polygram 0.7.0 → 0.7.2

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://anthropic.com/claude-code/plugin.schema.json",
   "name": "polygram",
-  "version": "0.7.0",
+  "version": "0.7.2",
   "description": "Telegram integration for Claude Code that preserves the OpenClaw per-chat session model. Migration target for OpenClaw users. Multi-bot, multi-chat, per-topic isolation; SQLite transcripts; inline-keyboard approvals. Bundles /polygram:status|logs|pair-code|approvals admin commands and a history skill.",
   "keywords": [
     "telegram",

package/README.md

@@ -364,7 +364,7 @@ foreign-chat clicks are rejected. Default-deny on IPC error.
 ## Development
 
 ```bash
-npm test           # 619 tests, 153 suites, node:test, no external services
+npm test           # 643 tests, 158 suites, node:test, no external services
 npm run coverage   # native test coverage (Node 22+, no devDeps)
 npm start -- --bot my-bot
 npm run split-db -- --config config.json --dry-run

package/lib/announces.js

@@ -12,28 +12,102 @@
  * or `config.chats.<id>.announceSubagents`), so existing chats see
  * no behavior change.
  *
- * Note: this is the minimal MVP of OpenClaw's full subagent-announce
- * queue (which has debounce, drop policies, multi-channel routing).
- * Polygram's "subagents" are in-process (Claude Code Task tool), so
- * the announce path is just a one-shot informational sendMessage.
+ * 0.7.1 redesign: factory-based with split read/write predicates
+ * (canAnnounce / markAnnounced) and lazy GC. Pre-0.7.1 had a
+ * module-scoped Map and a mutate-on-check `shouldAnnounce` predicate
+ * — both anti-patterns flagged in design review. The free-function
+ * API (`shouldAnnounce`, `announce`) is preserved for back-compat,
+ * delegating to a default singleton.
  */
 
 const SUBAGENT_DEBOUNCE_MS = 30_000;
 
 /**
- * Per-chat debounce so a turn that spawns 5 subagents back-to-back
- * doesn't post 5 announces. Module-scoped Map keyed by chatId →
- * timestamp of last announce.
+ * Per-chat debounce tracker. Returns:
+ *   - canAnnounce(chatId): true if this chat hasn't announced within
+ *     the debounce window. Pure read, NO mutation — safe for
+ *     speculative checks.
+ *   - markAnnounced(chatId): records `now` as the last announce time
+ *     for this chat. Caller invokes after a successful send.
+ *   - sweep(): drops entries older than `2 * debounceMs`. Called lazily
+ *     on every canAnnounce check past a soft threshold.
+ *   - size(): for tests / diagnostics.
+ *   - clear(): for test isolation.
  */
-const lastAnnounceByChat = new Map();
+function createAnnouncer({
+  debounceMs = SUBAGENT_DEBOUNCE_MS,
+  clock = Date.now,
+  sweepThreshold = 1000,
+} = {}) {
+  const lastAnnounceByChat = new Map();
 
+  function key(chatId) { return String(chatId); }
+
+  function sweep() {
+    const cutoff = clock() - 2 * debounceMs;
+    for (const [k, ts] of lastAnnounceByChat) {
+      if (ts < cutoff) lastAnnounceByChat.delete(k);
+    }
+  }
+
+  function canAnnounce(chatId) {
+    if (lastAnnounceByChat.size > sweepThreshold) sweep();
+    const prev = lastAnnounceByChat.get(key(chatId));
+    return prev == null || (clock() - prev) >= debounceMs;
+  }
+
+  function markAnnounced(chatId) {
+    lastAnnounceByChat.set(key(chatId), clock());
+  }
+
+  return {
+    canAnnounce, markAnnounced, sweep,
+    get size() { return lastAnnounceByChat.size; },
+    clear() { lastAnnounceByChat.clear(); },
+  };
+}
+
+// Default per-process state for the back-compat free-function API.
+// Pre-0.7.1, this was the only API. Long-running daemons should still
+// prefer createAnnouncer() for tests / multi-bot isolation, but
+// polygram.js's single-bot-per-process model means the singleton works
+// fine for the production path. The Map is pruned lazily inside
+// shouldAnnounce when it grows past the sweep threshold.
+const _defaultLastAnnouncements = new Map();
+const _DEFAULT_SWEEP_THRESHOLD = 1000;
+
+/**
+ * Back-compat: pre-0.7.1 callers used `shouldAnnounce(chatId, now?,
+ * debounceMs?)` which is a "predicate that mutates" — call site is
+ * `if (shouldAnnounce(id)) await sendAnnounce()`. The mutation happens
+ * eagerly. Preserved verbatim for callers; new code should use
+ * `createAnnouncer()` and the explicit canAnnounce/markAnnounced split.
+ *
+ * 0.7.1: added lazy sweep so the Map doesn't grow unbounded over a
+ * multi-week-uptime daemon.
+ */
 function shouldAnnounce(chatId, now = Date.now(), debounceMs = SUBAGENT_DEBOUNCE_MS) {
-  const prev = lastAnnounceByChat.get(String(chatId));
+  if (_defaultLastAnnouncements.size > _DEFAULT_SWEEP_THRESHOLD) {
+    const cutoff = now - 2 * debounceMs;
+    for (const [k, ts] of _defaultLastAnnouncements) {
+      if (ts < cutoff) _defaultLastAnnouncements.delete(k);
+    }
+  }
+  const key = String(chatId);
+  const prev = _defaultLastAnnouncements.get(key);
   if (prev != null && now - prev < debounceMs) return false;
-  lastAnnounceByChat.set(String(chatId), now);
+  _defaultLastAnnouncements.set(key, now);
   return true;
 }
 
+/**
+ * Reset the default singleton state (for tests). Not exported in
+ * production docs.
+ */
+function _resetDefaultAnnouncerForTests() {
+  _defaultLastAnnouncements.clear();
+}
+
 /**
  * Send a plain-text announce (no markdown processing, no reply linkage).
  * Caller passes `tg(bot, method, params, meta)` as `send` so we don't
@@ -67,4 +141,10 @@ async function announce({
   }
 }
 
-module.exports = { announce, shouldAnnounce, SUBAGENT_DEBOUNCE_MS };
+module.exports = {
+  announce,
+  shouldAnnounce,
+  createAnnouncer,
+  SUBAGENT_DEBOUNCE_MS,
+  _resetDefaultAnnouncerForTests,
+};

package/lib/deliver.js

@@ -37,8 +37,13 @@ async function deliverReplies({
   logger = console,
 }) {
   if (!Array.isArray(chunks) || chunks.length === 0) {
-    return { sent: [], failed: [] };
+    return { sent: [], failed: [], results: [] };
   }
+  // 0.7.1: results[] preserves correspondence with chunks[] — results[i]
+  // describes what happened to chunks[i]. sent[]/failed[] are projections
+  // for back-compat with callers that already use them; they no longer
+  // ambiguously mean "the i-th success/failure" vs "chunk i's outcome".
+  const results = [];
   const sent = [];
   const failed = [];
   for (let i = 0; i < chunks.length; i++) {
@@ -56,14 +61,18 @@ async function deliverReplies({
     try {
       const res = await send(bot, 'sendMessage', params, meta);
       const msgId = res?.message_id ?? null;
+      results.push({ index: i, status: 'ok', messageId: msgId });
       sent.push(msgId);
     } catch (err) {
       logger.error?.(`[deliver] chunk ${i + 1}/${chunks.length} failed: ${err.message}`);
+      results.push({ index: i, status: 'fail', error: err.message });
       failed.push({ index: i, error: err.message });
-      // Keep going — partial delivery is better than total loss.
+      // Keep going — partial delivery is better than total loss. Caller
+      // should inspect failed.length and surface a warning to the user
+      // (see polygram.js handleMessage's stream-redeliver event log).
     }
   }
-  return { sent, failed };
+  return { sent, failed, results };
 }
 
 module.exports = { deliverReplies };

package/lib/sent-cache.js

@@ -17,13 +17,54 @@
 
 const TTL_MS = 24 * 60 * 60 * 1000;
 const CLEANUP_THRESHOLD = 100;
+// 0.7.1: hard cap on per-chat Map size. CLEANUP_THRESHOLD only drops
+// EXPIRED entries — if a busy chat sends >100 fresh messages within
+// 24h, GC finds nothing to drop and the inner Map grows unbounded.
+// Cap evicts oldest entries past this point regardless of TTL.
+const MAX_PER_CHAT = 500;
+// 0.7.1: outer Map sweep — drop chats whose inner Map has been empty
+// long enough that we're sure no live caller still references it.
+const OUTER_SWEEP_THRESHOLD = 1000;
 
-function createSentCache() {
+function createSentCache({
+  ttlMs = TTL_MS,
+  cleanupThreshold = CLEANUP_THRESHOLD,
+  maxPerChat = MAX_PER_CHAT,
+  outerSweepThreshold = OUTER_SWEEP_THRESHOLD,
+  clock = Date.now,
+} = {}) {
   // chatKey → Map<msgId, ts>
   const sentMessages = new Map();
 
   function chatKey(chatId) { return String(chatId); }
 
+  function gcInner(entry) {
+    const cutoff = clock() - ttlMs;
+    for (const [id, ts] of entry) if (ts < cutoff) entry.delete(id);
+    // After TTL prune, if still over the hard cap, drop oldest entries
+    // (insertion order in Map iteration). This handles the busy-chat
+    // case where 1000 messages all sent within 24h would otherwise
+    // leak.
+    if (entry.size > maxPerChat) {
+      const dropCount = entry.size - maxPerChat;
+      let i = 0;
+      for (const id of entry.keys()) {
+        if (i >= dropCount) break;
+        entry.delete(id);
+        i += 1;
+      }
+    }
+  }
+
+  function gcOuter() {
+    // Drop chat entries that are entirely empty (their inner Map was
+    // drained by gcInner). Without this the outer Map's chatId set
+    // grows by one per ever-active-then-idle chat, slowly leaking.
+    for (const [k, entry] of sentMessages) {
+      if (entry.size === 0) sentMessages.delete(k);
+    }
+  }
+
   function record(chatId, messageId) {
     if (chatId == null || messageId == null) return;
     const key = chatKey(chatId);
@@ -32,13 +73,10 @@ function createSentCache() {
       entry = new Map();
       sentMessages.set(key, entry);
     }
-    entry.set(messageId, Date.now());
-    // Lazy GC: when the per-chat map gets crowded, drop expired
-    // entries. Cheap (O(n) over n ≤ 100 + a bit) and amortises to O(1).
-    if (entry.size > CLEANUP_THRESHOLD) {
-      const cutoff = Date.now() - TTL_MS;
-      for (const [id, ts] of entry) if (ts < cutoff) entry.delete(id);
-    }
+    entry.set(messageId, clock());
+    if (entry.size > cleanupThreshold) gcInner(entry);
+    // Periodic outer sweep — runs only when the outer Map gets crowded.
+    if (sentMessages.size > outerSweepThreshold) gcOuter();
   }
 
   function wasSent(chatId, messageId) {
@@ -48,8 +86,10 @@ function createSentCache() {
     if (!entry) return false;
     const ts = entry.get(messageId);
     if (ts == null) return false;
-    if (Date.now() - ts > TTL_MS) {
+    if (clock() - ts > ttlMs) {
       entry.delete(messageId);
+      // If we just emptied the inner Map, drop the outer entry too.
+      if (entry.size === 0) sentMessages.delete(key);
       return false;
     }
     return true;
@@ -61,11 +101,19 @@ function createSentCache() {
     return total;
   }
 
+  function chatCount() { return sentMessages.size; }
+
   function clear() {
     sentMessages.clear();
   }
 
-  return { record, wasSent, size, clear };
+  return { record, wasSent, size, chatCount, clear };
 }
 
-module.exports = { createSentCache, TTL_MS, CLEANUP_THRESHOLD };
+module.exports = {
+  createSentCache,
+  TTL_MS,
+  CLEANUP_THRESHOLD,
+  MAX_PER_CHAT,
+  OUTER_SWEEP_THRESHOLD,
+};

package/lib/stream-reply.js

@@ -1,24 +1,31 @@
 /**
  * Live streaming-reply state machine for a single turn.
  *
- * Lifecycle (0.7.0):
+ * Lifecycle:
  *   idle  -> (text >= minChars) -> live
  *   live  -> (subsequent chunks) -> live       (throttled edits)
+ *   live  -> flushDraft()         -> live      (drains pending edit)
  *   live  -> forceNewMessage()    -> idle      (next chunk = new bubble)
  *   live  -> discard()            -> finalized (bubble deleted)
  *   any   -> finalize(finalText)  -> finalized
  *
- * The streamer never talks to Telegram directly — callers inject `send(text)`,
- * `edit(msg_id, text)`, and (new in 0.7.0) optional `deleteMessage(msg_id)`.
+ * The streamer never talks to Telegram directly — callers inject
+ * `send(text)`, `edit(msg_id, text)`, and (optional) `deleteMessage(msg_id)`.
  * That keeps polygram.js in charge of transcript writes, sticker/reaction
  * routing, and error handling; this module is just a cadence machine.
  *
- * 0.7.0 finalize() returns rich result so the caller can decide whether the
- * preview's last edit IS the final reply, or whether to discard the preview
- * and redeliver via deliverReplies (overflow / final edit failed). This is
- * the OpenClaw pattern: short replies preview-becomes-final (no flicker),
- * long replies preview-deleted-redelivered (single coherent bubble flow at
- * chat bottom).
+ * `finalize()` returns a rich result so the caller can decide whether the
+ * preview's last edit IS the final reply, or whether to discard the
+ * preview and redeliver via deliverReplies (overflow / final edit failed):
+ *
+ *   { kind: implicit, see flags below }
+ *   { streamed: false }                                  — never went live
+ *   { streamed: true, finalEditOk: true }                — preview = final
+ *   { streamed: true, finalEditOk: false, overflow: true } — body too long
+ *   { streamed: true, finalEditOk: false, overflow: false } — edit failed
+ *
+ * Short replies preview-becomes-final (no flicker, single bubble); long
+ * replies preview-deleted-redelivered (chunks land at chat bottom).
  *
  * Test-friendly: inject `clock` (now() fn) and `schedule` (setTimeout-like)
  * so a fake clock can drive throttle timing deterministically.
@@ -50,12 +57,19 @@ function createStreamer({
   let lastEditTs = 0;
   let pendingEdit = null;   // timer id
   let flushPromise = null;  // ongoing edit promise (for back-pressure)
+  // 0.7.2: msg_ids of bubbles that have been superseded by
+  // forceNewMessage(). The caller (polygram.js handleMessage at
+  // end-of-turn) reads getArchived() and issues deleteMessage on
+  // each — matches OpenClaw's archivedAnswerPreviews cleanup so
+  // the user sees only the final answer's bubble, not every
+  // "thinking out loud" intermediate from a tool-heavy turn.
+  const archived = [];
 
-  // 0.7.0: this is the LIVE-EDIT truncation, used during streaming
-  // when latestText overshoots maxLen. The trailing "..." signals to
-  // the user that more is coming. At finalize time, we DON'T truncate
-  // — we either edit-to-final-as-is (caller already chunked correctly)
-  // or signal overflow back to the caller.
+  // LIVE-EDIT truncation only — used during streaming when latestText
+  // overshoots maxLen. The trailing "..." signals to the user that more
+  // is coming. Finalize doesn't truncate: overflow is handled by
+  // signalling the caller to discard-and-redeliver via chunkMarkdownText,
+  // which preserves all content without any byte-cut.
   function truncateForLive(s) {
     if (s.length <= maxLen) return s;
     return s.slice(0, maxLen - 3) + '...';
@@ -132,15 +146,20 @@ function createStreamer({
     if (flushPromise) { try { await flushPromise; } catch {} }
   }
 
-  // 0.7.0: reset bubble state so the next onChunk creates a NEW message.
-  // Used by the upcoming Phase 7 F (forceNewMessage on assistant-
-  // message-start) — when Claude emits a new top-level assistant message
-  // mid-turn (post tool-result), we want it in its own bubble below
-  // the previous one, not appended via edit.
+  // Reset bubble state so the next onChunk creates a NEW message.
+  // Used by `onAssistantMessageStart` in process-manager.js when Claude
+  // emits a new top-level assistant message mid-turn (post tool-result):
+  // we want it in its own bubble below the previous one, not appended
+  // via editMessageText to the original.
   function forceNewMessage() {
     if (pendingEdit) { cancel(pendingEdit); pendingEdit = null; }
     // Don't await flushPromise — the caller has decided to start a new
     // message; whatever the old bubble shows is "done".
+    // 0.7.2: track the previous bubble's msgId for end-of-turn cleanup.
+    // Without this, every intermediate "thinking out loud" assistant
+    // message in a tool-heavy turn leaves a permanent bubble in the
+    // chat — the user wants only the final answer's bubble visible.
+    if (msgId != null) archived.push(msgId);
     msgId = null;
     currentText = '';
     latestText = '';
@@ -232,6 +251,11 @@ function createStreamer({
     }
   }
 
+  // 0.7.2: snapshot of bubble msgIds that forceNewMessage() superseded.
+  // Returns a copy so callers can't mutate internal state. polygram.js
+  // reads this at end-of-turn and issues deleteMessage on each.
+  function getArchived() { return archived.slice(); }
+
   return {
     onChunk,
     finalize,
@@ -239,6 +263,7 @@ function createStreamer({
     forceNewMessage,
     discard,
     archive,
+    getArchived,
     // Introspection for tests:
     get state() { return state; },
     get msgId() { return msgId; },

package/lib/telegram-chunk.js

@@ -1,24 +1,26 @@
 /**
  * Markdown-aware chunking for Telegram-bound replies.
  *
- * Direct port of OpenClaw's chunkMarkdownText (`extensions/telegram` uses
- * `chunkerMode: 'markdown'`). The naive byte-cut chunker we shipped pre-0.7.0
- * landed boundaries mid-word and mid-HTML-tag, which Telegram's parse_mode=HTML
- * rejected with `400 can't parse entities` — bubbles froze and content got
- * dropped (see msg 10794 incident). This algorithm guarantees:
+ * Direct port of OpenClaw's chunkMarkdownText. The naive byte-cut
+ * chunker we shipped before this would land boundaries mid-word and
+ * mid-HTML-tag, which Telegram's parse_mode=HTML rejected with
+ * `400 can't parse entities` — bubbles froze and content got dropped.
+ *
+ * Guarantees:
  *
  *   1. No chunk exceeds `limit`.
  *   2. Breaks prefer newlines over whitespace over hard-cut.
- *   3. Code fences (```...```) are never broken silently — if a chunk would
- *      land inside a fence, we close it on chunk N and re-open with the same
- *      marker + language on chunk N+1, so each chunk is independently
- *      parseable.
- *   4. Parenthesised expressions `(...)` aren't broken at whitespace inside
- *      the parens (avoids splitting `[markdown link](http://example.com/...)`).
+ *   3. Code fences (```...```) are never broken silently — if a chunk
+ *      would land inside a fence, we close it on chunk N and re-open
+ *      with the same marker + language on chunk N+1, so each chunk is
+ *      independently parseable.
+ *   4. Parenthesised expressions `(...)` aren't broken at whitespace
+ *      inside the parens (avoids splitting markdown-link syntax like
+ *      `[label](http://example.com/...)`).
  *
- * Plain `chunkText` (no fence handling) is exported for callers that already
- * know the input has no markdown — primarily code paths handling raw user
- * input echoes or non-text payloads.
+ * Plain `chunkText` (no fence handling) is exported for callers that
+ * already know the input has no markdown — primarily code paths
+ * handling raw user input echoes or non-text payloads.
  */
 
 // ─── Code-fence span detection ──────────────────────────────────────
@@ -105,11 +107,19 @@ function scanParenAwareBreakpoints(window, isAllowed = () => true) {
 
 // ─── Chunkers ────────────────────────────────────────────────────────
 
-// Common early-out: empty / non-positive limit / fits-in-one returns
-// directly so the loop bodies can assume there's real work to do.
+// Common early-out: empty / fits-in-one returns directly so the loop
+// bodies can assume there's real work to do. `limit ≤ 0` is treated as
+// a programmer error and throws — silently returning [text] would let
+// a misread config pass through a body that exceeds Telegram's actual
+// 4096-char cap, which the chunker exists to prevent.
 function resolveChunkEarlyReturn(text, limit) {
-  if (!text) return [];
-  if (limit <= 0) return [text];
+  if (typeof limit !== 'number' || !Number.isFinite(limit) || limit <= 0) {
+    throw new RangeError(`chunk limit must be a positive number; got ${limit}`);
+  }
+  if (text == null || text === '') return [];
+  if (typeof text !== 'string') {
+    throw new TypeError(`chunk text must be a string; got ${typeof text}`);
+  }
   if (text.length <= limit) return [text];
   return undefined;
 }
@@ -119,7 +129,7 @@ function resolveChunkEarlyReturn(text, limit) {
 // Negative / out-of-range break indices fall back to hard-cut at limit.
 function chunkTextByBreakResolver(text, limit, resolveBreakIndex) {
   if (!text) return [];
-  if (limit <= 0 || text.length <= limit) return [text];
+  if (text.length <= limit) return [text];
   const chunks = [];
   let remaining = text;
   while (remaining.length > limit) {

package/lib/telegram.js

@@ -173,20 +173,32 @@ async function send({ bot, method, params, db = null, meta = {}, logger = consol
   let res;
   const rawAttempt = async (p) => bot.api.raw[method](p);
 
-  // safeAttempt wraps every API call with three OpenClaw-style fallbacks:
-  //   1. MESSAGE_NOT_MODIFIED on editMessageText → swallow as success.
-  //      The streamer's debounced edit can land on text that exactly
-  //      matches the bubble's current state (no-op edit). Telegram
-  //      returns 400; we treat it as success and skip the noise.
-  //   2. HTML parse error (`can't parse entities` etc) → retry the
-  //      same call as plain text, no parse_mode, original raw value
-  //      restored to the formatted field. Saves the call when our
-  //      markdown→HTML conversion produces malformed output (the
-  //      msg-10794 case: streamer truncate cut mid `**bold**` marker).
-  //   3. 429 rate limit → sleep retry_after seconds, retry once.
-  //      Telegram's per-bot limit is ~30 req/s; high-effort xhigh turns
-  //      with many parallel sessions can occasionally hit it. Honor
-  //      Telegram's hint rather than bombing the call.
+  // OpenClaw-style fallback layers, composing outermost-to-innermost:
+  //
+  //   withThreadFallback (outer)  — strips message_thread_id and
+  //                                 retries on TOPIC_DELETED / "thread
+  //                                 not found"
+  //   withPreConnectRetry         — single retry on transient pre-
+  //                                 connect errors (DNS flap, TCP
+  //                                 refused, ENETUNREACH); never
+  //                                 retries post-connect errors that
+  //                                 might have landed
+  //   safeAttempt                 — sleeps `retry_after` and retries
+  //                                 once on 429
+  //   tryOnce (innermost)         — handles two per-call recoveries:
+  //                                 (a) MESSAGE_NOT_MODIFIED on
+  //                                 editMessageText → synthetic
+  //                                 success (streamer debounce often
+  //                                 lands on no-op edits, Telegram
+  //                                 returns 400 we don't want to
+  //                                 propagate); (b) HTML parse error
+  //                                 → retry as plain text with the
+  //                                 raw pre-conversion field value
+  //                                 restored
+  //   rawAttempt                  — bot.api.raw[method](params)
+  //
+  // Each layer is a closure built per call; allocation cost is
+  // negligible vs. network RTT.
   const RETRY_AFTER_CAP_MS = 60_000;
   const tryOnce = async (p) => {
     try {
@@ -209,7 +221,20 @@ async function send({ bot, method, params, db = null, meta = {}, logger = consol
         const plainParams = { ...p };
         delete plainParams.parse_mode;
         plainParams[formatField] = rawFieldValue;
-        return await rawAttempt(plainParams);
+        try {
+          return await rawAttempt(plainParams);
+        } catch (plainErr) {
+          // 0.7.1: if the plain retry also fails, preserve BOTH errors
+          // in the message that propagates. Pre-fix, only `plainErr`
+          // bubbled up — operators investigating from markOutboundFailed
+          // saw e.g. "Forbidden: bot was kicked" and missed that the
+          // ORIGINAL failure was a markdown→HTML parse bug.
+          const origMsg = redactBotToken(err.message)?.slice(0, 200);
+          const wrapped = new Error(`plain-retry failed (after HTML parse error: ${origMsg}): ${plainErr.message}`);
+          if (plainErr.code) wrapped.code = plainErr.code;
+          if (plainErr.parameters) wrapped.parameters = plainErr.parameters;
+          throw wrapped;
+        }
       }
       throw err;
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "polygram",
-  "version": "0.7.0",
+  "version": "0.7.2",
   "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": {

package/polygram.js

@@ -1664,21 +1664,34 @@ async function handleMessage(sessionKey, chatId, msg, bot) {
     ...(linkPreview === false ? { linkPreview: false } : {}),
   };
 
+  // 0.7.2: only the FIRST bubble in a turn quotes the user's message
+  // via reply_parameters. When a tool-heavy turn produces multiple
+  // assistant messages (each spawning its own bubble via
+  // forceNewMessage), subsequent bubbles shouldn't re-quote the user
+  // — the chat would show N copies of the same quoted message stacked
+  // vertically. After the first send, the flag flips and subsequent
+  // initial-sends omit reply_parameters.
+  let firstBubbleSent = false;
   // Streaming is unconditional as of 0.4.0 — matches OpenClaw's model and
   // eliminates the "stuck at 15min typing" complaint from the non-streaming
   // code path. For short responses the streamer stays idle and we fall
   // through to the normal send path via finalize() returning streamed=false.
   const streamer = createStreamer({
-    send: async (text) => tg(bot, 'sendMessage', {
-      chat_id: chatId, text,
-      // allow_sending_without_reply: long-running turns give the user
-      // plenty of time to delete their original message. Without this
-      // flag, Telegram rejects the reply with MESSAGE_NOT_FOUND and the
-      // whole streamed answer is lost. With it, the reply simply lands
-      // as a standalone message.
-      reply_parameters: { message_id: msg.message_id, allow_sending_without_reply: true },
-      ...(threadId && { message_thread_id: threadId }),
-    }, outMetaBase),
+    send: async (text) => {
+      const params = {
+        chat_id: chatId, text,
+        ...(threadId && { message_thread_id: threadId }),
+      };
+      if (!firstBubbleSent) {
+        // allow_sending_without_reply: long-running turns give the user
+        // plenty of time to delete their original message. Without this
+        // flag, Telegram rejects the reply with MESSAGE_NOT_FOUND and the
+        // whole streamed answer is lost.
+        params.reply_parameters = { message_id: msg.message_id, allow_sending_without_reply: true };
+        firstBubbleSent = true;
+      }
+      return tg(bot, 'sendMessage', params, outMetaBase);
+    },
     edit: async (messageId, text) => {
       try {
         // Route edits through tg() so applyFormatting runs (MarkdownV2
@@ -1725,6 +1738,34 @@ async function handleMessage(sessionKey, chatId, msg, bot) {
   });
   // streamer is registered with this turn via pm.send's context (below)
 
+  // 0.7.2: clean up bubbles superseded by forceNewMessage() — the
+  // intermediate "thinking out loud" assistant messages that fired in
+  // a tool-heavy turn. Without this, every tool-result cycle leaves a
+  // permanent bubble in the chat (see the screenshot from the post-
+  // 0.7.1 deploy where six bubbles appeared for one logical turn).
+  // Matches OpenClaw's archivedAnswerPreviews end-of-turn cleanup.
+  // Call AFTER finalize/discard decisions so we never delete the
+  // bubble that's the final reply.
+  async function cleanupArchivedBubbles() {
+    const archived = streamer.getArchived?.() || [];
+    if (archived.length === 0) return;
+    for (const messageId of archived) {
+      try {
+        await tg(bot, 'deleteMessage', {
+          chat_id: chatId, message_id: messageId,
+        }, { source: 'bot-reply-archived-cleanup', botName: BOT_NAME });
+      } catch (err) {
+        // Non-fatal — message may be >48h old or already gone.
+        // Operator-visible only via the events table.
+        console.error(`[${label}] archived-cleanup ${messageId}: ${err.message}`);
+      }
+    }
+    logEvent('telegram-archived-cleanup', {
+      chat_id: chatId, msg_id: msg.message_id, count: archived.length,
+      bot: BOT_NAME,
+    });
+  }
+
   // Status reactions on the user's message: 👀 queued → 🤔 thinking →
   // 👨‍💻 coding / ⚡ web / 🔥 tool → 👍 done / 🤯 error. Silent (no
   // notifications), updates in place, one emoji per message. Uses
@@ -1748,12 +1789,13 @@ async function handleMessage(sessionKey, chatId, msg, bot) {
   // at which point we flip to THINKING (🤔).
   reactor.setState('QUEUED');
 
-  // Mark the inbound row terminal so boot replay doesn't pick it up again.
-  // Must fire down EVERY non-throwing exit path (early returns for error/
-  // NO_REPLY, streamed-reply early return, regular reply at end). 0.5.4
-  // hardened this — earlier versions only marked at the bottom of try, so
-  // streamed replies (which return at line ~1477) left handler_status
-  // stuck at 'dispatched' forever, causing replay loops on every restart.
+  // Mark the inbound row terminal so boot replay doesn't pick it up
+  // again. Must fire down EVERY non-throwing exit path (early returns
+  // for error / NO_REPLY, streamed-reply preview-becomes-final, the
+  // discard+redeliver branch, regular reply at end). Earlier versions
+  // only marked at the bottom of try, so streamed-reply early returns
+  // left handler_status stuck at 'dispatched' forever and the next
+  // boot replayed every long turn.
   const markReplied = () => dbWrite(() => db.setInboundHandlerStatus({
     chat_id: chatId, msg_id: msg.message_id, status: 'replied',
   }), 'set handler_status=replied');
@@ -1800,6 +1842,13 @@ async function handleMessage(sessionKey, chatId, msg, bot) {
     // those still markReplied silently.
     if (result.text === 'NO_REPLY') { markReplied(); return; }
     if (!result.text) {
+      // 0.7.1: if the fallback send itself fails, throw rather than
+      // silently markReplied — the user gets nothing AND the inbound
+      // is marked replied so boot replay won't redispatch. Same
+      // anti-pattern that caused msg-10794. Promote to a thrown error
+      // so dispatchHandleMessage's catch branches correctly:
+      //   shutdown   → 'replay-pending' (boot replay retries)
+      //   runtime    → 'failed' + user-visible apology via errorReplyText
       try {
         await tg(bot, 'sendMessage', {
           chat_id: chatId,
@@ -1808,7 +1857,12 @@ async function handleMessage(sessionKey, chatId, msg, bot) {
           reply_parameters: { message_id: msg.message_id, allow_sending_without_reply: true },
         }, { ...outMetaBase, source: 'empty-response-fallback' });
       } catch (err) {
-        console.error(`[${label}] empty-response fallback send failed: ${err.message}`);
+        reactor.setState('ERROR');
+        logEvent('telegram-empty-response-fallback-failed', {
+          chat_id: chatId, msg_id: msg.message_id, bot: BOT_NAME,
+          error: err.message?.slice(0, 200),
+        });
+        throw new Error(`empty-response fallback send failed: ${err.message}`);
       }
       logEvent('telegram-empty-response-fallback', {
         chat_id: chatId, msg_id: msg.message_id, bot: BOT_NAME,
@@ -1820,7 +1874,7 @@ async function handleMessage(sessionKey, chatId, msg, bot) {
     const parsed = parseResponse(result.text);
     const outMeta = { ...outMetaBase, sessionId: result.sessionId, costUsd: result.cost };
 
-    // 0.7.0 streamed text path: OpenClaw's preview-becomes-final flow.
+    // OpenClaw's preview-becomes-final flow:
     //
     //   1. flushDraft() — drain any pending throttled edit so the
     //      bubble's visible state is up-to-date before deciding.
@@ -1828,12 +1882,11 @@ async function handleMessage(sessionKey, chatId, msg, bot) {
     //      final body. Returns rich result describing whether the
     //      preview can stand as the final reply.
     //   3a. finalEditOk:true        → preview IS final, done.
-    //   3b. overflow OR !finalEditOk → discard preview, redeliver via
-    //      deliverReplies(chunkMarkdownText(...)). This is the path
-    //      that fixes msg-10794: if the live bubble couldn't render
-    //      the full body (size or parse error), we delete it cleanly
-    //      and send the proper chunks fresh at chat bottom — no
-    //      content lost, no stranded edit-failure bubble.
+    //   3b. overflow OR !finalEditOk → discard preview, redeliver
+    //      via deliverReplies(chunkMarkdownText(...)). The bubble
+    //      couldn't render the full body (size or parse error), so
+    //      we delete it cleanly and send the proper chunks fresh at
+    //      chat bottom — no content lost, no stranded bubble.
     if (parsed.text) {
       await streamer.flushDraft();
       const fin = await streamer.finalize(parsed.text);
@@ -1841,6 +1894,7 @@ async function handleMessage(sessionKey, chatId, msg, bot) {
         if (fin.finalEditOk) {
           // Preview was successfully edited to the final text.
           // No follow-up messages needed.
+          await cleanupArchivedBubbles();
           console.log(`[${label}] ${elapsed}s | ${result.text.length} chars | streamed | ${chatConfig.model}/${chatConfig.effort} | $${result.cost?.toFixed(4) || '?'}`);
           markReplied();
           return;
@@ -1868,7 +1922,25 @@ async function handleMessage(sessionKey, chatId, msg, bot) {
           delivered: r.sent.length, failed: r.failed.length,
           bot: BOT_NAME,
         });
-        console.log(`[${label}] ${elapsed}s | ${result.text.length} chars | streamed-redeliver(${reason}, ${chunks.length} chunks) | ${chatConfig.model}/${chatConfig.effort} | $${result.cost?.toFixed(4) || '?'}`);
+        // 0.7.1: surface partial-failure to the user. Without this,
+        // a chunk-3-of-5 failure leaves a coherent-looking reply with
+        // a silent gap (the user reads chunks 1, 2, 4, 5 unaware
+        // that chunk 3 was dropped). Append a warning + flip the
+        // reactor to ERROR so something visible signals "look here".
+        if (r.failed.length > 0) {
+          reactor.setState('ERROR');
+          try {
+            await tg(bot, 'sendMessage', {
+              chat_id: chatId,
+              text: `⚠️ ${r.failed.length} of ${chunks.length} message parts failed to deliver. The reply may be incomplete — please retry.`,
+              ...(threadId && { message_thread_id: threadId }),
+            }, { ...outMetaBase, source: 'partial-delivery-warning' });
+          } catch (warnErr) {
+            console.error(`[${label}] partial-delivery warning failed: ${warnErr.message}`);
+          }
+        }
+        await cleanupArchivedBubbles();
+        console.log(`[${label}] ${elapsed}s | ${result.text.length} chars | streamed-redeliver(${reason}, ${chunks.length} chunks${r.failed.length ? `, ${r.failed.length} failed` : ''}) | ${chatConfig.model}/${chatConfig.effort} | $${result.cost?.toFixed(4) || '?'}`);
         markReplied();
         return;
       }