ylstack-agents-stack/TODO.md at main · ylstackmain/ylstack-agents-stack · GitHub
Skip to content

TODO.md

Latest commit

 

History

History
53 lines (34 loc) · 4.26 KB

TODO.md

File metadata and controls

53 lines (34 loc) · 4.26 KB
Raw

ylstack-agents-stack — TODO

Running list of work beyond v1. Each section is scoped to hand off to a single agent.

2. OpenClaw-style onboarding flow

Current first-run experience: the four identity files get silently seeded with our generic text, and the user is dropped into an empty chat. OpenClaw (the reference project) has a much better onboarding that introduces the agent and walks the user through setting up their identity.

  • Research OpenClaw's onboarding (repo: https://github.com/zeroclaw-labs/zeroclaw and original OpenClaw) — what prompts it uses, what it asks the user, how it writes the initial SOUL.md / IDENTITY.md / USER.md.
  • Copy OpenClaw's onboarding prompts directly into our seed content in src/worker/agent/core-files.ts to start. Customize afterward.
  • Build a first-run UX: detect that USER.md is still empty (or has only the seed template), and route the user into an onboarding sequence that walks through filling in the four files — probably a guided multi-turn chat where Claw asks questions and writes to the files as it learns.

Agent brief: Read OpenClaw's source to extract its onboarding prompts. Replace our current seed content with OpenClaw's. Design and build a first-run guided onboarding flow that the agent drives.

6. Tiptap for markdown editing

Current editor is a plain <textarea> (src/components/markdown/MarkdownEditor.tsx). Upgrade to Tiptap for identity-file editing and for the onboarding flow.

  • Tiptap — https://tiptap.dev — is a headless rich-text editor framework on top of ProseMirror. Tiptap v3 supports React.
  • Use the StarterKit extension plus a Markdown-aware extension (either @tiptap/pm with a markdown serializer or a community tiptap-markdown package).
  • Reference the todo-app's TiptapTodoInput.tsx and its DateTokenDecoration extension for our in-house patterns, but we'll want full markdown support rather than single-line.
  • Research with context7: fetch current Tiptap docs for Markdown and React usage.

Agent brief: Research Tiptap's current markdown editing story (use context7 if available, otherwise WebFetch). Replace MarkdownEditor with a Tiptap-based editor while keeping the same { value, onChange } interface. Use it for the four identity files and the upcoming onboarding flow.

8. Voice input via Whisper transcription

Let users record a voice message in the chat and have it transcribed via Cloudflare Workers AI's Whisper model before sending.

  • Cloudflare Workers AI hosts OpenAI's Whisper — model IDs: @cf/openai/whisper (base, multilingual) and @cf/openai/whisper-large-v3-turbo (faster / more accurate for longer audio). Verify current options in the Workers AI catalog at https://developers.cloudflare.com/workers-ai/models/.
  • Client: add a microphone button next to the send button in src/components/chat/InputBox.tsx. Use MediaRecorder + navigator.mediaDevices.getUserMedia to capture audio. Show a waveform or recording timer while active. Stop-and-send or stop-and-review flow (decide based on UX).
  • Server: add an endpoint (POST /api/transcribe or an agent RPC method) that accepts the audio blob, calls env.AI.run("@cf/openai/whisper", { audio: [...bytes] }) or similar, and returns the transcript. See Workers AI docs for the exact input shape (expects a byte array of the raw audio).
  • Insert the transcript into the input textarea. User can then edit before submitting, or auto-submit — configurable.

Agent brief: Confirm the current Whisper model IDs and request shape in Workers AI docs. Build the POST /api/transcribe endpoint that wraps env.AI.run. Add a mic button to InputBox.tsx that records, uploads, and drops the transcript into the input. Handle permissions denial, recording errors, and empty transcripts gracefully.

Parking lot / future

  • Exa → configurable: allow swapping to Brave or Tavily via env var.
  • Model selector: UI to pick a Workers AI model without editing wrangler.jsonc.
  • Context-block compaction for very long threads (Think's configureSession supports this; skipped for v1).
  • Sub-agent Facets for deep research delegation.
  • Conversation branching (Think's tree-structured sessions).
  • Scheduled HEARTBEAT.md routines via DO Alarms.
  • Deploy-to-Cloudflare button + one-click onboarding for distribution.