polygram 0.8.0-rc.28 → 0.8.0-rc.29

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.28",
+  "version": "0.8.0-rc.29",
   "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/process-manager-sdk.js

@@ -195,6 +195,7 @@ class ProcessManagerSdk {
     onAssistantMessageStart = null,
     onCompactBoundary = null,
     onQueueDrop = null,
+    onThinking = null,
     queryCloseTimeoutMs = DEFAULT_QUERY_CLOSE_TIMEOUT_MS,
   } = {}) {
     if (!spawnFn) throw new Error('spawnFn required');
@@ -212,6 +213,7 @@ class ProcessManagerSdk {
     this.onAssistantMessageStart = onAssistantMessageStart;
     this.onCompactBoundary = onCompactBoundary;
     this.onQueueDrop = onQueueDrop;
+    this.onThinking = onThinking;
     this.procs = new Map();                    // sessionKey → entry
     this._lruWaiters = [];                     // [{ resolve, reject, timer }]
   }
@@ -380,6 +382,31 @@ class ProcessManagerSdk {
       return;
     }
 
+    // rc.29: stream_event with content_block_start of type='thinking'.
+    // Fires DURING extended-thinking phase, BEFORE any text or tool_use
+    // content appears. Without this, polygram's reactor stays at QUEUED
+    // (👀) for the full thinking duration (10+ s with effort=high),
+    // then transitions THINKING → CODING in <1s when the model finally
+    // emits text/tool. UX target: 👀 → 🤔 transition fires within
+    // 100-500ms of pm.send, matching Claude Code CLI's "Thinking..."
+    // spinner timing.
+    //
+    // Requires `includePartialMessages: true` in SdkOptions; without
+    // it, this branch is unreachable (we never receive stream_event).
+    if (msg.type === 'stream_event' && head && !head.thinkingFired) {
+      const ev = msg.event;
+      const isThinkingStart = ev?.type === 'content_block_start'
+        && ev?.content_block?.type === 'thinking';
+      if (isThinkingStart) {
+        head.thinkingFired = true;
+        if (this.onThinking) {
+          try { this.onThinking(entry.sessionKey, entry); }
+          catch (err) { this.logger.error?.(`[${entry.label}] onThinking: ${err.message}`); }
+        }
+      }
+      return;
+    }
+
     if (msg.type === 'system' && msg.subtype === 'compact_boundary') {
       // D6 / §5: surface compaction boundary to caller. Sequencing
       // guarantee — we await this callback before processing the
@@ -587,6 +614,7 @@ class ProcessManagerSdk {
         prompt,
         transientRetries: 0,
         firstAssistantSeen: false,
+        thinkingFired: false,                  // rc.29: extended-thinking → reactor THINKING
       };
 
       pending.fireFirstStream = () => {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "polygram",
-  "version": "0.8.0-rc.28",
+  "version": "0.8.0-rc.29",
   "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

@@ -959,6 +959,14 @@ 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 && {
@@ -3584,6 +3592,20 @@ 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,
+      });
+    },
     // 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