Proposed workflow · Quicksilver + Hermes

One conversation, every device.

A proposal for moving the Obi-Wan/Kenobi workflow from Claude/Kenobi toward Quicksilver running through Hermes: the same assistant reachable from the X1 keyboard, Framework, MacBook Air, iPhone, Apple Watch, ESP32, Rabbit R1, and Telegram.

View architecture Migration phases
North star

An ambient personal agent, not another app.

The point is simple: Léo should be able to ask Quicksilver for help from whichever device is closest, and Quicksilver should be able to answer back through the most useful channel: text, voice, Telegram, Obsidian, Fastmail, calendar, or a published page.

Architecture

Proposed Quicksilver/Hermes flow

Draft for review
Quicksilver Hermes Workflow Inputs flow through device adapters into obi-wan-core, then Hermes/Quicksilver, model routing, tools, memory and output channels. Apple Watch Shortcut → /message iPhone Shortcut / Share sheet ESP32-S3-BOX wake → PCM /talk Rabbit R1 OpenClaw WS Telegram Quicksilver bot Keyboards X1 / Framework / MBA obi-wan-core device adapters stay stable auth + routing STT/TTS bridge Local media layer Whisper STT :8002 Piper/Kokoro TTS :8888 Hermes Gateway Quicksilver persona current model router tools + skills + memory Context stores Obsidian / Open Brain Fastmail PIM project repos + files Model/tool layer ChatGPT 5.5 now GLM writer optional Qwen-coder optional local Ollama optional no Claude API dependency Outputs Telegram replies voice on device Obsidian notes email/calendar/pages authenticated input handoff model call read/write context deliver back speech/audio return path
Inputs

All roads lead to Quicksilver.

Apple WatchShortcut captures voice/text and posts to the existing obi-wan-core /message endpoint. Keep the fast wrist workflow; change the backend brain.
iPhoneShortcut and possible share-sheet actions send text, URLs, dictated notes, or files into the same conversation channel.
ESP32Wake word records PCM, posts to /talk, uses local Whisper for STT, Hermes for reasoning/action, and TTS for spoken replies.
Rabbit R1Keep the OpenClaw/WebSocket adapter as a device frontend; route its intents to Quicksilver instead of Claude-specific channels.
TelegramMove from Claude’s Kenobi account to the Quicksilver Telegram account as the main async command-and-response channel.
X1 / Framework / MacBook AirUse Hermes CLI/TUI directly for rich keyboard sessions, plus the same memory, skills, and project context as the mobile/voice flows.
Processing

Replace Claude API calls with Hermes calls.

Short conversational turns

obi-wan-core can keep routing device traffic, but the reasoning call should go to Hermes using its currently configured model, now ChatGPT 5.5.

Heavy delegated work

Replace spawn_claude_code with spawn_agent or spawn_hermes_task. Hermes gets a self-contained task, picks the right skill/toolset, and reports back.

Model specialization

Hermes becomes the model router. ChatGPT 5.5 can be default; GLM 5.1 can be selected for polished writing; Qwen-coder for focused coding.

Memory and context

Use Hermes memory, Open Brain, Obsidian, Fastmail, calendar, project files, and session recall as shared context rather than scattering state across bots.

Outputs

Replies should land where they are useful.

Immediate reply

Watch/iPhone/Telegram/R1 get concise conversational answers in-channel.

Spoken response

ESP32 and local voice paths return audio through the existing TTS service when voice is the right medium.

Durable artifact

Longer work writes to Obsidian, a project repo, Fastmail draft, calendar event, or a public page at pages.laporte.cloud.

Async callback

Background work should notify the Quicksilver Telegram account when complete, with links to any created artifacts.

Migration phases

Low-risk path from Kenobi/Claude to Quicksilver/Hermes.

Phase 1Switch Telegram from Claude’s Kenobi bot to the Quicksilver Telegram account. Keep message behavior the same.
Phase 2Replace direct Claude API calls in obi-wan-core with Hermes calls using Hermes’ active model configuration.
Phase 3Replace spawn_claude_code with a backend-neutral spawn_agent tool whose first new backend is Hermes.
Phase 4Add reliable async delivery: spawned Hermes tasks report completion to Telegram and/or write durable notes/artifacts.
Phase 5Decide whether obi-wan-core remains the device router or becomes a thin adapter into Hermes gateway/webhooks.
Preserve Watch Shortcut Preserve ESP32 firmware contract Use Quicksilver Telegram No Claude API dependency Hermes model router
Review questions

Decisions before implementation.

  • Should Telegram become the canonical async return channel for all devices?
  • Should the Watch/iPhone endpoint keep the name /message, or become /quicksilver/message later?
  • Should ESP32 responses always be spoken, or optionally also forwarded to Telegram?
  • Should Hermes run as a CLI subprocess first, or should we enable Hermes webhook/gateway support before coding?
  • Which model personalities should be explicit: default, writer, coder, researcher, local/private?