Agent console

The agent console has shipped its first two phases: the agent-console-server host, the Claude Agent SDK driver, the default theme, and a full-screen world under its console overlay at terminal parity. One page of the app now works every Claude Code session on a machine in place of the terminal. This proposal is what comes after those phases. It is judged by the workflow comparison: nothing here is built while the console still sends the person back to a terminal during a day's work.

Decisions

• Every view keeps the world's budget. A view builds only what is in view and rebuilds only what a change touched, the rules the Genshin engine keeps, so no cost grows with the repository behind it.
• The look is behind a theme. The default theme is the world with no character. A theme may add a palette, a scene in the world, an avatar, reactions and a voice, and Genshin is the first to add them (themes).
• Views are separate from themes. A view is a panel any theme can show, such as the collector harbour, so a repository's tooling is visualised whatever the console is dressed as.
• Other tooling joins through tiers, cheapest first. App routes open in a side pane with no code, external tools arrive through MCP Apps, and a first-party view is written only when a tool needs the scene (extensions).
• The world is Genshin's. The console's world is the Genshin recreation, built by its own program; the voxel world's open world, building, map, day, sound, atelier and codebase city are rejected with it.
• A second driver, for sessions the SDK cannot hold. A session a terminal already runs is attached to from the outside (terminal-mirror driver).
• The page manages; each host runs. The page is the manager of every session on every host it has paired — this computer's, and each remote connection — and runs nothing itself. A host holds its sessions, its Claude Code login and its repositories, and a page reloading, a hot reload in development, or a closed tab never ends a turn, since the page reconnects and is replayed what it missed.
• Both ends are live, and either can act. A session started, stopped or answered in the page shows in the host's window at once, and a host stopped from its own window tells every page before it exits (host installer). A host that dies without a word is the one case the page shows as not answering and retries.
• Nothing runs hidden. This computer's host and each of its sessions run in windows the reader can see and close (host installer, session windows): no login item, no tray, no background service. Stopping a session is closing its window, stopping the host is closing its own, and nothing is left running after either. A remote connection's sessions stay inside their host, since there is no desktop there to see them on, and are stopped from the page.
• Nothing that costs money, Windows first. The installer ships for Windows alone and unsigned, and every secret the app keeps is encrypted with a key the app already holds rather than a paid vault. Signing and the other platforms wait on signed host installers.
• TresJS first; raw Three.js only where TresJS and cientos have nothing, said on the page of the view that reaches for it, with the reason.

How it works

flowchart LR
  subgraph Browser
    W[Genshin world and panels] --> T{Theme}
    T -->|default — shipped| N[Notifications]
    T -->|Genshin| G[Palette, scene, avatar, voice]
    W --> V[Views: harbour]
    W --> X[Side pane: app routes, MCP Apps]
  end
  subgraph Host[agent-console-server — shipped]
    D{Driver} -->|Agent SDK — shipped| C[Claude Code session]
    D -->|terminal mirror| M[A terminal's own session]
  end
  W <-->|contracts| D

How the page and its hosts stay in step — the principle every connection sub-spec builds to:

sequenceDiagram
  participant P as Page — the manager
  participant A as Host A — its own window
  participant B as Host B — another window or machine
  P->>A: Start a session
  A-->>P: Session opened, events stream
  A-->>A: Window logs the session
  P->>B: Stop a session
  B-->>P: Sessions changed
  Note over P: The page reloads — hot reload, a crash
  P->>A: Reconnect with its credential
  A-->>P: Every open session's log replayed
  Note over B: The reader closes B's window
  B-->>P: Host stopping
  P-->>P: B shown stopped, its sessions closed
  Note over A: A dies without a word
  P-->>P: A shown not answering, retried with backoff

The pages of this proposal

Scope and order

1. One day's work in the console alone, now that terminal parity has no gap left, with every return to the terminal written into the workflow comparison. Connecting comes first, since it is the first thing every reader does: the host installer, session windows, device pairing and remote connections to a machine of the reader's own have shipped, and Anthropic-hosted sessions come later. The Code tab's lean core then closes the returns that comparison already predicts: the shell pane has shipped, and repository files, diff comments and side chat come next, with effort level and session titles as small changes to the composer and the sessions tab.
2. The Genshin theme: persona, voice, wish banner, then the ambience, all inside the world.
3. Views: the collector harbour.
4. The terminal-mirror driver, when a session started in a terminal needs to be picked up.

What this does not propose

• Agents hosted by Esposter. A hosted agent needs a sandboxed checkout, the user's key held server-side and compute the app would pay for. Every console surveyed runs the agent where the code already is, and so does this one.
• A renderer of the character's Live2D model. The desktop viewer draws it (own Live2D renderer, deferred).
• A new chat product. The console works the session the code is in; it is not a chatbot beside it.
• The rest of the Code tab. Worktree sessions, remote sessions, split sessions, cross-session messages and the session archive are deferred behind their triggers. A browser pane and its other previews, a pull request bar, view modes, computer use, scheduled tasks, a customize panel, an environment editor and a pane layout are rejected.

Key files

Notes

• The trade is stated rather than hidden. The terminal costs nothing to keep, while the console is a host, a page and a theme layer to maintain. It is taken because the terminal cannot show a diff, a context gauge, parallel sessions or a scene, and because the console can be played.
• This supersedes the page half of chat into the session: with a driver holding the session, a line typed at the character is simply a prompt. The channel stays only as the terminal-mirror driver's input.