opencode and native tasks

seed/tasks/transcribe-audio-file/TASK.md

@@ -26,7 +26,7 @@ Transcribe an audio file to text using whisper.cpp.
 
 ## Steps
 
-1. Determine the user's root directory by navigating one level up from the current working directory. Read `settings.json` from the root directory and extract the `languages` section.
+1. Starting from the directory containing the audio file, walk up parent by parent until you find a directory whose name is an email address. This is the user's root directory. Read `settings.json` from it and extract the `languages` section.
 2. Detect the language of the audio file using the whisper.cpp skill with `detect_language=true` and `response_format=verbose_json`.
 3. Compare the detected language against the user's `languages.spoken` list. If the detected language is in the list, skip translation. Otherwise, set `translate=true`.
 4. Use the whisper.cpp skill to transcribe the audio file at `file_path`, passing the detected language as the `language` parameter and the `translate` flag from the previous step.

src/servers/api/opencode/websocket.ts

@@ -2,7 +2,8 @@ import type { ServerWebSocket } from 'bun';
 import { mkdir, rename } from 'node:fs/promises';
 import { join } from 'node:path';
 import type { ClientMessage, ServerMessage, ImageData, TaskInfo } from '@@/api/chat-types';
-import { getTmpAttachmentsDir, getAttachmentsDir } from '@@/data-path';
+import { getTmpAttachmentsDir, getAttachmentsDir, getNativeSkillsDir, getGlobalSkillsDir, getUserSkillsDir } from '@@/data-path';
+import { readSkillDirs, parseFrontmatter } from '@@/api/skills/skills';
 import { createTaskLog, appendToLog, finalizeLog } from '@@/api/task-logger';
 
 const OPENCODE_PORT = process.env.OPENCODE_PORT ?? '10006';
@@ -303,6 +304,29 @@ function handleSSEEvent(ws: ServerWebSocket<WSData>, event: string, data: any) {
   }
 }
 
+async function buildSkillsPrompt(email: string): Promise<string> {
+  const nativeSkills = await readSkillDirs(getNativeSkillsDir());
+  const globalSkills = await readSkillDirs(getGlobalSkillsDir());
+  const userSkills = await readSkillDirs(getUserSkillsDir(email));
+
+  const merged = new Map(nativeSkills);
+  for (const [name, path] of globalSkills) merged.set(name, path);
+  for (const [name, path] of userSkills) merged.set(name, path);
+
+  if (merged.size === 0) return '';
+
+  const lines = await Promise.all(
+    Array.from(merged.entries()).map(async ([dirName, filePath]) => {
+      const raw = await Bun.file(filePath).text();
+      const { frontmatter } = parseFrontmatter(raw);
+      const name = frontmatter.name || dirName;
+      return `- ${name}: ${frontmatter.description} (read ${filePath} for full instructions)`;
+    }),
+  );
+
+  return `\n\nYou have access to the following skills. When a user's request matches a skill, read its SKILL.md file for detailed instructions before proceeding.\n\nAvailable skills:\n${lines.join('\n')}`;
+}
+
 type HandleChatParams = {
   ws: ServerWebSocket<WSData>;
   prompt: string;
@@ -368,6 +392,9 @@ async function handleChat({ ws, prompt, sessionId, model, attachmentIds, images,
       }
     }
 
+    const skillsAppend = await buildSkillsPrompt(ws.data.email);
+    const fullPrompt = skillsAppend ? `${skillsAppend}\n\n${prompt}` : prompt;
+
     const parts: Record<string, unknown>[] = [];
     if (images?.length) {
       for (const img of images) {
@@ -378,7 +405,7 @@ async function handleChat({ ws, prompt, sessionId, model, attachmentIds, images,
         });
       }
     }
-    parts.push({ type: 'text', text: prompt });
+    parts.push({ type: 'text', text: fullPrompt });
 
     const promptBody: Record<string, unknown> = { parts };
     if (modelSelection.modelId) {