thinkpool-pair 0.7.66 → 0.7.68

package/bridge.mjs

@@ -39,7 +39,7 @@ import { createClient } from '@supabase/supabase-js'
 import { startClaudeSession } from './claude-session.mjs'
 import { turnInFlight } from './update-gate.mjs'
 import { saveSession, flushSession, deleteSession, loadAll, canResume, loadPtyId, savePtyId, loadNames, saveNames } from './session-store.mjs'
-import { stampEvent, makeSeqCounter, maxSeq, seqable } from './event-id.mjs'
+import { stampEvent, makeSeqCounter, maxSeq, seqable, capReplayEvents } from './event-id.mjs'
 import { makeThrottledTrack } from './presence.mjs'
 
 // Public client creds (the same anon values the web app ships — safe to embed).
@@ -454,8 +454,37 @@ const webPeerPresent = () => {
   return false
 }
 
+// Realtime broadcast silently DROPS frames over its size limit AND can break the
+// socket for the rest of the tick — the 2026-06-26 reload-freeze (a 5.4MB replay).
+// This is the universal backstop: ANY oversized frame (replay OR a giant live
+// tool_result/assistant blob) is (a) LOGGED loudly so it can never be a silent black
+// hole again, and (b) defanged by truncating the largest string field so the frame
+// still delivers instead of taking the channel down. capReplayEvents bounds bulk
+// replays before they reach here; this catches the single-event case + is the alarm.
+const FRAME_SOFT_LIMIT = 200000
+const defendFrame = (event, payload) => {
+  let json
+  try { json = JSON.stringify(payload) } catch { return payload }
+  if (!json || json.length <= FRAME_SOFT_LIMIT) return payload
+  process.stderr.write(`\n  ⚠ oversized ${event} frame: ${json.length}B (term ${String(payload?.term || '?').slice(0, 8)}) — truncating to fit so delivery isn't dropped\n`)
+  try {
+    const p = JSON.parse(json)
+    const evt = p.evt
+    if (evt && typeof evt === 'object') {
+      for (const k of ['text', 'content', 'output', 'stdout', 'stderr']) {
+        if (typeof evt[k] === 'string' && evt[k].length > 20000) {
+          const orig = evt[k].length
+          evt[k] = evt[k].slice(0, 20000) + `\n…[truncated ${orig - 20000} bytes — too large to broadcast]`
+        }
+      }
+    }
+    return p
+  } catch { return payload }
+}
+
 const bcast = (event, payload) => {
   if (event === 'pty-out' || event === 'code-event') lastActivity = Date.now()
+  payload = defendFrame(event, payload)
   try {
     if (channel.channelAdapter?.canPush?.() ?? true) {
       channel.send({ type: 'broadcast', event, payload })
@@ -975,9 +1004,11 @@ channel
     flushAll()
     for (const [id, t] of terms) {
       if (!t.scrollback) continue
+      // Cap scrollback too — a huge PTY buffer would blow the same frame limit.
+      const sb = t.scrollback.length > 100000 ? t.scrollback.slice(-100000) : t.scrollback
       bcast('pty-replay', {
         to: payload?.to ?? null, term: id,
-        b64: Buffer.from(t.scrollback, 'utf8').toString('base64'),
+        b64: Buffer.from(sb, 'utf8').toString('base64'),
       })
     }
     // Structured sessions replay their event log (reader rebuilds from it).
@@ -990,7 +1021,10 @@ channel
     for (const [id, s] of sessions) {
       if (!s.log.length) continue
       const from = Number(payload?.cursors?.[id]) || 0
-      const events = from > 0 ? s.log.filter((e) => e.seq == null || e.seq > from) : s.log
+      const tail = from > 0 ? s.log.filter((e) => e.seq == null || e.seq > from) : s.log
+      // Cap to a frame-safe size — the full log on a fresh-reload (empty cursor) is
+      // multi-MB and gets dropped by Realtime, freezing the terminal. See capReplayEvents.
+      const events = capReplayEvents(tail)
       bcast('code-replay', { to: payload?.to ?? null, term: id, events })
     }
     // Re-send any still-pending permission/question cards. They ride a one-shot

package/event-id.mjs

@@ -67,3 +67,29 @@ export function maxSeq(events) {
   for (const e of events) if (e && typeof e.seq === 'number' && e.seq > m) m = e.seq
   return m
 }
+
+// Cap a replay event list to a byte budget so the broadcast frame fits under the
+// Realtime size limit. THE BUG (2026-06-26, room XE3IQ): on a fresh reload the client
+// sends replay-request with an EMPTY cursor (seqHi=0), so the bridge replied with the
+// ENTIRE per-term log — multi-thousand events / multiple MB for an established terminal.
+// That frame exceeds Realtime's broadcast size limit and is silently DROPPED (and breaks
+// the socket for the rest of the tick, taking small terminals' replays + the CHANNEL_ERROR
+// churn down with it). The terminal then never refills and every live event is gap-dropped
+// → "messages vanish, agent never responds, only brand-new terminals work".
+// Keep the MOST RECENT events that fit: the client's DB transcript snapshot (code_messages
+// tx:, last 200) already renders older history, and seqDedup.classifyReplay explicitly
+// accepts the seq eviction-jump, so the client's live cursor (seqHi) still reaches current.
+// Empirically a ~266KB frame delivered while multi-MB did not; 150KB is a safe margin.
+// Always keeps at least the newest event so seqHi can advance even if it alone is oversized.
+export const REPLAY_BYTE_BUDGET = 150000
+export function capReplayEvents(events, budget = REPLAY_BYTE_BUDGET) {
+  if (!Array.isArray(events) || events.length === 0) return events
+  let total = 0, i = events.length
+  while (i > 0) {
+    const sz = JSON.stringify(events[i - 1]).length + 1
+    if (i < events.length && total + sz > budget) break
+    total += sz
+    i--
+  }
+  return i === 0 ? events : events.slice(i)
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "thinkpool-pair",
-  "version": "0.7.66",
+  "version": "0.7.68",
   "description": "Share a local coding-agent CLI (Claude Code, Codex, Gemini, Aider, …) into a ThinkPool Code room, live.",
   "type": "module",
   "bin": {