An ACP agent inside the BEAM, for driving a Managoat.ACP.Peer in tests.

It is the other end of the transport seam: writer/1 is the function a test hands to Peer.start/1, and every frame the peer writes arrives here, is decoded, and is answered the way a well-behaved adapter would answer it. Replies go back into the peer through Managoat.ACP.Peer.stdout/2 once connect/2 has told the agent which peer it is talking to; frames written before that (the initialize the peer sends on start) are held and answered on connect.

Ships in lib/, like the other Managoat libraries' fakes, so a host's own tests can run a real peer against it without a sandbox, a port or a stub.

{:ok, agent} = ScriptedAgent.start_link(updates: [text_chunk("hello")])

{:ok, peer} =
  Peer.start(
    owner: self(),
    writer: ScriptedAgent.writer(agent),
    ref: make_ref(),
    prompt: "say hello",
    mode: :run,
    session_id: nil
  )

:ok = ScriptedAgent.connect(agent, peer)

assert_receive {:acp, _, {:session, "scripted-session"}}
assert_receive {:acp, _, {:lines, "acp", line}}
assert_receive {:acp, _, {:done, "end_turn", nil}}

What it answers

A :permission is %{"toolCall" => …, "options" => […]} in the protocol's own shape; the agent sends it as a request and waits for the peer's answer before streaming the updates, exactly as a blocked adapter would. The answer's outcome is reported to the observer.

What the observer sees

Every frame the peer writes is reported to :observer (the process that started the agent, by default) as {:scripted_agent, :wrote, decoded}, so a test can assert on the wire as well as on what the owner received; and a permission answer as {:scripted_agent, :permission_answered, outcome}. Requests the agent sends carry ids from its own counter, which starts at 0 per turn — claude-agent-acp's measured habit, and the reason the peer mints public request ids.