Automate Without Code: Visual Agent Workflow Builder

Summary
A drag-and-drop workflow editor in the console where users chain blocks — trigger → fetch/scrape → LLM step → condition → write file / send message — into repeatable automations, no programming required.

Component(s) Affected
done
Core / Backend (app, agents, config, providers, utils, local_models)
done
Console (frontend web UI)
not done
Channels (DingTalk, Feishu, QQ, Discord, iMessage, etc.)
done
Skills
not done
CLI
done
Documentation (website)
not done
Tests
not done
CI/CD
not done
Scripts / Deploy
Problem / Motivation
Today, multi-step automation in QwenPaw means writing a skill (code) or hoping the agent improvises the right tool sequence every time. Non-developers — exactly the audience the desktop app courts — can't build "every morning, pull these three pages, summarize changes, and message me on Telegram" without help. Deterministic, repeatable pipelines also waste fewer tokens than free-form agent loops re-deciding the plan each run.

Proposed Solution
Flow editor UI: nodes for Trigger (manual / cron / channel message), HTTP request, Web scrape, LLM instruction (pick agent + model), Condition/branch, Read/Write file, Run skill, Call MCP tool, Send to channel.
Variables passed between nodes; per-node output preview for debugging; run history with per-step logs.
Flows saved as JSON in the workspace (versionable, shareable) and executable headless via cron — synergy with the existing crons API and 18 channels: flows that end in "send result to WhatsApp/Telegram" become a unique capability.
Export/import so the community can share flows.
Alternatives Considered
Skills: powerful but code-only; excludes most desktop users.
Prompt-engineering an agent to follow steps: non-deterministic, token-hungry, and fragile across model changes.
External automation platforms: heavyweight, and they can't reach QwenPaw's agents, skills, or channels natively.
Additional Context
The building blocks exist (agents, skills, MCP tools, cron, channels); this request is the glue layer plus UI. Even a minimal first version (linear flows, 6 node types) would unlock enormous value.

Willing to Contribute
not done
I am willing to open a PR for this feature (after discussion).