Moonkale

Agent Sessions and Profiles

5 min read

Built in Milestone 15 - Agents, Profiles and Connections from Prompt24. Background: LLM and RAG, Claude Code Extension.

Saved agents (profiles)

A saved agent is a name plus a language model: provider, model, endpoint or command, secret name, provider options (Claude Code’s permission mode and allowed tools). They live in settings:

{
  "llm":    { "provider": "openai", "model": "mistral-code-latest", "base_url": "https://api.mistral.ai/v1", "secret": "mistral" },
  "agents": [
    { "name": "Claude", "provider": "claude-code", "options": { "permission_mode": "acceptEdits" } },
    { "name": "Local",  "provider": "ollama", "model": "qwen2.5-coder" }
  ],
  "agent":  { "default": "Claude" }
}
  • The flat llm block is the profile called Default — every settings file, environment override (MOONKALE_LLM, OPENAI_BASE_URL, …) and test from before keeps its meaning. agents[] adds more. Both scopes may contribute; a workspace entry with the same name overlays the user one field by field.
  • agent.default names the one that runs unless a session says otherwise; an unknown name means Default. The resolved Settings.llm is the default agent’s settings, so search embeddings, server sessions and everything else that asked for “the LLM” are unchanged.
  • The embedding model belongs to search, not to an agent: it is set once (Settings → Search) and served by Default’s provider; every profile carries it.
  • Keys are never in settings: the secret name is, the key comes from MOONKALE_SECRET_<NAME>, the classic environment variables or the secrets file (Security).

Settings → Agents shows one card per profile: name (Default’s is fixed), runs by default, provider, model, endpoint/command, secret; a Claude Code card also shows the CLI’s status and a Log in button. Add agent appends Agent N (mock) to edit; Remove forgets a card.

Sessions

The Agent panel holds any number of sessions. Each has its own transcript, agent (moonkale_llm::Agent with the conversation), provider built from its saved agent, busy flag, pending approval and audit. The toolbar: the session select (● marks one with a turn in flight; below a line, the folder’s saved sessions), New, the profile select for the current session (disabled while it runs), Activity, Save (as a markdown page), Clear, × (close; the session stays saved).

Concurrency. A session runs one turn at a time — its compose box is closed meanwhile — and any number of sessions run at once. A turn is a root-owned task (spawn_forever), so closing the panel, switching sessions or opening another file does not stop it; an approval waits in that session until you come back and answer. With Claude Code, each running session is its own claude process in the folder. The same rule holds on the server: one thread per turn, one turn per session (Remote and Server Modes).

History. When a turn ends, the session is written to the store of the folder’s host (Milestone 18 phase 5.15: a row of agent_sessions under folder · "local" · id, Internal State; before that, and without a store, .moonkale/agent-sessions/local/<id>.json, which is imported once) — its messages (what the model saw), the rendered items, profile, title (the first message) and time. Opening a folder lists these under saved in this folder in the session select; picking one restores the transcript, reconnects its profile and continues the conversation. The folder source’s ls query dialect lists that hidden directory (Children skips hidden and ignored entries on purpose). The server’s own sessions (agent.on_server) are a head row plus one row per transcript item in the server’s store (formerly the append logs <id>.jsonl) — a different shape, because the server records approvals and streams; a served folder lists those, not the local snapshots.

Claude Code: status and login

The claude-code provider reports its readiness (Provider::status): claude --version and claude auth status --json → not installed (with the install command), not logged in, or logged in as … (claude.ai). Settings shows it on the profile card, with Check again. Log in runs claude auth login as a terminal tab (WorkspaceConfig::spawn_program, desktop): the CLI opens the browser, you sign in and paste the code into that tab — the flow Daniel expected in Prompt24, without Moonkale ever touching the credential. On the web the CLI runs and logs in on the server (/api/llm/status reports its state; the button is absent there).

Saved SSH connections

Open Remote Folder… keeps connections in the user file (remote.saved[]: name, the host as typed after ssh, the folder): a select fills the fields, Save stores the current ones under a name, Forget drops one. The user file only — a workspace file lives inside the folder that is being opened remotely.

Where things are

  • ext-api/settings.rs: AgentProfileFile, RemoteFile/SavedConnection, AgentProfile, DEFAULT_AGENT, Settings::agent(name), the merge (overlay_llm).
  • ext-api/workspace.rs: remote_saved/save_remote/forget_remote, spawn_program/run_in_terminal, write_text_at/list_at.
  • editors/agent/src/panel.rs: Chats (root context), Session, new_session/connect/run_turn/persist/load_saved/restore; server_panel.rs: the profile select feeding TurnSettings.llm.
  • llm/src/provider.rs: ProviderStatus; claude_code.rs: cli_status; server/src/llm.rs: llm_status.
  • shell/src/settings_panel.rs: AgentProfileCard, ClaudeStatus; shell/src/remote_dialog.rs: the saved-connection rows.
  • Tests: ext-api unit tests (merge, default, connections), project-fs (ls), llm (cli_status on the mock CLI), shell/tests/remote_dialog.rs (harness), E2E agents.mjs.

Graph View

Backlinks