or their phone. With no
+ * redirect_uri at all, claude.ai falls back to its in-page code display
+ * (the same flow that `claude setup-token`'s "Paste code here if prompted"
+ * Ink input is built to consume), and the user can paste the code back into
+ * our modal — which works whether they signed in on phone or desktop.
+ *
+ * Done with regex rather than `new URL()` because the URL constructor
+ * normalizes the path (which can break Claude's strict redirect check)
+ * and re-encodes spaces/special chars in other params.
+ */
+function stripRedirectUri(url) {
+    const before = url;
+    // Three cases: leading param (?redirect_uri=...&), middle/trailing (&redirect_uri=...),
+    // and only param (?redirect_uri=...). Order matters so cleanup leaves the URL well-formed.
+    url = url.replace(/&redirect_uri=[^&]*/g, '');
+    url = url.replace(/\?redirect_uri=[^&]*&/g, '?');
+    url = url.replace(/\?redirect_uri=[^&]*$/g, '');
+    if (before !== url) {
+        console.log('🔑 Stripped localhost redirect_uri from OAuth URL — claude.ai will show a pasteable code instead of redirecting');
+    }
     return url;
 }
 // ─────────────────────────────────────────

package/dist/claude-llm.js

@@ -12,11 +12,7 @@ import { EventEmitter } from 'events';
 import { saveSessionMetadata, getSessionWorkspace } from './config.js';
 import { getResearchSystemPrompt, getDirectModeResearchPrompt } from './prompts.js';
 import { existsSync, readdirSync, readFileSync } from 'node:fs';
-import { join, dirname } from 'node:path';
-import { fileURLToPath } from 'node:url';
-// Directory of this module — used to locate co-located prompt files (e.g., turn-shape reminder).
-const __claudeLlmDir = dirname(fileURLToPath(import.meta.url));
-const TURN_SHAPE_REMINDER_PATH = join(__claudeLlmDir, 'prompts', 'turn-shape-reminder.md');
+import { join } from 'node:path';
 /**
  * Strip markdown formatting for TTS (text-to-speech)
  * Removes **bold**, ##headers, ```code```, etc. so TTS doesn't read them literally
@@ -621,7 +617,7 @@ export class ClaudeLLM extends llm.LLM {
                             callbacks.eventEmitter.emit('assistant_text', { text: block.text });
                             const ttsChunk = stripMarkdownForTTS(block.text);
                             if (ttsChunk.trim()) {
-                                console.log(`🔊 TTS say (${ttsChunk.length} chars): "${ttsChunk}"`);
+                                console.log(`🔊 TTS say (${ttsChunk.length} chars): "${ttsChunk.substring(0, 60)}..."`);
                                 callbacks.eventEmitter.emit('tts_say', { text: ttsChunk });
                             }
                         }
@@ -736,7 +732,7 @@ class ClaudeLLMStream extends llm.LLMStream {
                 });
                 return;
             }
-            console.log(`🎤 User (${userText.length} chars): "${userText}"`);
+            console.log(`🎤 User: "${userText.substring(0, 100)}${userText.length > 100 ? '...' : ''}"`);
             // Build Claude Agent SDK options
             const resumeSessionId = this.#opts.resumeSessionId;
             const continueSession = this.#opts.continueSession;
@@ -856,35 +852,6 @@ class ClaudeLLMStream extends llm.LLMStream {
                                     this.#eventEmitter.emit('tool_result', { name: toolName, input: toolInput, response: toolResponse });
                                     return {};
                                 }]
-                        }],
-                    // Per-turn behavioral re-anchor. Fires on EVERY user message that reaches Claude
-                    // (initial requests, follow-ups, mid-flight steering, resumed-session messages).
-                    // Reads the reminder text from disk every call, so it's hot-editable just like the
-                    // main prompt — edit agent/src/prompts/turn-shape-reminder.md, reconnect, next message
-                    // sees the new reminder. The SDK injects `additionalContext` alongside the user's actual
-                    // message so the model sees both the literal user input AND the reminder, weighing them
-                    // together. This is what fights JSONL-history-overrides-system-prompt drift on resumed
-                    // sessions: the conductor pattern gets re-asserted on every turn instead of being
-                    // anchored only at session-init time.
-                    UserPromptSubmit: [{
-                            matcher: '.*',
-                            hooks: [async (input) => {
-                                    try {
-                                        const reminder = readFileSync(TURN_SHAPE_REMINDER_PATH, 'utf-8');
-                                        const promptPreview = String(input?.prompt || '').substring(0, 60).replace(/\n/g, ' ');
-                                        console.log(`📌 UserPromptSubmit: injected turn-shape reminder (${reminder.length} chars) for prompt="${promptPreview}..."`);
-                                        return {
-                                            hookSpecificOutput: {
-                                                hookEventName: 'UserPromptSubmit',
-                                                additionalContext: reminder,
-                                            },
-                                        };
-                                    }
-                                    catch (err) {
-                                        console.error('⚠️ UserPromptSubmit: failed to load turn-shape-reminder.md:', err instanceof Error ? err.message : err);
-                                        return { hookSpecificOutput: { hookEventName: 'UserPromptSubmit' } };
-                                    }
-                                }]
                         }]
                 },
                 // Named sub-agents — Haiku overseer delegates to these specialists.
@@ -1109,12 +1076,12 @@ class ClaudeLLMStream extends llm.LLMStream {
                                 if (this.#opts.skipTTSQueue) {
                                     // Direct mode: emit event for session.say() — bypasses LiveKit's
                                     // BufferedTokenStream which causes stuck/delayed/out-of-order audio
-                                    console.log(`🔊 TTS say (${ttsChunk.length} chars): "${ttsChunk}"`);
+                                    console.log(`🔊 TTS say (${ttsChunk.length} chars): "${ttsChunk.substring(0, 60)}..."`);
                                     this.#eventEmitter.emit('tts_say', { text: ttsChunk });
                                 }
                                 else {
                                     // Realtime mode: use LLM stream queue (framework handles TTS)
-                                    console.log(`🔊 TTS stream (${ttsChunk.length} chars): "${ttsChunk}"`);
+                                    console.log(`🔊 TTS stream (${ttsChunk.length} chars): "${ttsChunk.substring(0, 60)}..."`);
                                     this.queue.put({
                                         id: requestId,
                                         delta: { role: 'assistant', content: ttsChunk },
@@ -1134,11 +1101,11 @@ class ClaudeLLMStream extends llm.LLMStream {
                         const ttsText = stripMarkdownForTTS(rawResult);
                         if (ttsText.trim()) {
                             if (this.#opts.skipTTSQueue) {
-                                console.log(`🔊 TTS say result (${ttsText.length} chars): "${ttsText}"`);
+                                console.log(`🔊 TTS say result (${ttsText.length} chars): "${ttsText.substring(0, 60)}..."`);
                                 this.#eventEmitter.emit('tts_say', { text: ttsText });
                             }
                             else {
-                                console.log(`🔊 TTS result (${ttsText.length} chars): "${ttsText}"`);
+                                console.log(`🔊 TTS result (${ttsText.length} chars): "${ttsText.substring(0, 60)}..."`);
                                 this.queue.put({
                                     id: requestId,
                                     delta: { role: 'assistant', content: ttsText },

package/dist/codex-llm.js

@@ -97,7 +97,7 @@ class CodexLLMStream extends llm.LLMStream {
                 });
                 return;
             }
-            console.log(`🎤 User (${userText.length} chars): "${userText}"`);
+            console.log(`🎤 User: "${userText.substring(0, 100)}${userText.length > 100 ? '...' : ''}"`);
             // Create or reuse thread
             if (!this.#thread) {
                 console.log('🆕 Starting new Codex thread');

package/dist/index.js

@@ -227,47 +227,6 @@ function startApiServer(workingDir, port) {
             setTimeout(() => process.exit(0), 150);
             return;
         }
-        // GET /events — Server-Sent Events heartbeat for cloud-sandbox keepalive.
-        //
-        // This endpoint is the single thing preventing Sprites' CRIU-based
-        // hibernation from freezing osborn's Node.js event loop and dropping our
-        // LiveKit WebSocket mid-session. Short HTTP pings don't work: Sprites'
-        // warm state serves /health responses from a process snapshot without
-        // actually resuming the event loop, so background timers (including
-        // LiveKit heartbeats) stop firing after a few seconds. That causes the
-        // LiveKit server to drop osborn's participant, delete the room, and
-        // leave any future user joins stuck at "Connecting..." forever.
-        //
-        // An OPEN long-lived TCP connection keeps the sprite in 'running' state.
-        // The frontend opens this endpoint on chat page mount and holds it open
-        // for the entire voice session. While open, osborn's event loop ticks
-        // continuously, LiveKit heartbeats fire, and the room stays alive.
-        //
-        // For local (non-cloud) dev, this endpoint is harmless — it just idles
-        // on a client that may never connect. Zero cost when unused.
-        if (req.method === 'GET' && url.pathname === '/events') {
-            res.writeHead(200, {
-                'Content-Type': 'text/event-stream',
-                'Cache-Control': 'no-cache',
-                'Connection': 'keep-alive',
-                // Disable proxy buffering (nginx-style) so each ping is flushed
-                // through Sprites' reverse proxy immediately rather than batched.
-                'X-Accel-Buffering': 'no',
-            });
-            res.write(`: sprite-keepalive connected at ${new Date().toISOString()}\n\n`);
-            const heartbeat = setInterval(() => {
-                try {
-                    res.write(`: ping ${Date.now()}\n\n`);
-                }
-                catch { }
-            }, 10_000);
-            req.on('close', () => {
-                clearInterval(heartbeat);
-                console.log('[events] SSE client disconnected');
-            });
-            console.log('[events] SSE client connected');
-            return;
-        }
         res.writeHead(404, { 'Content-Type': 'application/json' });
         res.end(JSON.stringify({ error: 'Not found' }));
     });
@@ -447,43 +406,6 @@ async function main() {
     let currentSession = null;
     let currentAgent = null; // For updateChatCtx() context injection
     let currentLLM = null;
-    /**
-     * Hard-kill the in-flight Claude SDK query AND the persistent subprocess.
-     *
-     * Why this exists: the persistent ClaudeLLM session is deliberately kept alive
-     * across user messages to avoid JSONL replay (see CLAUDE.md "Persistent Session
-     * Architecture"). When the participant disconnects, simply nulling `currentLLM`
-     * drops the JS reference but does NOT kill the underlying Claude Code subprocess
-     * — the SDK keeps draining the MessageChannel, running tools, and pushing TTS
-     * calls into a now-null voice session. Visible in logs as repeated:
-     *   "⚠️ tts_say fired but currentSession is null — text dropped"
-     * followed by orphaned `🔧 Claude: Bash` calls and `📍 Checkpoint captured` lines
-     * that nobody is listening to. Wasted compute, wasted tokens, possible side effects.
-     *
-     * The right cleanup is `abortQuery()` (on ClaudeLLM directly) or `abortAgent()`
-     * (on PipelineDirectLLM, which wraps ClaudeLLM). They both call into
-     * `closeSession()` → kills the subprocess. We duck-type to handle both class
-     * shapes since `currentLLM` can hold either, depending on voice mode.
-     */
-    function killCurrentLLM(reason) {
-        if (!currentLLM)
-            return;
-        try {
-            const llm = currentLLM;
-            if (typeof llm.abortQuery === 'function') {
-                llm.abortQuery();
-            }
-            else if (typeof llm.abortAgent === 'function') {
-                llm.abortAgent();
-            }
-            else {
-                console.warn(`⚠️ killCurrentLLM(${reason}): no abort method on currentLLM`);
-            }
-        }
-        catch (err) {
-            console.error(`❌ killCurrentLLM(${reason}) failed:`, err instanceof Error ? err.message : err);
-        }
-    }
     let localParticipant = null;
     let agentState = 'initializing';
     // Session-level always-allow list: paths the user has approved for this session without prompting
@@ -557,7 +479,7 @@ async function main() {
         // fullText is what was being spoken when interrupted (passed from tts_say handler).
         // No word-level cutoff for say() — only generateReply pipeline has that — but Claude
         // knows its own output from JSONL, so the full block is enough context.
-        console.log(`🔇 Speech interrupted. Was speaking (${fullText.length} chars): "${fullText}"`);
+        console.log(`🔇 Speech interrupted. Was speaking: "${fullText.substring(0, 80)}..."`);
         // Read last 10 assistant messages from JSONL (Claude's full untruncated output).
         // SessionMessage.text is pre-joined from all text content blocks.
         let recentMessages = '';
@@ -933,7 +855,7 @@ async function main() {
         });
         // Wire up Claude text output - RAW text goes to frontend for chat bubbles
         directLLM.events.on('assistant_text', (data) => {
-            console.log(`💬 Claude text (${data.text?.length || 0} chars): ${data.text || ''}`);
+            console.log(`💬 Claude text: ${data.text?.substring(0, 60)}...`);
             sendToFrontend({
                 type: 'claude_output',
                 text: data.text,
@@ -943,7 +865,7 @@ async function main() {
         });
         // Wire up Claude final result - RAW result goes to frontend
         directLLM.events.on('assistant_result', (data) => {
-            console.log(`📋 Claude result (${data.text?.length || 0} chars): ${data.text || ''}`);
+            console.log(`📋 Claude result: ${data.text?.substring(0, 60)}...`);
             sendToFrontend({
                 type: 'claude_output',
                 text: data.text,
@@ -1061,7 +983,7 @@ async function main() {
         directLLM.events.on('tts_say', (data) => {
             // Guard: session must be alive — TTS errors can kill the session while background query runs
             if (!currentSession) {
-                console.warn(`⚠️ tts_say fired but currentSession is null — text dropped (${data.text?.length || 0} chars): "${data.text || ''}"`);
+                console.warn(`⚠️ tts_say fired but currentSession is null — text dropped: "${data.text?.substring(0, 60)}"`);
                 return;
             }
             if (!data.text?.trim()) {
@@ -1069,7 +991,7 @@ async function main() {
                 return;
             }
             const sayId = Date.now(); // simple ID to correlate start/end logs
-            console.log(`🗣️ [${sayId}] session.say START (${data.text.length} chars): "${data.text}"`);
+            console.log(`🗣️ [${sayId}] session.say START (${data.text.length} chars): "${data.text.substring(0, 60)}..."`);
             try {
                 const handle = currentSession.say(data.text);
                 if (handle && typeof handle.addDoneCallback === 'function') {
@@ -1197,7 +1119,7 @@ async function main() {
             }
         });
         realtimeClaudeHandler.events.on('assistant_result', (data) => {
-            console.log(`📋 Claude result (${data.text?.length || 0} chars): ${data.text || ''}`);
+            console.log(`📋 Claude result: ${data.text?.substring(0, 60)}...`);
             sendToFrontend({
                 type: 'claude_output',
                 text: data.text,
@@ -1689,9 +1611,6 @@ async function main() {
         lastCompletedResearch = null;
         currentSession = null;
         currentAgent = null;
-        // Same disconnect-leak fix as the other two cleanup sites — kill the Claude SDK
-        // subprocess BEFORE dropping the reference. See killCurrentLLM() for full context.
-        killCurrentLLM('disconnected_cleanup');
         currentLLM = null;
         clearFastBrainSession();
         clearPipelineFastBrainSession();
@@ -1734,9 +1653,6 @@ async function main() {
             catch { }
             currentSession = null;
             currentAgent = null;
-            // Same disconnect-leak fix — kill the previous user's Claude subprocess
-            // before binding currentLLM to the new user's session below.
-            killCurrentLLM('previous_session_cleanup');
             currentLLM = null;
         }
         // Extract voice architecture, provider, and sessionId from participant metadata (sent by frontend)
@@ -1905,7 +1821,7 @@ async function main() {
                 // (Gemini v1.0.51: userInput in generateReply creates a user conversation item)
                 if (normalized.startsWith('[SCRIPT]') || normalized.startsWith('[PROACTIVE]') || normalized.startsWith('[NOTIFICATION]'))
                     return;
-                console.log(`📝 User (${source}, ${transcript.length} chars): "${transcript}"`);
+                console.log(`📝 User (${source}): "${transcript.substring(0, 60)}..."`);
                 sendToFrontend({ type: 'user_transcript', text: transcript });
                 lastSentUserTranscript = normalized;
             }
@@ -1915,7 +1831,7 @@ async function main() {
                 const normalized = text.trim().replace(/\s+/g, ' ');
                 if (normalized === lastSentAgentTranscript)
                     return;
-                console.log(`💬 Agent (${source}, ${text.length} chars): "${text}"`);
+                console.log(`💬 Agent (${source}): "${text.substring(0, 60)}..."`);
                 sendToFrontend({ type: 'assistant_response', text });
                 lastSentAgentTranscript = normalized;
             }
@@ -2373,9 +2289,6 @@ async function main() {
             })();
         }
         currentAgent = null;
-        // Kill the Claude SDK subprocess BEFORE dropping the reference, otherwise the
-        // persistent session keeps running tools and pushing TTS into a dead session.
-        killCurrentLLM('participant_disconnected');
         currentLLM = null;
         clearFastBrainSession();
         clearPipelineFastBrainSession();
@@ -2421,10 +2334,10 @@ async function main() {
                             fullContent += `\n\n[Image attached: ${f.name}]`;
                         }
                     }
-                    console.log(`📝 Text + ${files.length} file(s) (${fullContent.length} chars): "${fullContent}"`);
+                    console.log(`📝 Text + ${files.length} file(s): "${fullContent.substring(0, 100)}"`);
                 }
                 else {
-                    console.log(`📝 Text (${fullContent.length} chars): "${fullContent}"`);
+                    console.log(`📝 Text: "${fullContent.substring(0, 100)}"`);
                 }
                 // Skip interrupt for Gemini — disrupts state machine (hangs in speaking state)
                 if (currentProvider !== 'gemini') {

package/dist/pipeline-direct-llm.js

@@ -82,7 +82,7 @@ export class PipelineDirectLLM extends llm.LLM {
                 break;
             }
         }
-        console.log(`📥 [pipeline] chat() call #${callN} (${userText.length} chars): "${userText}"`);
+        console.log(`📥 [pipeline] chat() call #${callN}: "${userText.substring(0, 60)}"`);
         // Check for pending interruption context — enrich user message if interrupted
         const interruptCtx = this.#opts.getAndConsumeInterruptionContext?.();
         if (interruptCtx && userText.trim()) {