polygram 0.8.0-rc.30 → 0.8.0-rc.32

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.30",
+  "version": "0.8.0-rc.32",
   "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/package.json

@@ -1,6 +1,6 @@
 {
   "name": "polygram",
-  "version": "0.8.0-rc.30",
+  "version": "0.8.0-rc.32",
   "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

@@ -948,27 +948,9 @@ function buildSdkOptions(sessionKey, ctx) {
       })
     : null;
 
-  // rc.29 (refined): map effort → thinking. SDK's `effort` knob alone
-  // does NOT enable extended thinking — it only "works WITH" thinking.
-  // Without an explicit `thinking` config, the model emits no thinking
-  // content blocks, our onThinking callback never fires, and the
-  // reactor stays at 👀 until first text/tool. Mapping:
-  //   low    → disabled (fastest replies)
-  //   medium → adaptive (Claude decides when to think)
-  //   high   → adaptive (same; effort biases depth)
-  //   xhigh  → adaptive
-  //   max    → adaptive
-  // Chat configs can override by setting `thinking` directly in
-  // chatConfig (composeSdkOptions passes it through).
-  const effortValue = chatConfig.effort || config.defaults.effort;
-  const derivedThinking = effortValue === 'low'
-    ? { type: 'disabled' }
-    : { type: 'adaptive' };
-
   const baseOpts = {
     model: chatConfig.model || config.defaults.model,
-    effort: effortValue,
-    thinking: derivedThinking,
+    effort: chatConfig.effort || config.defaults.effort,
     cwd: chatConfig.cwd,
     env: childEnv,
     // permissionMode 'default' when canUseTool is wired (so the SDK
@@ -977,14 +959,6 @@ function buildSdkOptions(sessionKey, ctx) {
     permissionMode: useCanUseTool ? 'default' : 'bypassPermissions',
     allowDangerouslySkipPermissions: !useCanUseTool,
     ...(useCanUseTool && { canUseTool: makeCanUseTool(sessionKey) }),
-    // rc.29: enable partial messages so pm-sdk can detect thinking
-    // blocks during the extended-thinking phase (before any text or
-    // tool_use arrives). Without this, the reactor stays at 👀 for
-    // the full thinking duration. With this, pm-sdk's onThinking
-    // callback fires within ~100-500ms of pm.send, giving the user
-    // a fast 👀 → 🤔 transition matching Claude Code CLI's
-    // "Thinking..." spinner.
-    includePartialMessages: true,
     hooks: {
       PostToolBatch: [{ hooks: [postToolBatchHook] }],
       ...(sessionStartHook && {
@@ -2440,16 +2414,23 @@ async function handleMessage(sessionKey, chatId, msg, bot) {
     availableEmojis,
     logError: (m) => console.error(`[${label}] ${m}`),
   });
-  // Start at QUEUED (👀) so user sees their message was received but
-  // not yet being worked on. pm calls context.onActivate when this
-  // pending becomes the queue head (Claude is actually starting it),
-  // at which point we flip to THINKING (🤔).
-  // 0.7.4 (item G): if voice ack (👂) was just emitted by
-  // transcribeVoiceAttachments, skip QUEUED — its 👀 would overwrite the
-  // ack within milliseconds, wasting an API call and flickering. Let 👂
-  // stay until THINKING flips it to 🤔 when Claude actually starts work.
+  // rc.32: skip QUEUED (👀) entirely for first-message-in-chain. Go
+  // straight to THINKING (🤔). The 👀 → 🤔 two-hop didn't add
+  // user-readable signal — Telegram's ✓✓ already conveys "delivered",
+  // and the technical "received-but-not-started vs thinking"
+  // distinction is operator-debugging context, not user UX.
+  //
+  // Follow-up messages during an in-flight turn still go through the
+  // autosteer path (✍ AUTOSTEERED state) — that's the visual that
+  // means "captured while bot is busy, will incorporate." This
+  // change ONLY affects the trigger message of a fresh turn:
+  // previously 👀 → (300ms timer) → 🤔, now just 🤔 immediately.
+  //
+  // 0.7.4 (item G) voice-ack guard preserved: if 👂 is up from
+  // voice transcription, don't overwrite it. Let onFirstStream
+  // promote to 🤔 when Claude actually starts work.
   if (!voiceAck.ackEmitted) {
-    reactor.setState('QUEUED');
+    reactor.setState('THINKING');
   }
 
   // Mark the inbound row terminal so boot replay doesn't pick it up
@@ -3610,20 +3591,13 @@ async function main() {
       const r = head?.context?.reactor;
       if (r && typeof r.heartbeat === 'function') r.heartbeat();
     },
-    // rc.29: extended-thinking → 🤔 transition. Fires the moment the
-    // model's first content_block_start with type='thinking' arrives
-    // via stream_event. Pre-rc.29 we waited for first text/tool_use,
-    // which under effort=high means the reactor sat at 👀 for 10+ s
-    // before flipping. Now the user sees 👀 → 🤔 within ~100-500ms
-    // of pm.send, matching Claude Code CLI's "Thinking..." UX.
-    onThinking: (sessionKey, entry) => {
-      const head = entry.pendingQueue?.[0];
-      const r = head?.context?.reactor;
-      if (r && typeof r.setState === 'function') r.setState('THINKING');
-      logEvent('thinking-started', {
-        chat_id: entry.chatId, session_key: sessionKey,
-      });
-    },
+    // rc.29 onThinking removed — replaced by simpler timer-based
+    // approach in handleMessage (post-QUEUED setState). The
+    // SDK-thinking-block detection added complexity (partial-message
+    // bandwidth, thinking config mapping) without solving the actual
+    // user-visible problem ("show me the bot is working"). The lib-
+    // side handler in pm-sdk stays for any future caller; polygram
+    // doesn't wire it.
     // 0.8.0 Phase 2 step 5: SDK auto-compaction observability. Fires
     // when SDK emits SDKCompactBoundaryMessage (between turns or
     // mid-turn — see Phase 0 gate 8.5). Surfaces a quiet system