radiantlogicinc/fastworkflow/integrate-chat-agent

Integrate a fastWorkflow Chat Agent

Improve the integrated workflow

Use fastworkflow run_chatbot as the developer workbench for the application workflow. In an existing session, inspect benchmark → experiment → conversation (or ad-hoc conversations by UTC date) on the left, and turns/components on the right. The right breadcrumb retains the full path. Add component-level Human feedback to record what should improve.

For repeatable verification, use create-workflow-benchmarks to author benchmarks and task prompts and create experiment identities. The UI does not execute those tasks: connect an application runner using the experiment and task IDs. Use debug-workflow-conversations for diagnosis and optimize-workflow-with-feedback to verify changes against recorded outcomes. Human annotations do not automatically update the agent or train models. When a family of requests needs the same several commands in a known order, write a skill (_skills/<name>.py) with author-workflow-skills.

Mission

Your mission is to integrate agentic chat UI within the application's UI that allows natural

language conversations to execute the business logic functionality and provide human readable

answers to questions. All information that is exposed via the UI should also be available via

the chat interface. Use fastWorkflow (uv add fastworkflow) to implement the backend for this

chat functionality.

This chat window should pop up when an overlaid chat icon is clicked in the web UI. Users should

be able to start new chats and also continue previous chats. The chat window should stream

internal fastWorkflow conversations between workflow and assistant (same UX as fastWorkflow cli

invoked via fastWorkflow run) before giving the final answer.

fastWorkflow AI-enables an existing Python application by wrapping its classes and methods with an intent-detection + parameter-extraction + execution pipeline. The chat backend is hosted as the fastWorkflow FastAPI service, which exposes streaming HTTP endpoints the chat UI consumes.

Track all work in beads

Use beads (`bd`) for issue tracking — not markdown TODOs. Install it if missing (npm install -g beads or see beads docs). Create one epic for this integration and tasks under it that mirror the workflow phases below. Mark tasks in_progress / closed as you go.

bd create "Integrate fastWorkflow chat agent" -t epic -p 1 --json
bd create "Build & train fastWorkflow from app business logic" -t task -p 1 --deps discovered-from:<epic-id> --json
bd create "Host fastWorkflow FastAPI streaming service" -t task -p 1 --deps discovered-from:<epic-id> --json
bd create "Build popup chat UI with streaming + conversation history" -t task -p 1 --deps discovered-from:<epic-id> --json

Workflow

Copy this checklist and track progress:

- [ ] Step 1: Discover the application's business logic to expose
- [ ] Step 2: Install fastWorkflow and scaffold the workflow directory
- [ ] Step 3: Write command files for each app capability
- [ ] Step 4: Set up env files — PAUSE for the user to add API keys
- [ ] Step 5: Train the workflow's intent models
- [ ] Step 6: Smoke-test the agent from the CLI
- [ ] Step 7: Host the fastWorkflow FastAPI streaming service
- [ ] Step 8: Build the popup chat UI (new/continue chats + live trace streaming)
- [ ] Step 9: End-to-end verification

Step 1: Discover the business logic to expose

Enumerate every capability the app's UI exposes (and any read-only "questions" users may ask). The hard requirement: everything available in the UI must also be reachable through chat. Map each UI action / data view to a Python class + method or function in the app's codebase. Group related operations into classes (contexts) that hold state. This mapping becomes the command set.

If the app's backend is not Python, fastWorkflow still runs as a separate Python sidecar service; the commands call into the app via its existing API/SDK/DB client instead of in-process imports.

Step 2: Install fastWorkflow and scaffold

uv add fastworkflow            # or: pip install "fastworkflow[server]"
fastworkflow examples fetch hello_world   # provides fastworkflow.env + fastworkflow.passwords.env templates

Create the workflow directory next to the app code:

<app>/chat_workflow/
├── application/        # symlink or thin wrappers calling the app's real business logic
├── _commands/          # generated + hand-edited command files
├── fastworkflow.env
└── fastworkflow.passwords.env

Add ___workflow_contexts, ___command_info, ___convo_info to .gitignore.

Step 3: Write the command files

For each capability mapped in Step 1, create a command file in _commands/<command_name>.py that invokes the app's business logic, and declare it in _commands/context_inheritance_model.json. Group commands into contexts (classes) that hold state. Use Pydantic Input/Output signatures with strong Field descriptions, examples, and a default="NOT_FOUND" so missing parameters are detected rather than hallucinated. See reference.md for the command-file structure and context-model format.

Strong field descriptions are the floor, not the finish. Four colocated skills cover the features that decide whether the agent can actually complete a multi-step task, and are worth reading before writing the first command rather than after the first bad demo:

The first three form one pipeline per parameter: available_from says where a missing value comes from, db_lookup fixes one that arrived nearly right, and validate_extracted_parameters rejects or normalises one that arrived unusable. A parameter worth taking is usually worth declaring in at least one of the three.

Step 4: Set up env files — PAUSE for API keys

Copy fastworkflow.env and fastworkflow.passwords.env into the workflow directory. The service needs LLM API keys to run. Stop here, tell the user the absolute path of `fastworkflow.passwords.env`, and ask them to add the keys. Do not attempt to invent or commit keys. Minimum keys (Mistral small / OpenRouter free tiers work):

LITELLM_API_KEY_SYNDATA_GEN=...
LITELLM_API_KEY_PARAM_EXTRACTION=...
LITELLM_API_KEY_RESPONSE_GEN=...
LITELLM_API_KEY_PLANNER=...
LITELLM_API_KEY_AGENT=...
LITELLM_API_KEY_CONVERSATION_STORE=...   # FastAPI conversation topic/summary

Step 5: Train

fastworkflow train <app>/chat_workflow <app>/chat_workflow/fastworkflow.env <app>/chat_workflow/fastworkflow.passwords.env

Training generates synthetic utterances and trains intent models into ___command_info/. It takes several minutes and the first run also downloads HuggingFace models.

Step 6: Smoke-test from the CLI

fastworkflow run <app>/chat_workflow <app>/chat_workflow/fastworkflow.env <app>/chat_workflow/fastworkflow.passwords.env

Ask "what can you do?" and exercise a few real commands. Prefix a command with / to force deterministic (non-agentic) execution. Confirm the agent reaches every UI capability from Step 1.

Step 7: Host the FastAPI streaming service

The chat backend is the bundled FastAPI-MCP service (requires the server extra). It exposes JWT auth, conversation persistence, and the streaming endpoint that powers the live trace UX.

python -m fastworkflow.run_fastapi_mcp \
  --workflow_path <app>/chat_workflow \
  --env_file_path <app>/chat_workflow/fastworkflow.env \
  --passwords_file_path <app>/chat_workflow/fastworkflow.passwords.env \
  --port 8000

Run this alongside the app's backend (e.g. as a sidecar/separate process) and proxy it from the app's server, or call it directly from the frontend. Full endpoint contracts (request/response shapes, auth flow, streaming format) are in reference.md.

Step 8: Build the popup chat UI

Required UX (matches the fastWorkflow run CLI experience):

1. Overlaid chat icon in the app UI that toggles a chat window.
2. New chat button → POST /new_conversation.
3. Continue previous chats → GET /conversations lists past conversations;

POST /activate_conversation restores one into the window.

1. Stream internal conversation → send each user message to POST /invoke_agent_stream and

render the trace events live (the back-and-forth between workflow and assistant) before rendering the final output event. This is the core differentiating UX — do not collapse it into a single final answer.

Auth flow: call POST /initialize once per user/session to get a JWT, send it as Authorization: Bearer <token> on subsequent calls, and refresh via POST /refresh_token.

Step 9: End-to-end verification

• Click the chat icon → window opens.
• Send a natural-language command → trace events stream, then a human-readable final answer.
• Verify every UI capability from Step 1 is reachable through chat.
• Start a new chat, send messages, then reopen a previous conversation and confirm history restores.
• Close any open bd tasks with a reason.

Key references

• HTTP API contract, auth, streaming format, command-file & context-model structure: reference.md
• fastWorkflow concepts, CLI, env vars: project README.md