Agents

The Agent pane is a streaming chat surface. Behind the picker, Loom routes prompts to one of two provider families:

Family	Providers today	How auth works
CLI subprocess	Claude Code (Codex, Gemini stubbed)	Piggybacks on the CLI's existing OAuth login. No API key in Loom.
HTTP streaming	Ollama, OpenAI-compatible, Anthropic API	Configurable base URL, optional bearer token in Keychain.

Picker

The picker in the Agent pane's header groups agents by source:

• Default — Claude Code's vendor default.
• Plugin agents / Built-in agents — sub-agents discovered via claude agents list.
• Local · <endpoint name> — one entry per Ollama model, or one per OpenAI-compatible endpoint, fed by Settings → Providers.

Click the refresh icon next to the picker to re-query — useful after claude agents adds a new sub-agent or after ollama pull lands a new model.

Streaming behavior

Provider	Streaming	Cancel
Claude Code	One-shot (full response when subprocess exits)	Stop button terminates the process
Ollama	Live token stream	Stop button cancels the URLSession task
OpenAI-compatible	Live token stream (SSE)	Stop button cancels the URLSession task
Anthropic API	Live token stream (SSE)	Stop button cancels the URLSession task

The "..." placeholder in an empty assistant bubble means the request is in flight.

Conversation history

• CLI providers retain context server-side via --resume <session-id>. The session id is shown in the header (Claude Code · 3328C421).
• HTTP providers are stateless. Loom replays the full chat history with each turn, so context survives but you'll see longer requests as the conversation grows.

Switching between providers inside one workspace creates a new logical conversation — the new provider doesn't get the prior history of the previous one. Open a fresh workspace to start clean.

Per-workspace state

Each workspace renders its own Agent pane with local message state. Closing and reopening a workspace preserves the message log (in-memory; not persisted across app relaunches today).

Workspace-aware proposals

The Agent pane reads the active workspace + active tab via WorkspaceContext. In Ideas workspaces this surfaces an inline proposal card under list-shaped responses, letting one click commit the items as new note tabs or append them to the open one. See Ideas workspace → Save-to-tab proposal cards for the user-facing behavior.

Streaming itself is also fast-pathed: token deltas update an isolated StreamingBubble view (not the whole message list), so the chat stays smooth even on Ollama's faster local models.

Workspace context block

Every prompt the Agent pane sends carries a workspace snapshot so the model can ground its answer in the project you're actually sitting in. The snapshot is rebuilt at send-time and includes:

• Workspace name + kind (Loom (Ideas), vendetta (Prompt), …).
• Project folder path when the workspace has one configured.
• Project memory — Loom reads CLAUDE.md, AGENTS.md, GUIDE.md, and README.md from the workspace folder (priority order; capped at ~5 KB total) and pastes them in.
• Active idea tab name + body (Ideas workspaces only).
• Sibling idea tabs — titles plus a short excerpt of each, so the agent doesn't propose ideas that are already captured.

Wiring is per-provider:

• Anthropic / Ollama / OpenAI-compatible receive the snapshot via the system prompt every turn.
• Claude Code / Codex / Gemini subprocess agents don't take a separate system message, so Loom prepends the snapshot under a ## Loom workspace context heading and ends the prompt with a ## User request section before the user's question.

Result: in an Ideas workspace pointed at the Loom folder, asking "give me 8 ideas" returns ideas grounded in Loom's README.md + GUIDE.md instead of generic suggestions, and the proposal card lands them straight in the active tab.