polygram 0.12.0-rc.18 → 0.12.0-rc.19

package/lib/process/cli-process.js

@@ -362,6 +362,29 @@ class CliProcess extends Process {
       this._startPingLoop();
     } catch (err) {
       await this._teardownOnStartFailure();
+      // Self-heal an unresumable stale session. A `--resume` spawn that dies
+      // during startup means the persisted session can't be resumed cleanly.
+      // The production case (shumorobot general chat, session minted 2026-05-27):
+      // claude resumes a week-old session → renders the aged-session "Resuming
+      // the full session… Resume from summary?" dialog → the gate confirms it →
+      // the /compact resume exits code 0 → the gate reports TMUX_SESSION_GONE.
+      // Retrying the SAME id loops forever (the chat sat dead for days). Clear it
+      // and respawn ONCE with a fresh session: a fresh spawn has no aged-session
+      // dialog and starts clean — exactly what made the Music topic healthy once
+      // its cwd change minted a new session. The `_freshRetry` guard means a
+      // genuine fresh-spawn failure throws instead of looping. onInit re-upserts
+      // the fresh id on success, so the dead id leaves the DB without extra wiring.
+      if (this._shouldRefreshAfterStartupFailure(err, opts)) {
+        this._logEvent('cli-stale-session-reset', {
+          stale_session_id: opts.existingSessionId,
+          code: err.code,
+        });
+        this.logger.warn?.(
+          `[${this.label}] channels: startup failed resuming ${opts.existingSessionId} `
+          + `(${err.code}) — clearing stale session, respawning fresh.`,
+        );
+        return this.start({ ...opts, existingSessionId: null, _freshRetry: true });
+      }
       throw err;
     }
 
@@ -415,6 +438,27 @@ class CliProcess extends Process {
     }
   }
 
+  /**
+   * Decide whether a start() failure warrants a one-shot fresh-session
+   * retry. True only when ALL hold:
+   *   - we were RESUMING a session (`opts.existingSessionId` set) — a fresh
+   *     spawn that fails is a real failure, not a poisoned resume;
+   *   - we haven't already retried (`_freshRetry` guard) — prevents an
+   *     infinite respawn loop if the fresh spawn also dies;
+   *   - the failure is a startup-gate death: claude exited
+   *     (`TMUX_SESSION_GONE`) or the dialog gate timed out
+   *     (`CHANNELS_DIALOG_TIMEOUT`) — the signature of an unresumable aged
+   *     session (the "Resume from summary?" /compact path that exits code 0).
+   * Any other error (bridge-handshake timeout on a healthy session, config
+   * error, etc.) throws as-is.
+   */
+  _shouldRefreshAfterStartupFailure(err, opts) {
+    const code = err && err.code;
+    return !!(opts && opts.existingSessionId)
+      && !(opts && opts._freshRetry)
+      && (code === 'TMUX_SESSION_GONE' || code === 'CHANNELS_DIALOG_TIMEOUT');
+  }
+
   /**
    * M1 refactor: socket-server lifecycle delegated to ChannelsBridgeServer.
    * This class wires the event surface (bridge-ready, bridge-message,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "polygram",
-  "version": "0.12.0-rc.18",
+  "version": "0.12.0-rc.19",
   "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": {