polygram 0.8.0-rc.55 → 0.8.0-rc.56

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://anthropic.com/claude-code/plugin.schema.json",
   "name": "polygram",
-  "version": "0.8.0-rc.55",
+  "version": "0.8.0-rc.56",
   "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/lib/context-format.js

@@ -1,5 +1,5 @@
 /**
- * Pure formatters for /context command output and the 85%-full hint.
+ * Pure formatters for /context command output and the context-full hint.
  *
  * Lifted from polygram.js so the formatting can be unit-tested without
  * spinning up the full handleMessage stack. Both functions are pure —
@@ -12,11 +12,22 @@
  *   0-1 ratio and multiplied by 100, which displayed "7700% full" and
  *   skipped the 85% hint threshold. The formatters below assume the
  *   0-100 scale; do not multiply or divide.
+ *
+ * rc.56 threshold change: default lowered from 85 → 70.
+ *   Background: at 85% the SDK has typically already auto-compacted
+ *   mid-turn, so polygram's post-turn check sees a low percentage
+ *   and the hint never fires. Production data showed 0 user-visible
+ *   hint triggers across 15 auto-compactions in May 2026, all of
+ *   which fired at pre_tokens 167-262k (≈85% of Sonnet's 200k
+ *   window). Lowering to 70% means polygram warns ~30k tokens
+ *   before the SDK auto-compacts, giving the user 1-3 turns of
+ *   headroom to choose /new vs /compact vs continue. Configurable
+ *   per-bot or per-chat via `contextHintThreshold` (number, 0-100).
  */
 
 'use strict';
 
-const HINT_THRESHOLD_PCT = 85;
+const HINT_THRESHOLD_PCT = 70;
 
 /**
  * Format a getContextUsage() result into a multi-line chat reply.
@@ -55,14 +66,20 @@ function formatContextReply(usage) {
 }
 
 /**
- * Decide whether to send the 85% hint and return the hint text if so.
+ * Decide whether to send the context-full hint and return the hint
+ * text if so.
  *
  * @param {object} usage  — same shape as formatContextReply input.
- * @returns {string|null} the hint text to send, or null when below threshold.
+ * @param {object} [opts]
+ * @param {number} [opts.threshold] — override the default percent
+ *   threshold (rc.56). Caller resolves per-chat / per-bot config
+ *   and passes it in. Defaults to HINT_THRESHOLD_PCT (70).
+ * @returns {string|null} the hint text to send, or null when below
+ *   threshold.
  */
-function maybeContextFullHint(usage) {
+function maybeContextFullHint(usage, { threshold = HINT_THRESHOLD_PCT } = {}) {
   const pct = usage?.percentage ?? 0;
-  if (pct < HINT_THRESHOLD_PCT) return null;
+  if (pct < threshold) return null;
   return [
     `📚 Context window ${pct.toFixed(0)}% full. Three options:`,
     '',

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "polygram",
-  "version": "0.8.0-rc.55",
+  "version": "0.8.0-rc.56",
   "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

@@ -2859,14 +2859,22 @@ async function handleMessage(sessionKey, chatId, msg, bot) {
       // result, the followup messages (if any) get their own SDK
       // pause to absorb at, no special handling needed.
 
-      // 0.8.0 Phase 2 step 4: 85%-context-full live hint. After a
+      // 0.8.0 Phase 2 step 4: context-full live hint. After a
       // successful turn, peek at SDK's getContextUsage(); if past
-      // 85%, post a quiet hint so the user knows /new will help.
-      // SDK pm only — CLI pm has no equivalent (no Query object,
-      // no getContextUsage). OPT-IN per-chat or per-bot
-      // (rc.12+) — most chats don't want the noise. Per-chat takes
+      // the threshold, post a quiet hint so the user knows /new
+      // will help. SDK pm only — CLI pm has no equivalent (no
+      // Query object, no getContextUsage). OPT-IN per-chat or
+      // per-bot — most chats don't want the noise. Per-chat takes
       // precedence over per-bot so admins (Ivan DM) can opt in
       // without forcing it on every other chat.
+      //
+      // rc.56: threshold default lowered to 70% (was 85%) because
+      // the SDK auto-compacts mid-turn at ~85% — by the time
+      // polygram queries getContextUsage post-turn, the percentage
+      // has already dropped and the hint never fires. 70% gives
+      // the user 1-3 turns of headroom before SDK compaction.
+      // Configurable via `contextHintThreshold` (number, 0-100)
+      // per-chat or per-bot. Same precedence rule as contextHint.
       const chatCtxHint = chatConfig.contextHint != null
         ? chatConfig.contextHint
         : config.bot?.contextHint;
@@ -2874,8 +2882,13 @@ async function handleMessage(sessionKey, chatId, msg, bot) {
         const entry = pm.get(sessionKey);
         const q = entry?.query;
         if (q && typeof q.getContextUsage === 'function') {
+          const threshold = chatConfig.contextHintThreshold != null
+            ? chatConfig.contextHintThreshold
+            : (config.bot?.contextHintThreshold != null
+              ? config.bot.contextHintThreshold
+              : undefined);
           q.getContextUsage().then((usage) => {
-            const text = maybeContextFullHint(usage);
+            const text = maybeContextFullHint(usage, threshold != null ? { threshold } : undefined);
             if (!text) return;
             return tg(bot, 'sendMessage', {
               chat_id: chatId,