Autoloop Acp

Autoloop over ACP

autoloop acp presents an Agent Client Protocol (ACP) interface over stdio. The external system (an editor, harness, or another agent) is the ACP client; autoloop is the ACP agent. This skill is for setting up that connection and driving autoloop through it. Parse $ARGUMENTS to determine what the user needs; if ambiguous, ask.

Direction matters. This skill is about other systems driving autoloop (autoloop is the agent). The inverse — autoloop consuming an external ACP agent as a backend — is the separate -b kiro option and is not what this covers.

Launching the agent

# Start the ACP agent on stdio (the client spawns this as a subprocess)
autoloop acp

# Surface debug logs as agent thoughts
autoloop acp --verbose      # or -v

# Usage
autoloop acp --help         # or -h

Transport rules (important):

• Frames are NDJSON over stdio. Stdout is the protocol channel — nothing else may write to it. All logs go to stderr.
• The client therefore launches autoloop acp as a subprocess and speaks ACP on stdin/stdout. Don't pipe other output into stdout.

Working directory (REQUIRED): per session. The ACP client must set it via the session/new cwd field, as the absolute path of the target project directory. autoloop anchors all run state (journal, registry, runs/) on cwd. If cwd is missing or blank, session/new is rejected with a JSON-RPC invalid params error — autoloop does not silently fall back to its own launch directory, because doing so would misroute the entire run into the wrong project. --project-dir / AUTOLOOP_PROJECT_DIR only affect dashboard control; they are not a fallback for the session working directory.

Representative client config

ACP clients differ, but they spawn the agent as a command + args. A generic entry looks like:

{
  "command": "autoloop",
  "args": ["acp"]
}

Add "--verbose" to args if needed. The client must send a valid cwd on session/new (see Working directory above). No authentication step is required — authenticate is a no-op because autoloop runs locally under the caller's own credentials.

What the client sees

On initialize, autoloop reports:

• agentInfo: { name: "autoloop", version: "0.1.0" }
• agentCapabilities: { loadSession: false } — sessions are not resumable; one session lives for the life of the connection and can launch many runs and quick commands over successive prompt turns.

On session/new, autoloop requires a valid cwd (rejecting the session otherwise) and advertises its verbs as slash commands (available_commands_update), each with a description and an input hint.

Prompt behaviour

Each prompt turn's text is dispatched as an autoloop command line:

• Bare objective (no recognised verb) → runs the default preset: run autocode "<your prompt>". So sending build the login page starts an autocode loop on that objective.
• Explicit verb or /slash command always overrides the default (e.g. run autofix "...", /loops, inspect scratchpad).
• A single outer XML-ish wrapper is stripped before parsing — <user_message>, <user>, <message>, <query>, or <prompt>. This makes model-driven clients that wrap prompts work transparently. Unrecognised tags are left untouched.
• Tokenising honours single and double quotes, so quoted objectives with spaces survive: run autocode "Fix the login bug".
• An empty prompt returns help text.

Command surface

All 14 verbs work both as bare prompt lines and as /slash commands. They run in one of three modes:

Stream mode — long-running loops, streamed as ACP tool calls

These execute a loop and bridge its events onto ACP session/update notifications as tool calls. The turn ends with a summary message.

Capture mode — quick commands, returns captured text

These run synchronously with stdout/stderr captured and return the text as a single agent message.

Control mode — dashboard lifecycle

Handled directly by the agent. On start, the dashboard URL is returned on its own line so clients can linkify it.

Streaming & cancellation

• Stream verbs (run, chain) emit live session/update notifications as the loop runs, then a final summary with a stop reason.
• Capture verbs return their full captured output in one message.
• Cancel an in-flight turn with the ACP session/cancel notification — it aborts the active turn (and the running loop) for that session.

Quick reference