Tools are being wiped with iterations

[superbeagle]

In using the Agentspan demo environment it appears tools were being 'forgotten'. Using AI it appears it found the possible issue in AgentService:

Look at the start() method — it always calls metadataDAO.updateWorkflowDef(def) which overwrites the stored workflow definition with a freshly compiled one on every agent/start call. That's why tools get wiped.

[superbeagle]

What may be happening at least according to Claude:

_From the UI:

UI calls POST /api/agentspan/agent/start with the agentConfig from demos.ts
The agentConfig in demos.ts has tools: [] (we emptied it) or incomplete tool schemas
AgentService.start() compiles a new workflow from that config and calls metadataDAO.updateWorkflowDef(def) — overwriting the correct definition the Python worker deployed
The LLM gets tools: [] and answers from memory

From the API (/api/workflow/thirteen_f_analyst):

Bypasses AgentService.start() entirely
Goes directly to Conductor's workflow executor
Uses whatever workflow definition is already stored in the database
If the worker deployed correctly, tools are intact and everything works

So the UI path is fundamentally broken for this agent because agent/start is designed to redeploy on every call — which works fine when the frontend has the complete tool definitions, but breaks when tools: []._

I can replicate at will and can provide steps on how to replicate