fix(desktop): unify agent tools and floating pills

AGENTS.md

@@ -152,7 +152,7 @@ Rules:
 Agents can and should self-test the running app — don't stop at a successful compile. The fast path skips the slow parts (web login, sidebar click-through):
 
 1. **Build + launch a named bundle:** `cd desktop/macos && OMI_APP_NAME="omi-<feature>" ./run.sh` (add `OMI_SKIP_TUNNEL=1` for a local backend without a tunnel; `OMI_SKIP_BACKEND=1 OMI_DESKTOP_API_URL=…` to point at a remote backend).
-2. **Boot signed-in (no browser):** sign into "Omi Dev" once, then clone the session into the named bundle **before launch** (UserDefaults is read at startup):
+2. **Boot signed-in (no browser):** sign into "Omi Dev" once; `./run.sh` auto-clones auth/onboarding into named bundles **before launch** (UserDefaults is read at startup). To do it manually:
    ```bash
    cd desktop/macos && ./scripts/omi-auth-dump.sh                  # capture the Omi Dev session
    ./scripts/omi-auth-seed.sh com.omi.omi-<feature>          # replay into the test bundle

desktop/macos/AGENTS.md

@@ -127,9 +127,9 @@ This creates `/Applications/omi-fix-rewind.app` with bundle ID `com.omi.omi-fix-
 - NEVER use the default `./run.sh` (which overwrites "Omi Dev") when testing a specific feature — always set `OMI_APP_NAME`
 - **ALWAYS prefix the name with `omi-`** (e.g., `omi-fix-rewind`, `omi-6512-polling`, `omi-vision-test`) so named bundles are visually grouped in `/Applications/` alongside "Omi Dev" and "Omi Beta"
 - Keep the name short and descriptive (it becomes both the app name and bundle ID suffix)
-- The named bundle gets its own permissions, database, and auth state — the user may need to re-grant permissions and sign in
+- The named bundle gets its own permissions and database. `./run.sh` auto-seeds auth/onboarding from "Omi Dev" unless `OMI_SKIP_AUTH_SEED=1` is set.
 - To connect agent-swift: `agent-swift connect --bundle-id com.omi.omi-fix-rewind`
-- **Skip the web login:** sign into "Omi Dev" once, then `./scripts/omi-auth-dump.sh && ./scripts/omi-auth-seed.sh com.omi.omi-fix-rewind` clones the session so the named bundle boots signed-in
+- **Skip the web login:** sign into "Omi Dev" once; named bundles launched by `./run.sh` clone that session before launch.
 - **Jump to a screen without clicking:** the automation bridge auto-enables on non-prod bundles — `./scripts/omi-ctl navigate <screen>` (e.g. `rewind`, `memories`, `settings rewind`). See "Fast-Path for Local Iteration" in `e2e/SKILL.md`.
 
 ### After Implementing Changes

desktop/macos/Desktop/Sources/Chat/ChatPrompts.swift

@@ -273,6 +273,11 @@ struct ChatPrompts {
       * "Do I like dogs?" → get_memories_tool (FACT/preference)
       * "When did a dog bite me?" → search_conversations_tool (EVENT)
 
+    **Task-Agent Awareness:**
+    - Omi can run local task-chat agents/subagents in the desktop task panel. If the user says "your subagents", "task agents", "running agents", or mentions task-agent errors/timeouts, do NOT deny that you have subagents.
+    - Call get_task_agent_status to inspect the current Omi task-agent registry before answering.
+    - If a task agent timed out or failed, summarize the affected task and error, then suggest a concrete retry/recovery step.
+
     **Limit and Transcript Guidance:**
     - **Summarization queries** ("summarize my week", "recap my month"): set limit=5000, include_transcript=false to get all conversations without flooding context.
     - **Normal queries**: limit=20 (default) is fine for most questions.
@@ -469,65 +474,7 @@ struct ChatPrompts {
     </critical_accuracy_rules>
 
     <tools>
-    You have 7 tools. ALWAYS use them before answering — don't guess when you can look it up.
-
-    **execute_sql**: Run SQL on the local omi.db database.
-    - Supports: SELECT, INSERT, UPDATE, DELETE
-    - SELECT auto-limits to 200 rows. UPDATE/DELETE require WHERE. DROP/ALTER/CREATE blocked.
-    - Use for: personal facts, app usage stats, time queries, task management, aggregations, anything structured.
-    - Supports FTS5 MATCH queries for keyword search (see schema for FTS tables and patterns).
-
-    **semantic_search**: Vector similarity search on screen history.
-    - Use for: fuzzy conceptual queries where exact SQL keywords won't work.
-    - e.g. "reading about machine learning", "working on design mockups"
-    - Parameters: query (required), days (default 7), app_filter (optional)
-
-    **search_tasks**: Vector similarity search on tasks (action_items + staged_tasks).
-    - Use for: finding tasks by meaning, not just keywords.
-    - e.g. "tasks about shopping", "anything related to the presentation"
-    - Parameters: query (required), include_completed (default false)
-    - More reliable than hand-writing MATCH queries for task search.
-
-    **get_daily_recap**: Pre-formatted activity recap (apps, conversations, tasks, focus, memories, observations).
-    - Use for: "what did I do today/yesterday/this week" — single tool call, much faster than multiple SQL queries.
-    - Parameters: days_ago (0=today, 1=yesterday, 7=past week, default: 1)
-
-    **complete_task**: Toggle a task's completion status.
-    - Takes: task_id (the backendId from action_items table)
-    - Use for: marking tasks done or uncompleting them
-    - First use execute_sql to find the task, then use this tool with its backendId
-
-    **delete_task**: Delete a task permanently.
-    - Takes: task_id (the backendId from action_items table)
-    - Use for: removing tasks the user no longer needs
-    - First use execute_sql to find the task, then use this tool with its backendId
-
-    **save_knowledge_graph**: Save a knowledge graph of entities and relationships extracted from the user's data.
-    - Parameters: nodes (array of {id, label, node_type, aliases}), edges (array of {source_id, target_id, label})
-    - node_type must be one of: person, organization, place, thing, concept
-    - Use when: exploring the user's files during onboarding to build their knowledge graph
-    - Deduplication is handled automatically — just provide all entities you find
-
-    **CRITICAL — When to use tools proactively:**
-    The <user_facts> section above only contains a SAMPLE of {user_name}'s memories. The full set is in the database.
-    For ANY personal question (age, preferences, relationships, habits, past events, "what do you know about me", etc.):
-    1. FIRST check <user_facts> — if the answer is there, use it directly.
-    2. If NOT in <user_facts>, ALWAYS query the memories table before saying you don't know.
-    3. For questions about past events or conversations, query transcription_sessions/transcription_segments.
-    NEVER say "I don't know" or "I don't have that info" without checking the database first.
-
-    **When to use which tool:**
-    - "how old am I?" / "what's my name?" / personal facts → execute_sql (query memories table)
-    - "what did I do yesterday?" → get_daily_recap (single tool call, returns formatted summary)
-    - "what apps did I use most?" → execute_sql (GROUP BY appName, COUNT)
-    - "find where I was reading about AI" → semantic_search (conceptual)
-    - "find tasks about shopping" → search_tasks (semantic task search)
-    - "create a task to buy milk" → execute_sql (INSERT INTO action_items)
-    - "what are my tasks?" → execute_sql (SELECT FROM action_items)
-    - "complete the first task" → execute_sql to find backendId, then complete_task
-    - "delete that task" → execute_sql to find backendId, then delete_task
-    - "show my conversations" → execute_sql (SELECT FROM transcription_sessions)
-    - "what did I talk about with John?" → execute_sql (search transcription_segments)
+    \(DesktopCapabilityRegistry.desktopToolPrompt)
 
     {database_schema}