use agent manager for subagent

[yingjiehe-xyz]

Pull Request Description

Use the agent manager added in #4684 for the subagent task execution

Test:

( O)> use the subagent to summarize my slacks channels g2-help

I'll use a subagent to summarize your Slack channel `g2-help`. Let me create a task for this.
─── create_task | dynamic_task ──────────────────────────
task_parameters:
    -
        instructions: Summarize the Slack channel 'g2-help'. Get recent messages from this channel and provide a comprehensive summary of the discussions, key topics, common issues, and any important information shared. Focus on the most recent activity and highlight any recurring themes or important announcements.
        title: Summarize g2-help Slack Channel
        extensions:
            - slack
        return_last_only: true


─── execute_task | subagent ──────────────────────────
task_ids: 1a070d51-3d2d-4f57-8c19-cad83b66418d

# Summary of G2-Help Slack Channel

Based on the recent messages from the **g2-help** channel, here's a comprehensive summary of the discussions, key topics, and common issues:

## **Key Topics & Recurring Themes**

### **1. Integration Requests & MCP Extensions**
- **Zendesk Integration**: Request for Zendesk integration for Bitkey's customer service management
- **Sentry Integration**: On the roadmap and being prioritized
- **Salesforce & Looker**: Targeting Q4 for Salesforce, working on timeline for Looker
- **Square Wise Integration**: Request for technical teams integration
- **Multiple Notion Spaces**: Users want ability to connect to multiple Notion workspaces simultaneously

### **2. Technical Issues & Troubleshooting**

#### **Context Limit Errors**
- Users experiencing "conversation has exceeded the model's context limit" errors
- **Solution**: Start a new session to reset token count
- Progress bar shows token usage during chat sessions
- Future context summarization feature planned

#### **Google Calendar SSL Issues**
- SSL/TLS connection failures when connecting to Google Calendar API
- Error: `[SSL] record layer failure (_ssl.c:2580)`
- Issue appears intermittent and may resolve on its own

......

[DOsinga]

I have some preliminary questions but like the deleting

[DOsinga]

this looks pretty decent, but I think we need to tighten up the typing here and make the compiler our friend.

my main concern here is how we run the agent here. we have a parent_id but we're not doing all that much with it. the TaskConfig has a provider and a parent_id and the provider is the provider of the parent. maybe it should copy that on instantiation? same for working_dir I think. extensions is set separately again, but I feel like this should sometimes default to whatever the agent had at the time? like if you run a subrecipe where you don't specify extensions

[yingjiehe-xyz]

this looks pretty decent, but I think we need to tighten up the typing here and make the compiler our friend.

my main concern here is how we run the agent here. we have a parent_id but we're not doing all that much with it. the TaskConfig has a provider and a parent_id and the provider is the provider of the parent. maybe it should copy that on instantiation? same for working_dir I think. extensions is set separately again, but I feel like this should sometimes default to whatever the agent had at the time? like if you run a subrecipe where you don't specify extensions

for the extension, I think it would be better to start with specified extension to reduce the prompt length and keep it consistent with recipe

[DOsinga]

for the extension, I think it would be better to start with specified extension to reduce the prompt length and keep it consistent with recipe

sorry, but that is not how I read the code. I think we ask the AI to generate a spec to run with an empty list for extensions for no extensions and none for inheriting the extensions. /cc @tlongwell-block

[tlongwell-block]

for the extension, I think it would be better to start with specified extension to reduce the prompt length and keep it consistent with recipe

sorry, but that is not how I read the code. I think we ask the AI to generate a spec to run with an empty list for extensions for no extensions and none for inheriting the extensions. /cc @tlongwell-block

Right now, if the parent agent does not specify the extensions parameter, the subagent it creates gets all of the same extensions as the parent agent

If the parent agent specifies an empty extensions parameter, the subagent it creates gets no extensions

[tlongwell-block]

─── create_task | dynamic_task ──────────────────────────
task_parameters:
    -
        title: Test 4a: Weather
        instructions: Tell me a fun fact about weather
        return_last_only: true
    -
        title: Test 4b: Joke
        instructions: Tell me a knock-knock joke
        return_last_only: true


─── execute_task | subagent ──────────────────────────
task_ids: 87641b31-7d63-4081-95cf-6c5e957f5f3e, e8b0b3ed-a30a-4174-b521-636e95ab49fc
execution_mode: parallel


🎯 Task Execution Dashboard
═══════════════════════════

📊 Progress: 2 total | ⏳ 0 pending | 🏃 0 running | ✅ 1 completed | ❌ 1 failed

❌ 87641b31-7d63-4081-95cf-6c5e957f5f3e (inline_recipe)
   ⏱️  0.0s
   ⚠️  Inline recipe execution failed: -32603: Failed to execute task: Failed to cre...

✅ e8b0b3ed-a30a-4174-b521-636e95ab49fc (inline_recipe)
   ⏱️  4.0s
   📄 {
  "result": "Here's a knock-knock joke for you:\n\nKnock, knock.  \nWho's there?  \nInterrupting cow.  \nInterrupting cow w—  \n**MOOOOO!**\n\n😄 The joke works because the cow interrupts before you can finish saying \"who?\""
}

Execution Complete!
═══════════════════════
Total Tasks: 2
✅ Completed: 1
❌ Failed: 1
📈 Success Rate: 50.0%

❌ Failed Tasks:
   • 87641b31-7d63-4081-95cf-6c5e957f5f3e
     Error: Inline recipe execution failed: -32603: Failed to execute task: Failed to create a session for sub agent: error returned from database: (code: 5) database is locked

📝 Generating summary...
-32603: 1/2 tasks failed:
Task '87641b31-7d63-4081-95cf-6c5e957f5f3e' (ID: 87641b31-7d63-4081-95cf-6c5e957f5f3e): Inline recipe execution failed: -32603: Failed to execute task: Failed to create a session for sub agent: error returned from database: (code: 5) database is locked

Output: No output captured

Interesting! We hit a database locking issue with parallel execution. Let me update the TODO and retry:

[DOsinga]

also let me know @yingjiehe-xyz when you want me to have another look

[yingjiehe-xyz]

also let me know @yingjiehe-xyz when you want me to have another look

yes, I think it should be ready for the review