fix(telegram): remove the autonomous-silence notice — it only ever landed as a DM nag (DIVE-3910)

plugins/telegram/hooks/stop-reply-check.ts

@@ -40,7 +40,6 @@ import { sendMessage, getToken } from './lib/telegram'
 import { emitBlock } from './lib/output'
 import { TG_TOOL_PREFIX, typingStopFile } from './lib/paths'
 import { getAllowedChatIds, getCallerChat, getGroupTopics, type CallerChat } from './lib/access'
-import { nextSilentRun, composeSilentRunNotice } from './lib/autonomous-silence'
 import { parseResetEpoch } from './lib/time'
 import type { HookPayload, TranscriptEntry, TranscriptContentBlock } from './lib/types'
 
@@ -211,43 +210,11 @@ const a = analyzeTurn(entries, TG_TOOL_PREFIX)
 // already false for such a turn (the guard below would exit too), but assert
 // it explicitly so the golden tripwire can lock the invariant against a future
 // analyzeTurn refactor that might change turn-boundary detection.
-// DIVE-3422: BEFORE the a2a and inbound exits, because the case this covers
-// is the one that leaves through them. An autonomous session (cron, `/goal`,
-// any turn with no <channel> inbound) has no caller chat, so every branch
-// below correctly exits and NOTHING the agent writes reaches anyone — the
-// agent meanwhile reads its own transcript as evidence that it reported.
-// Track the run of consecutive silent turns and, past a threshold, speak once
-// into the agent's own group topic. Every arm of the decision is in
-// ./lib/autonomous-silence; the negative control (any turn that reaches the
-// channel resets the run to zero) lives there and is unit-locked.
-const silentRunFile = join(tmpdir(), `5dive-tg-silentrun-${lockKey}.count`)
-if (getToken()) {
-  const prevRun = (() => {
-    try { return parseInt(readFileSync(silentRunFile, 'utf8').trim(), 10) || 0 } catch { return 0 }
-  })()
-  const decision = nextSilentRun(prevRun, {
-    hadInbound: a.hadInbound,
-    a2aTurn: a.a2aTurn,
-    hadSend: a.hadSend,
-    hasText: a.texts.some(t => t.trim().length > 0),
-  })
-  if (decision.count !== prevRun) {
-    try { writeFileSync(silentRunFile, String(decision.count)) } catch { /* stale count only */ }
-  }
-  if (decision.notify) {
-    // The agent's own forum topic first, allowed chats only if no group is
-    // configured. An autonomous session reporting into its own topic is
-    // information; the same text in a human's DM is a nag, and a nag gets
-    // muted — which is the silence this exists to end.
-    const topics = getGroupTopics()
-    const targets: CallerChat[] = topics.length > 0
-      ? topics
-      : getAllowedChatIds().map(chatId => ({ chatId }))
-    const lastText = [...a.texts].reverse().find(t => t.trim().length > 0) ?? ''
-    const notice = composeSilentRunNotice(decision.count, lastText)
-    await Promise.all(targets.map(t => sendMessage(t.chatId, notice, t.threadId)))
-  }
-}
+// DIVE-3910: the autonomous-silence notice was REMOVED here (lodar, 2026-09-02:
+// "I feel this works wrongly ... I never seen its been useful ... it just confuses
+// clients"). It fell back to a human's DM whenever no group topic was configured,
+// which is every seat here, so the one shape its own comment called a nag was the
+// only shape it ever took. The agent-side silence-watchdog is untouched.
 
 // DIVE-1323 (continued): suppress the human-DM reflex on an inter-agent turn.
 if (a.a2aTurn) process.exit(0)