agent: sign in to an agent from the chat, through ACP

[foxnne]

Part of fizzyedit/plans plans/AGENTS_ANYWHERE_PLAN.md, step 5 (sections 3–4).

What changes

• initialize now reads the agent's authMethods and agentCapabilities.auth.logout, and
  advertises clientCapabilities.auth.terminal while some plugin registers terminal.run.
• When session/new, session/resume, session/load, or a prompt after a sign-out comes back
  auth_required (-32000), the chat shows the agent's methods as buttons through chat.ask. It
  no longer prints "The agent refused to start". Any other refusal, or auth_required from an
  agent with no methods, still gets the old note.
• agent method: the chat sends authenticate { methodId } and shows "Waiting for to
  sign in...". On success it opens the session again, resuming the saved one if there was one, and
  anything queued is sent. On failure it says why and shows the buttons again.
• terminal method: the chat runs the configured agent command plus the method's args,
  with its env, through terminal.run. When that exits 0 the chat restarts the agent. Any other
  exit shows the buttons again. A terminal method is never sent to authenticate.
• Sign out: a new palette command, "Chat: Sign Out of the Agent" (agent.sign_out), sends
  logout. It is enabled only while a running agent has auth.logout.
• Stop while signing in stops the agent; the next message starts it again. A prompt refused for
  a sign-in is sent again once signed in. Terminal methods are shown only while terminal.run
  exists.
• The chat never sees a token: the agent keeps its own credentials.
• acp.Inbound.response now carries the error code.

Named seam: terminal.run

No plugin runs a program in a terminal pane yet: ghostty's plugin registers no commands. So
src/auth.zig defines the contract a terminal plugin fills:

• Call: terminal.run { title, argv, env, cwd, done, tag }.
• On exit, the terminal plugin calls the done command with { tag, code }.

Until a plugin offers it, the chat does not advertise auth.terminal. The spec then says the
agent should not offer terminal methods. If an agent offers them anyway, the chat shows them,
and the chat leaves it out of the buttons. If that leaves no buttons, the chat says to sign in
with the agent's own CLI.

Claude's ACP adapter (0.93.0) offers only terminal methods (claude-ai-login,
console-login: --cli auth login …). So signing a logged-out Claude in needs the ghostty
follow-up. A logged-in Claude is unaffected.

Checked against the spec

agentclientprotocol.com/protocol/authentication and the schema's ErrorCode
(auth_required = -32000), on 2026-10-19:

• type absent means agent.
• A terminal method's env is an object map.
• logout takes {}, and is sent only when agentCapabilities.auth.logout is non-null.

Verify

• zig build test: new tests in src/auth.zig cover method parsing (unknown types left out,
  terminal args/env, logout), the initialize / authenticate / logout request writers, and
  auth_required → choose → authenticate → retry, including a refusal and a terminal exit.
• Plugin builds (Debug).
• Live check, macOS, in a --quiet Debug fizzy from main, on its own profile, driven through
  the control socket:
  • Claude's ACP adapter (0.93.0), already logged in: "Connected." with Mode / Model / Effort
    shown. Its initialize advertises auth.logout and two terminal methods.
  • A scripted ACP agent (an agent method and a terminal one; it refuses session/new and
    prompts with -32000 until authenticate):
    • It showed one button, the terminal method left out with no terminal.run.
    • Waiting... → Signed in → Connected.
    • Stop during sign-in → "Stopped signing in", and the next message started over.
    • Sign Out → "Signed out".
    • The next prompt was refused, the person signed in again, and the same prompt was sent again
      and answered.
  • Gemini CLI is not installed here, so it was not tried.

🤖 Generated with Claude Code