lumbridge-code/docs/RESEARCH.md

# Research map

The working checkouts live outside this repository at:

```text
Desktop/Lumbridge Code/Research/
```

They are shallow snapshots for study, not dependencies or vendored source.

## Multiplexers and terminal engines

- `manaflow-ai/cmux`: the current cmux product; native workspace/pane model,
  automation socket, embedded browser, agent hooks, and terminal UX. The
  captured tree is GPL-3.0-or-later, so it is behavior/UX study only.
- `Helvesec/rmux`: Rust multiplexer engine, daemon, typed SDKs, tmux surface.
- `tmux/tmux`: the durable client/server and command model to remain compatible
  with where useful.
- `zellij-org/zellij`: Rust workspace, layouts, plugins, collaboration, web client.
- `wez/wezterm`: mature Rust terminal emulator, multiplexer, rendering, and PTY
  portability reference.

## Agent IDEs and orchestration products

- `stablyai/orca`: parallel worktrees, many harnesses, terminals, usage tracking,
  remote execution, and review surfaces.
- `get-bb/bb`: self-drivable agent IDE with desktop, web, CLI, and API surfaces.
- `coder/xum`: isolated parallel agent workspaces, remote execution, review, and
  cost/context UX; AGPL, so treat implementation as study-only.
- `zed-industries/zed`: high-performance Rust editor, GPUI, terminal, and ACP
  client. Review the dual-license boundary file by file.
- `aaif-goose/goose`: Rust desktop/CLI/API agent, many providers, MCP, and ACP.
- `block/buzz`: Rust human/agent collaboration plus production ACP harness design.
- `bbarit/terminal`: Rust agent terminal/IDE ideas. No top-level license file was
  present in the captured snapshot, so do not reuse implementation code.
- `OnlyTerp/opengrok`: provider-wire maps, behavioral verification of reasoning
  controls, local-key posture, drift detection, and fail-closed provider UX;
  MIT, but its Grok Bot patching is product-specific reference work.
- `elie222/rakazo`: persistent AI teammates, delegation, routines, computer
  providers, browser/terminal/desktop tools, and deterministic sandbox testing;
  Apache-2.0. Its Electron/React/Postgres architecture is research-only because
  Lumbridge's shell and local runtime remain native Rust and SQLite.

## Protocols

- `agentclientprotocol/agent-client-protocol`: canonical ACP schema and spec.
- `agentclientprotocol/rust-sdk`: official high-level Rust client/agent runtime.

## Harnesses

- `openai/codex`
- `deepseek-ai/deepseek-harness`
- `google-gemini/gemini-cli`
- `badlogic/pi-mono`
- `NousResearch/hermes-agent`
- `openclaw/openclaw`
- `anomalyco/opencode`

DeepSeek Harness at the captured revision is MIT-licensed experimental software,
but it is a large Node/TypeScript workspace with a React web client, Cordis
plugin tree, native/platform payloads, and a broad third-party closure. Its own
safety notice says it is unaudited and can access processes, files, credentials,
and the network exposed to it. We will study its service seams, durable event
log, approval flow, ACP profile, and agent lifecycle. We will not embed its web
shell into Lumbridge or treat its sandbox as a security boundary. Any branded
fork is a separate repository/process with preserved notices and an explicit
dependency and payload review.

## Research questions

1. Which terminal core gives us correctness without inheriting a product-shaped
   dependency or incompatible license?
2. Can the runtime maintain smooth human input while agents read/write other
   panes through automation?
3. Which UI stack meets latency, IME, accessibility, and packaging requirements?
4. What is the smallest honest common denominator across ACP harnesses?
5. Which official provider/harness surfaces expose subscription quota, and what
   must remain measured or unavailable?
6. Can one event model cover PTY, ACP, Git, usage, permissions, and recovery
   without turning into untyped JSON everywhere?