polygram 0.12.0-rc.19 → 0.12.0-rc.20

package/lib/error/classify.js

@@ -195,6 +195,18 @@ const CODES = {
     isTransient: false,
     autoRecover: null,
   },
+  // TMUX_SESSION_GONE: claude exited during spawn so the tmux session vanished
+  // before the channel went live (the startup-gate's captureWide hit "can't
+  // find pane"). Usual cause: an unresumable aged session whose "Resume from
+  // summary?" /compact exits code 0. The dispatcher poison-clears the session
+  // on this code, so a resend genuinely starts fresh and works — hence the
+  // calm "send it again" copy instead of the old raw "[startup-gate]…" leak.
+  TMUX_SESSION_GONE: {
+    kind: 'tmuxSessionGone',
+    userMessage: '🔄 That chat got stuck starting up, so I reset it. Send your message again and I\'ll pick it up fresh.',
+    isTransient: false,
+    autoRecover: null,
+  },
   // TURN_TIMEOUT: 10-min wall-clock cap on a single channels turn. Mirror
   // of the tmux wall-clock ceiling — typically a runaway, not a wedge.
   // Not transient (auto-retry would just runaway again).

package/lib/handlers/dispatcher.js

@@ -178,6 +178,26 @@ function createDispatcher({
         aborted: wasAborted || undefined,
         replay: isReplay || undefined,
       });
+      // Startup-gate death (claude exited during spawn / the dialog gate timed
+      // out) of a likely-aged RESUMED session — the persisted claude_session_id
+      // can't be resumed cleanly (shumorobot general chat 2026-06-01→03: a
+      // week-old session renders claude's "Resume from summary?" dialog whose
+      // /compact resume exits code 0 → TMUX_SESSION_GONE → the chat re-resumes
+      // the same dead id on every message, stuck for days). Poison-clear so the
+      // NEXT message spawns a FRESH session — same recovery the auto-resume path
+      // does for BRIDGE_DISCONNECTED below. clearSessionId is a no-op DELETE when
+      // there's no row (a genuine fresh-spawn failure), so this is safe; and
+      // unlike an in-process recursive retry it never reuses a closed instance.
+      if ((err.code === 'TMUX_SESSION_GONE' || err.code === 'CHANNELS_DIALOG_TIMEOUT')
+          && typeof db.clearSessionId === 'function') {
+        dbWrite(
+          () => db.clearSessionId(sessionKey),
+          `clearSessionId: poisoned by ${err.code} on startup`,
+        );
+        logEvent('session-reset-after-startup-gate', {
+          chat_id: chatId, session_key: sessionKey, msg_id: msg?.message_id, code: err.code,
+        });
+      }
       // rc.55: surface replay failures with a meaningful message.
       // Pre-rc.55 any boot-replay turn that failed for ANY reason
       // was silently dropped. The rc.51-onward boot-replay path is

package/lib/process/cli-process.js

@@ -362,29 +362,6 @@ 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;
     }
 
@@ -438,27 +415,6 @@ 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.19",
+  "version": "0.12.0-rc.20",
   "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": {