@pi-unipi/notify 2.16.0 → 2.16.1

package/README.md

@@ -120,6 +120,22 @@ Settings stored at `~/.unipi/config/notify/config.json`. Edit via `/unipi:notify
 
 Per-event platform routing lets you control where each event type goes. The settings overlay shows all events with platform toggles.
 
+### Recap (thinking models)
+
+Recap summarizes the last assistant message into a one-line push notification (100-token budget). Thinking models served by llama.cpp or vLLM can spend that entire budget on reasoning and return nothing, falling back to a plain 100-character truncation. If your recap endpoint supports chat-template kwargs, set `recap.disableThinking` to skip reasoning tokens:
+
+```json
+{
+  "recap": {
+    "enabled": true,
+    "model": "localhost/gemma-4-e4b",
+    "disableThinking": true
+  }
+}
+```
+
+This sends `chat_template_kwargs: { enable_thinking: false, preserve_thinking: false }` with the request. Keep it `false` (the default) for strict OpenAI-compatible endpoints — they reject unknown params. Anthropic models are unaffected (thinking is opt-in there).
+
 ## License
 
 MIT

package/events.ts

@@ -352,7 +352,9 @@ function registerAgentNotification(
             .then((apiKeyResult) => {
               const apiKey = apiKeyResult.ok ? (apiKeyResult as { apiKey?: string }).apiKey : undefined;
               if (apiKey) {
-                return summarizeLastMessage(lastText, apiKey, model.baseUrl, model.api, modelId)
+                return summarizeLastMessage(lastText, apiKey, model.baseUrl, model.api, modelId, {
+                  disableThinking: config.recap.disableThinking,
+                })
                   .then((recap) => sessionName ? `${sessionName}: ${recap}` : recap);
               }
               return buildAgentLifecycleMessage(eventKey, sessionName);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-unipi/notify",
-  "version": "2.16.0",
+  "version": "2.16.1",
   "description": "Cross-platform notification extension for Pi — native OS, Gotify, and Telegram notifications for agent lifecycle events",
   "type": "module",
   "main": "index.ts",
@@ -34,7 +34,7 @@
     "access": "public"
   },
   "dependencies": {
-    "@pi-unipi/core": "2.16.0",
+    "@pi-unipi/core": "2.16.1",
     "node-notifier": "^10.0.1"
   },
   "peerDependencies": {

package/settings.ts

@@ -45,6 +45,7 @@ export const DEFAULT_CONFIG: NotifyConfig = {
   recap: {
     enabled: false,
     model: "openrouter/openai/gpt-oss-20b",
+    disableThinking: false,
   },
   silenceAfterInput: {
     enabled: false,

package/summarize.ts

@@ -12,6 +12,17 @@ const MAX_TOKENS = 100;
 const TIMEOUT_MS = 10_000;
 const FALLBACK_TRUNCATE_CHARS = 100;
 
+/** Options for summarizeLastMessage */
+export interface SummarizeOptions {
+  /**
+   * Send `chat_template_kwargs: { enable_thinking: false, preserve_thinking: false }`
+   * on OpenAI-compatible requests (llama.cpp / vLLM chat templates) so thinking
+   * models don't burn the token budget on reasoning (issue #36). Anthropic
+   * ignores this — thinking is opt-in there already.
+   */
+  disableThinking?: boolean;
+}
+
 /**
  * Summarize a message using an LLM.
  *
@@ -20,6 +31,7 @@ const FALLBACK_TRUNCATE_CHARS = 100;
  * @param baseUrl - Provider base URL (from Model.baseUrl)
  * @param api - API type (from Model.api, e.g. "openai-completions")
  * @param modelId - Model ID to use
+ * @param opts - Optional summarization options
  * @returns Summarized text, or truncated original on failure
  */
 export async function summarizeLastMessage(
@@ -28,6 +40,7 @@ export async function summarizeLastMessage(
   baseUrl: string,
   api: string,
   modelId: string,
+  opts?: SummarizeOptions,
 ): Promise<string> {
   // Truncate input if too long
   const input =
@@ -41,7 +54,7 @@ export async function summarizeLastMessage(
       return await callAnthropic(baseUrl, apiKey, modelId, input);
     }
     // Default: OpenAI-compatible (covers openai-completions, openai-responses, etc.)
-    return await callOpenAICompatible(baseUrl, apiKey, modelId, input);
+    return await callOpenAICompatible(baseUrl, apiKey, modelId, input, opts);
   } catch {
     return fallbackSummary(messageText);
   }
@@ -53,6 +66,7 @@ async function callOpenAICompatible(
   apiKey: string,
   modelId: string,
   input: string,
+  opts?: SummarizeOptions,
 ): Promise<string> {
   const url = `${baseUrl.replace(/\/$/, "")}/chat/completions`;
   const controller = new AbortController();
@@ -68,6 +82,14 @@ async function callOpenAICompatible(
       body: JSON.stringify({
         model: modelId,
         max_tokens: MAX_TOKENS,
+        ...(opts?.disableThinking
+          ? {
+              chat_template_kwargs: {
+                enable_thinking: false,
+                preserve_thinking: false,
+              },
+            }
+          : {}),
         messages: [
           { role: "system", content: SYSTEM_PROMPT },
           { role: "user", content: input },

package/types.ts

@@ -69,6 +69,15 @@ export interface RecapConfig {
   enabled: boolean;
   /** Model to use for recap (e.g. "openrouter/openai/gpt-oss-20b") */
   model: string;
+  /**
+   * Send `chat_template_kwargs: { enable_thinking: false, preserve_thinking: false }`
+   * with recap requests so llama.cpp/vLLM-style servers skip reasoning tokens.
+   * Without this, a thinking model can burn the entire 100-token budget on
+   * reasoning and return no summary (issue #36). Only enable for endpoints
+   * that accept these params — strict OpenAI-compatible servers reject them.
+   * The Anthropic path ignores this (thinking is opt-in there already).
+   */
+  disableThinking?: boolean;
 }
 
 /** Quiet listed platforms after recent terminal input */