# TODO

Ordered backlog. `CURRENT.md` tracks the live session state.

## Where the project finished (2026-09-13)

- The validated core bridge is live: a fixed action list, input validation, structured errors, and a
  hard outbound rate limit. See `docs/core-bridge.md`.
- The daemon uses the core bridge. QR login, pacing, the trigger log, trust tiers, context, the
  response policy, the continuation window, the persistent action log, and the read-only voice
  state all work.
- Human pacing is live: a fast first send (30-200 ms), burst sends (500-2800 ms), and a variable
  5-20 s burst reset. Reads take a small gap. See `docs/daemon.md`.
- User search reads the client cache and guild members. `voice.get` searches all guilds without a
  `guildId`. `voice.overview` lists occupied channels per guild.
- The web UI follows the Nocturne design handoff: the 7a debug shell with the guild rail, message
  view, voice membership, agent console, Takeover read-only gate, and a Settings slide-over for
  core process, login, visible guilds, and the action log. Messages fetch the latest 50 on open;
  attachments use CDN links; avatars and guild icons load from the CDN. Drafts survive reloads.
  Roadmap surfaces show as placeholders. See `docs/webui.md`.
- `dm.list` lists cached DM and group DM channels. `list_dms` exposes it over MCP.
- The daemon starts, stops, and restarts the client, and the Settings panel runs the QR login.
- Mock end-to-end tests pass: tiers, continuation, sensitive replies, batch MCP, and voice state.
- Repo note: use `zib/Veskslop-t` (remote `veskslop-t`) until the original repo's server hook is
  fixed.

## Paused

- The `poll` command. `POLL_ENABLED` is `false` in `poc/src/puppet.ts`. The parser and tests stay.
- The setup command flow. The reaction-based multiple-choice design is on hold.
- The LLM attachment and MCP event delivery. Parked by the owner.

## Deferred

- `stay @user`, invite-link joins, audio in and out, and video streaming. The plan is in
  `docs/voice.md`. The basic voice actions exist; auto-leave and audio do not.
- Client lightening. Low priority. The client must stay stable.
- Web UI authentication and remote accounts. Accepted for now: no auth, LAN or Tailscale only.

## Next

1. Voice auto-leave (owner-based) is implemented. `stay @user` is deferred.
2. Live verification on the audio host: TTS local + API, ducking, capture, and `voice.join`/
   `voice.leave` end to end.
3. Browser microphone puppeting: a mix sink so the browser mic and TTS share `veskslop_mic`, a
   `pacat` input path, and echo handling.
4. Speech to text: Vosk (small CPU model in `poc/models/`) is wired and verified; add more
   languages or a larger model when needed. `VESKSLOP_STT_CMD`/`_API_URL` stay available.
5. Notification parsing (edits, deletes, reactions, replies) and the reply-to-puppet trigger: done.
6. Per-user rate limit and command cooldowns.
7. Collector account: stripped-down read-only modules. Use it when the puppet cannot find a user's
   voice channel or the channel members.
8. Local agent harness (`agent/`, Go + Eino): scaffold done and smoke-tested. Next: daemon task
   push (free-form triggers only; commands stay local), tool allowlist, per-conversation
   serialization with cancel-and-replace, delivery IDs, and tests against a fake model.
9. Remote agent MCP path: a notification tool set for external systems, a documented LAN/Tailscale
   bind, and optional token auth on the MCP listener. See `docs/agent.md`.
10. Mobile web UI: first draft at 390px with the four tabs (Users, Chat, Call, Activity) is
    implemented and rendered headless. Next: move the transcript into the Call tab, put the TTS
    box in the Call tab as in 1c, and verify with real data on a phone-width viewport.
11. TTS VOX preference: a `tts.vox` toggle that exports `VESKSLOP_TTS_VOX` (like `words`), an
    optional `VOX_CLEAN`, and a help line in the TTS panel. Requested in `TTS-DESIGN.md`; the
    wrapper side is ready and verified. Inline `/rate` and `/pitch` need no UI.

## Web UI follow-ups

Ordered after the voice backend unless the owner asks first.

- Favourites, presence dots, and real member counts.
- MCP tool-call tracing in the console Tool calls tab.
- Transcript tab fed by the audio capture and STT.
- Yield (stop mid-line) for TTS when a human speaks.
- DM scope in the header (the guild picker becomes "Direct messages").
- Message editing, deletion, reactions, and history pagination.
- The remaining mockup screens: 7b/7c console details, 6b history window, 4b group chat, 2c state
  legend.

## Open decisions

- **Web UI auth.** Accepted: no auth, LAN or Tailscale only. Revisit before any wider exposure.
- **MCP event delivery.** Parked. Options when it returns:
  1. Poll `get_message_log` and `get_action_log` with a `since` cursor. No new transport work.
  2. A stateful Streamable HTTP session. The server can send notifications.
  3. A separate local event stream over SSE or a WebSocket.
- **Untrusted logging.** Untrusted interactions are ignored at this time. The separate 30-entry
  buffer exists but receives nothing. Decide whether to log untrusted DMs and mentions again.
- **Privacy pass.** Only token-shaped and `mfa.` strings are redacted, and only in the persistent
  action log. Decide the fuller policy (retention, cache clearing, message redaction) later.