Choose a Code Agent Path
Yeaft exposes three execution paths in one browser. A connected Agent may run all three, but each path keeps its own runtime and compatibility boundary.
| Path | Best when you need | Runtime / boundary |
|---|---|---|
| Claude Code | Exact Claude Code tools, skills, MCP, compact/clear commands, sub-agent events, and resume behavior | One locally installed and authenticated Claude Code CLI process per conversation |
| GitHub Copilot | Copilot entitlement, ACP tool-permission prompts, and Copilot's model catalog | One locally installed and authenticated copilot --acp process per conversation |
| Yeaft Code Agent | Native provider routing, 1..N VPs, 33 built-in tools, scoped memory, Projects, sub-agents, and Work Center | Native engine inside yeaft-agent; it does not emulate every vendor CLI command |
Claude Code
Choose Claude Code when compatibility with the Claude Code CLI matters more than provider neutrality.
- The Agent starts a CLI conversation and normalizes its stream-json events for the Web UI.
- Claude Code owns its CLI session and command semantics.
- The Yeaft Web UI can display streaming text, tools, files, sub-agents, context actions, and resume history exposed by that provider.
- The Expert Panel is a Claude Code conversation helper; it is not the native multi-VP Session system.
GitHub Copilot
Choose Copilot when the Agent machine already has an eligible Copilot setup and you want ACP behavior.
- The Agent starts
copilot --acpand translates ACP events to the shared browser renderer. - Available models follow the live Copilot catalog and the local account entitlement.
- Tool calls can require an allow-once, allow-always, or deny response.
- Copilot persistence and unsupported commands follow the installed CLI, not Claude Code behavior.
Yeaft Code Agent
Choose the native engine when you need product-level orchestration rather than exact CLI compatibility.
- A Session has 1..N reusable VPs and one durable timeline.
@mentionscan fan one turn out to several VPs;RouteForwardrecords explicit peer handoffs.- Native turns use a bounded window of their own Session history; legacy Dream/H2-AMS memory and sibling Session summaries are not injected.
- Native providers route through Anthropic Messages or OpenAI Responses adapters, including supported GitHub Copilot dynamic credentials and compatible gateways.
- The current built-in registry has 33 tools; Skills and MCP may add more.
- A Session can create an Agent-level WorkItem for durable, planned, recoverable work.
How to create each one
Use New chat in the unified sidebar:
- Select the Agent that owns the target directory.
- Select Claude Code, Copilot, or Yeaft.
- Enter the working directory and runtime-specific options.
- For Yeaft, choose the Session roster/default VP. After creation, choose model/effort in the composer and edit the announcement in Session settings.
The sidebar catalog can show conversations from multiple Agents. Runtime identity always includes the Agent; the same Session ID on two Agents is not the same Session.
The unified Projects / Recents layout is present from the first paint. While the catalog is pending, it shows a loading state rather than a temporary legacy list. Reconnecting with the same authentication retains the existing groups and rows, but creation, selection, and changes wait for a fresh snapshot. Changing accounts or re-authenticating clears the previous catalog. Accepted inventories from older Servers are adapted to the same layout; an unconfirmed catalog stays loading instead of showing a misleading empty list.
Common choices
- Exact Claude Code workflow or Claude-specific skills → Claude Code.
- Existing Copilot subscription and ACP permission flow → GitHub Copilot.
- One provider-neutral coding assistant with memory → Yeaft Session with one VP.
- Parallel developer/reviewer/research roles → Yeaft Session with several VPs.
- Persistent goal with Action planning, waiting, retry, or recovery → Work Center, often created from a Yeaft Session.