orchestrix-yuri 2.3.4 → 2.5.0

package/lib/gateway/engine/claude-tmux.js

@@ -143,24 +143,28 @@ function tmuxSafe(cmd) {
 
 // ── Claude Code TUI Indicators ─────────────────────────────────────────────────
 //
-// Claude Code uses three circle symbols as primary state indicators:
+// Claude Code TUI state indicators vary by version and statusline config:
 //
-//   ○  (U+25CB) IDLE       — Claude is waiting for user input
-//   ●  (U+25CF) PROCESSING — Claude is actively generating a response
-//   ◐  (U+25D0) APPROVAL   — Claude is waiting for permission approval
+//   Idle indicators (any of these = ready for input):
+//     ○  (U+25CB) — circle idle indicator (shown with certain statusline configs)
+//     ❯           — prompt cursor (always shown when idle, most reliable)
 //
-// During processing, a Braille spinner animates: ⠋ ⠙ ⠹ ⠸ ⠼ ⠴ ⠦ ⠧ ⠇ ⠏
-// with cycling verbs like "Baking...", "Computing...", "Thinking..."
+//   Processing indicators:
+//     ●  (U+25CF) — filled circle, active generation
+//     Braille spinner: ⠋ ⠙ ⠹ ⠸ ⠼ ⠴ ⠦ ⠧ ⠇ ⠏
 //
-// Completion message format (past-tense verb + duration):
-//   "Baked for 31s", "Worked for 2m 45s", "Cooked for 1m 6s"
-//   Pattern: /[A-Z][a-z]*ed for \d+/
+//   Status line elements (NOT state indicators):
+//     ◐  — effort level indicator (e.g. "◐ medium · /effort"), NOT approval prompt
+//
+//   Completion message (past-tense verb + duration):
+//     "Baked for 31s", "Worked for 2m 45s"
+//     Pattern: /[A-Z][a-z]*ed for \d+/
 //
 // ────────────────────────────────────────────────────────────────────────────────
 
 const BRAILLE_SPINNER = /[⠋⠙⠹⠸⠼⠴⠦⠧⠇⠏]/;
 const COMPLETION_RE = /[A-Z][a-z]*ed for \d+/;
-const IDLE_RE = /○/;
+const IDLE_RE = /[○❯]/;
 const PROCESSING_RE = /●/;
 
 /**
@@ -200,43 +204,31 @@ function paneTail(name, n) {
 }
 
 /**
- * Detect Claude Code's current state from pane output.
- *
- * Note: We launch with --dangerously-skip-permissions, so approval prompts (◐)
- * should never appear. The detection is kept for robustness but no auto-approve
- * action is taken — sending blind 'y' keystrokes is dangerous.
+ * Check if Claude Code has started up and is ready for input.
+ * Used ONLY during session initialization — looks for the ❯ input prompt
+ * which appears once Claude Code has fully loaded.
  *
- * @returns {'idle'|'processing'|'complete'|'unknown'}
+ * DO NOT use this for response completion detection — ❯ is always visible.
  */
-function detectState(name) {
+function isStarted(name) {
   const tail = paneTail(name, 15);
+  return IDLE_RE.test(tail);
+}
 
-  // Priority 1: Completion message — most reliable signal
-  // e.g. "Baked for 31s", "Worked for 2m 45s"
-  if (COMPLETION_RE.test(tail)) {
-    return 'complete';
-  }
-
-  // Priority 2: Idle indicator — waiting for input
-  if (IDLE_RE.test(tail)) {
-    return 'idle';
-  }
-
-  // Priority 3: Processing indicator — still working
-  if (PROCESSING_RE.test(tail) || BRAILLE_SPINNER.test(tail)) {
-    return 'processing';
-  }
-
-  return 'unknown';
+/**
+ * Check if a completion message is present in the pane output.
+ * e.g. "Baked for 31s", "Worked for 2m 45s"
+ * This is the most reliable signal that Claude has finished responding.
+ */
+function hasCompletionMessage(text) {
+  return COMPLETION_RE.test(text);
 }
 
 /**
- * Detect if Claude Code is idle (ready for input).
- * Checks for ○ idle indicator or completion message.
+ * Check if Claude Code is actively processing (● spinner visible).
  */
-function isIdle(name) {
-  const state = detectState(name);
-  return state === 'idle' || state === 'complete';
+function isProcessing(text) {
+  return PROCESSING_RE.test(text) || BRAILLE_SPINNER.test(text);
 }
 
 // ── Context Management ─────────────────────────────────────────────────────────
@@ -304,7 +296,7 @@ async function proactiveCompact(name) {
   log.tmux('Proactive /compact triggered');
   injectMessage(name, '/compact focus on the most recent user conversation and any pending operations');
 
-  const ok = await waitForIdle(name, 120000); // compact can take up to 2min
+  const ok = await waitForReady(name, 120000); // compact can take up to 2min
   if (ok) {
     _messageCount = 0;
     log.tmux('Proactive /compact completed');
@@ -354,7 +346,7 @@ async function createSession(engineConfig) {
   // Default 60s — Claude Code needs time to load CLAUDE.md, connect MCP servers, etc.
   const startupTimeout = engineConfig.startup_timeout || 60000;
   log.tmux(`Waiting for Claude Code to become idle (timeout: ${startupTimeout / 1000}s)...`);
-  const started = await waitForIdle(sessionName, startupTimeout);
+  const started = await waitForReady(sessionName, startupTimeout);
   if (!started) {
     // Don't kill session on failure — let user debug with tmux attach
     const tail = paneTail(sessionName, 10);
@@ -371,7 +363,7 @@ async function createSession(engineConfig) {
   if (l1) {
     log.tmux('Injecting L1 context...');
     await injectMessage(sessionName, l1);
-    await waitForIdle(sessionName, 120000); // allow up to 2min for L1 processing
+    await waitForReady(sessionName, 120000); // allow up to 2min for L1 processing
   }
 
   _sessionReady = true;
@@ -379,10 +371,12 @@ async function createSession(engineConfig) {
 }
 
 /**
- * Wait for Claude Code to become idle.
- * @returns {Promise<boolean>} true if idle detected, false if timeout
+ * Wait for Claude Code to be ready (❯ prompt visible).
+ * Used for session init and after /compact — NOT for response capture.
+ *
+ * @returns {Promise<boolean>} true if ready detected, false if timeout
  */
-function waitForIdle(name, timeoutMs) {
+function waitForReady(name, timeoutMs) {
   const pollInterval = 2000;
   return new Promise((resolve) => {
     const deadline = Date.now() + timeoutMs;
@@ -394,7 +388,7 @@ function waitForIdle(name, timeoutMs) {
         return resolve(false);
       }
 
-      if (isIdle(name)) {
+      if (isStarted(name)) {
         return resolve(true);
       }
       setTimeout(poll, pollInterval);
@@ -422,10 +416,14 @@ function injectMessage(name, text) {
 /**
  * Capture the response after injecting a message.
  *
- * Detection priority:
+ * Detection strategy (❯ prompt is always visible, so we CANNOT use idle detection):
  *   P1: Completion message — "[Verb]ed for [N]s/m" (e.g. "Baked for 31s")
- *   P2: Idle indicator — ○ appears in pane tail
- *   P3: Content stability — 3 consecutive polls with identical MD5 hash
+ *       Most reliable signal. Appears exactly once when Claude finishes.
+ *   P2: Content stability — pane output unchanged for N consecutive polls.
+ *       Fallback for edge cases where completion message is missed.
+ *
+ * We also track whether content has changed since injection (via marker)
+ * to avoid returning before Claude has even started responding.
  */
 async function captureResponse(name, marker, engineConfig) {
   const timeout = engineConfig.timeout || 300000;
@@ -435,7 +433,12 @@ async function captureResponse(name, marker, engineConfig) {
   const deadline = Date.now() + timeout;
   let lastHash = '';
   let stableCount = 0;
-  let sawProcessing = false;
+  let contentChanged = false;
+
+  // Capture baseline right after injection
+  const baselineRaw = capturePaneRaw(name, 500);
+  const baselineHash = crypto.createHash('md5').update(baselineRaw).digest('hex');
+  lastHash = baselineHash;
 
   return new Promise((resolve) => {
     const poll = () => {
@@ -451,31 +454,22 @@ async function captureResponse(name, marker, engineConfig) {
         return resolve({ reply: '❌ Claude Code session terminated unexpectedly.', raw: '' });
       }
 
-      const state = detectState(name);
       const raw = capturePaneRaw(name, 500);
       const hash = crypto.createHash('md5').update(raw).digest('hex');
 
-      // Track that Claude has started processing (● appeared)
-      // This prevents premature completion detection if ○ is still visible
-      // from the previous idle state before Claude begins processing.
-      if (state === 'processing') {
-        sawProcessing = true;
-        stableCount = 0;
-        lastHash = hash;
-        return setTimeout(poll, pollInterval);
+      // Track if content has changed since injection
+      if (hash !== baselineHash) {
+        contentChanged = true;
       }
 
       // P1: Completion message — most reliable done signal
-      if (state === 'complete' && sawProcessing) {
-        return resolve(extractResponse(raw, marker));
-      }
-
-      // P2: Idle indicator — done if we saw processing start
-      if (state === 'idle' && sawProcessing) {
+      // Only check after content has changed (Claude has started responding)
+      if (contentChanged && hasCompletionMessage(paneTail(name, 15))) {
         return resolve(extractResponse(raw, marker));
       }
 
-      // P4: Content stability fallback
+      // P2: Content stability — pane unchanged for N polls
+      // Only trigger after content has changed from baseline
       if (hash === lastHash) {
         stableCount++;
       } else {
@@ -483,7 +477,7 @@ async function captureResponse(name, marker, engineConfig) {
         lastHash = hash;
       }
 
-      if (stableCount >= stableThreshold && sawProcessing) {
+      if (contentChanged && stableCount >= stableThreshold) {
         log.tmux('Response detected via content stability');
         return resolve(extractResponse(raw, marker));
       }
@@ -492,7 +486,6 @@ async function captureResponse(name, marker, engineConfig) {
     };
 
     // Initial delay: give Claude time to start processing
-    // before first poll (avoids false-positive idle detection)
     setTimeout(poll, Math.max(pollInterval, 3000));
   });
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "orchestrix-yuri",
-  "version": "2.3.4",
+  "version": "2.5.0",
   "description": "Yuri — Meta-Orchestrator for Orchestrix. Drive your entire project lifecycle with natural language.",
   "main": "lib/installer.js",
   "bin": {