polygram 0.12.0-rc.10 → 0.12.0-rc.11

package/lib/process/cli-process.js

@@ -118,6 +118,16 @@ const STREAMING_HINT_RE = /esc to interrupt/i;
 // — false positives surface as no-op telemetry, false negatives surface
 // as the idle-ceiling timeout (~10min).
 const UNKNOWN_PROMPT_HEURISTIC_RE = /(\?\s*$|\(y\/N\)|Yes\/No|❯\s|^\s*[12345]\.\s)/im;
+// Dead-bridge signal. Claude Code prints "<source>  no MCP server configured
+// with that name" when a channel source references an MCP server that isn't
+// registered. Music topic incident (2026-06-01): seen mid-turn after an
+// auto-/compact of a large resumed session redrew the TUI and dropped the
+// `server:polygram-bridge` binding — the turn could no longer deliver its
+// reply and hung. The bridge name precedes the error on the same line, so
+// anchor on it: matching the bare phrase risks a false hit on prose that
+// merely quotes the error, while the healthy "polygram-bridge: <polygram-info>
+// …" connection line never contains "no MCP server configured".
+const BRIDGE_DEAD_RE = /polygram-bridge[^\n]*no MCP server configured/i;
 // Per-pattern rate limit so a dialog that lingers across multiple polls
 // doesn't spam sendControl/event emissions. Aligned with the 5s poll cadence.
 const MID_TURN_DEDUP_WINDOW_MS = 30_000;
@@ -1848,11 +1858,11 @@ class CliProcess extends Process {
    * landing just before the disconnect would otherwise leave a stray
    * timer on the dead instance).
    */
-  _handleBridgeDisconnected() {
+  _handleBridgeDisconnected(reason = 'socket-close') {
     this.bridgeReady = false;
     this.mcpReady = false;
     if (this.closed) return;
-    this.logger.warn?.(`[${this.label}] channels: bridge disconnected unexpectedly`);
+    this.logger.warn?.(`[${this.label}] channels: bridge disconnected unexpectedly (${reason})`);
     // L6: clear the interrupt grace timer alongside the rest of the lifecycle.
     if (this._interruptGraceTimer) {
       clearTimeout(this._interruptGraceTimer);
@@ -1875,7 +1885,7 @@ class CliProcess extends Process {
     this.pendingQueue.length = 0;
     this.inFlight = false;
     this.emit('bridge-disconnected');
-    this._logEvent('bridge-disconnected', { reason: 'socket-close' });
+    this._logEvent('bridge-disconnected', { reason });
   }
 
   async _doKill(reason) {
@@ -2328,6 +2338,36 @@ class CliProcess extends Process {
     }
     if (!pane) return;
 
+    // Wedge recovery (Music topic incident, 2026-06-01). Claude Code's
+    // channel source can lose its MCP-server registration mid-turn — most
+    // often after an auto-/compact of a large resumed session redraws the
+    // TUI and the `server:polygram-bridge` binding fails to re-resolve
+    // ("polygram-bridge  no MCP server configured with that name"). The
+    // bridge SOCKET stays up, so the socket-close path
+    // (bridgeServer 'bridge-disconnected' → _handleBridgeDisconnected) never
+    // fires — but claude can no longer deliver the reply, so the turn
+    // orphans and would hang until the wall-clock cap while this watchdog
+    // logged cli-mid-turn-unknown-prompt every 30s and recovered nothing.
+    // Detect it from the pane and route through the SAME recovery as a real
+    // socket disconnect: reject pending turns (user gets the 🔌 resend
+    // message via BRIDGE_DISCONNECTED) and emit 'bridge-disconnected' so
+    // process-manager kills + lazy-respawns the dead instance (next msg
+    // recovers the conversation via `claude --resume`). Gated on
+    // pendingTurns.size>0 by the early return above, so it can't fire on an
+    // idle session that merely has the string in scrollback.
+    if (BRIDGE_DEAD_RE.test(pane)) {
+      this.logger.warn?.(
+        `[${this.label}] cli: channels MCP registration lost mid-turn ` +
+        `(pane-detected) — recovering ${this.pendingTurns.size} orphaned turn(s)`,
+      );
+      this._logEvent('cli-bridge-detached-midturn', {
+        pending_count: this.pendingTurns.size,
+        session_id: this.claudeSessionId,
+      });
+      this._handleBridgeDisconnected('mcp-registration-lost');
+      return;
+    }
+
     const now = Date.now();
 
     // 0.12 Phase 3.2: liveness heartbeat. The TUI prints "esc to interrupt"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "polygram",
-  "version": "0.12.0-rc.10",
+  "version": "0.12.0-rc.11",
   "description": "Telegram daemon for Claude Code that preserves the OpenClaw per-chat session model. Migration path for OpenClaw users moving to Claude Code.",
   "main": "lib/ipc/client.js",
   "bin": {