polygram 0.12.0-rc.24 → 0.12.0-rc.26

package/lib/handlers/dispatcher.js

@@ -24,6 +24,13 @@
 
 const CONCURRENT_WARN_THRESHOLD_DEFAULT = 20;
 
+// Startup auto-retry (option a, 2026-06-04): a short breath before silently
+// re-dispatching a message whose first attempt died in the dev-channels startup
+// gate (TMUX_SESSION_GONE). Long enough that a host under momentary load isn't
+// hammered with a back-to-back respawn, short enough that a transient flake
+// still recovers fast enough to feel instant to the user.
+const STARTUP_RETRY_DELAY_MS = 1500;
+
 function createDispatcher({
   config,
   db,
@@ -48,6 +55,9 @@ function createDispatcher({
   // the historic 4096 for back-compat in synthetic test runs that pass
   // pre-formatted text.
   chunkBudget = 4096,
+  // Delay before a silent startup auto-retry re-dispatches (TMUX_SESSION_GONE).
+  // Injected so tests can drive it to 0; production uses STARTUP_RETRY_DELAY_MS.
+  startupRetryDelayMs = STARTUP_RETRY_DELAY_MS,
   // State accessors (need late binding because polygram.js mutates):
   getIsShuttingDown,              // () → boolean
   logger = console,
@@ -217,6 +227,35 @@ function createDispatcher({
       //   - shutting down ("Process killed" isn't a real error),
       //   - user just /stop'd (already saw their abort ack).
       if (!wasAborted && !isReplay && !isShuttingDown) {
+        // Startup auto-retry (option a, 2026-06-04). TMUX_SESSION_GONE = claude
+        // exited INSIDE the startup gate, before the dev-channels channel went
+        // live — so the user's message was NEVER delivered to claude. That makes
+        // a re-send idempotent BY CONSTRUCTION (unlike a mid-turn drop, where
+        // claude might still be slowly processing). The session_id was just
+        // poison-cleared above, so re-dispatching the SAME message spawns a FRESH
+        // session and delivers it. Silent: a transient startup flake (recurs
+        // ~once/9h on the channels backend) never reaches the user — instead of
+        // the "🔄 reset it, resend" papercut, polygram just retries. One-shot
+        // (_startupRetried) so a host that genuinely can't start claude surfaces
+        // the friendly reset reply (below) after EXACTLY one retry, never a loop.
+        // Scoped to TMUX_SESSION_GONE only: CHANNELS_DIALOG_TIMEOUT is a real
+        // blocking dialog (usage-limit / permission) a retry would just re-hit,
+        // so it keeps its "please resend" copy.
+        if (err.code === 'TMUX_SESSION_GONE' && !msg._startupRetried) {
+          logEvent('startup-auto-retry', {
+            chat_id: chatId, session_key: sessionKey, msg_id: msg?.message_id,
+          });
+          // Re-dispatch a COPY carrying the one-shot marker — never mutate the
+          // caller's msg (the boot-replay path shares/re-reads it). unref the
+          // best-effort timer so a pending retry can't pin the daemon alive
+          // (the Telegram long-poll already keeps the loop running).
+          const retryMsg = { ...msg, _startupRetried: true };
+          setTimeout(
+            () => dispatchHandleMessage(sessionKey, chatId, retryMsg, bot),
+            startupRetryDelayMs,
+          ).unref?.();
+          return;
+        }
         // rc.54: auto-resume on 300s no-activity timeout. The
         // resume turn itself runs through sendToProcess directly
         // (not handleMessage), so its errors don't re-enter this
@@ -309,4 +348,5 @@ function createDispatcher({
 module.exports = {
   createDispatcher,
   CONCURRENT_WARN_THRESHOLD_DEFAULT,
+  STARTUP_RETRY_DELAY_MS,
 };

package/lib/process/cli-process.js

@@ -115,13 +115,20 @@ const STREAMING_HINT_RE = /esc to interrupt/i;
 
 // 0.12.0 background-work lifecycle: claude's TUI mode line shows a live
 // background-shell COUNT while a `run_in_background:true` Bash outlives its turn,
-// e.g. `⏵⏵ auto mode on · 1 shell · ← for agents · ↓ to manage`. Confirmed on
-// claude 2.1.158 (P0 spike — docs/0.12.0-background-work-lifecycle-plan.md): the
-// count is always-present in the viewport mode line while shells run and clears
-// IN-PLACE within ~3s when they exit (no stale scrollback). Anchored to the
-// `auto mode on` line and matched only against the captured TAIL so a scrolled-off
-// history line never trips it. R1: re-validate on each pinned-claude bump.
-const BACKGROUND_SHELL_RE = /auto mode on[^\n]*·\s*(\d+)\s+shells?\b/i;
+// e.g. `⏵⏵ bypass permissions on · 1 shell · ← for agents · ↓ to manage`.
+// Confirmed on claude 2.1.158 (P0 spike — docs/0.12.0-background-work-lifecycle-
+// plan.md): the count is always-present in the viewport mode line while shells run
+// and clears IN-PLACE within ~3s when they exit (no stale scrollback).
+//
+// MODE-INDEPENDENT (prod regression fix, 2026-06-04): the original regex anchored
+// on "auto mode on", but EVERY shumorobot session runs "⏵⏵ bypass permissions on"
+// — the spike happened to be captured in auto mode. So the detector never matched
+// in prod and bg-work-status fired zero times. Anchor instead on the `⏵⏵` mode-
+// line glyph (present in auto / bypass / accept-edits modes alike); only the mode
+// label between it and `· N shell` varies. Still matched only against the captured
+// TAIL so a scrolled-off history line never trips it. R1: re-validate on each
+// pinned-claude bump (glyph + `N shell` wording).
+const BACKGROUND_SHELL_RE = /⏵⏵[^\n]*·\s*(\d+)\s+shells?\b/i;
 // How long a detached background shell may run AFTER its turn resolved (claude
 // idle) before the stall-watchdog fires one read-only self-check. Override via
 // the constructor (tests use a small value).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "polygram",
-  "version": "0.12.0-rc.24",
+  "version": "0.12.0-rc.26",
   "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": {