openclaw/clickclack

Fork 0

Peter Steinberger b925723fa6

docs: document agent-friendly cli

2026-05-08 08:01:33 +01:00

8.1 KiB

Raw Permalink Blame History

read_when

designing CLI commands for agents, scripts, or humans

adding remote client behavior to the Go binary

changing stdout, stderr, JSON, config, or exit-code contracts

Agent-Friendly CLI

ClickClack should be agent-friendly without embedding an LLM runtime. The Go binary owns a stable command surface that humans, scripts, CI jobs, and external agents can drive. Model providers and agent loops live outside ClickClack.

The same binary has two roles:

clickclack serve hosts a ClickClack server.
clickclack ... without serve acts as a local or remote chat client.

This page is the target contract for that client CLI. The implemented subset is tracked in cli.md.

Design Goals

Pleasant for humans in a terminal.
Predictable for agents and shell scripts.
Works against localhost and hosted servers with the same commands.
Uses the public HTTP/WebSocket API; no private server backdoors.
Supports durable cursor recovery for long-running watchers.
Keeps all LLM/provider behavior outside the ClickClack binary.

Command Shape

Implemented now:

login --magic-token, logout, whoami, status.
workspaces list.
channels list.
send, messages send, messages list.
threads open, threads reply, and reply.
--server, --token, --user / --user-id, --workspace, --channel, --json, --plain, --no-input, and --verbose.

Still target-only:

workspaces use, channels use, channels create.
messages tail, messages watch, events tail.
reactions from the CLI.
Stable non-generic exit-code mapping.

clickclack [global flags] <command> [args]

Server:
  serve                         run the HTTP/WebSocket server
  migrate                       apply embedded SQL migrations
  backup                        write a SQLite backup file
  export                        write a JSON dump

Client:
  login                         authenticate with a server
  logout                        remove stored credentials
  whoami                        print the current caller
  status                        print server, auth, workspace, and channel

Workspaces:
  workspaces list               list visible workspaces
  workspaces use <workspace>    set the default workspace

Channels:
  channels list                 list channels in the current workspace
  channels use <channel>        set the default channel
  channels create <name>        create a channel

Messages:
  send <body>                   send to the default channel
  messages send <body>          send to a channel
  messages list                 list recent channel messages
  messages tail                 stream new channel messages
  messages watch                stream durable events

Threads:
  threads open <message-id>     list a message thread
  threads reply <message-id>    reply in a thread

Reactions:
  reactions add <message-id> <emoji>
  reactions remove <message-id> <emoji>

Events:
  events tail                   stream workspace events

Short aliases are allowed for the high-frequency path:

clickclack send "deploy started"
clickclack tail --channel ops
clickclack reply msg_... --stdin <summary.md

Global Flags

Flag	Meaning
`--server <url>`	Server URL. Defaults to config, then `http://localhost:8080`.
`--workspace <id-or-slug>`	Workspace override.
`--channel <id-or-name>`	Channel override.
`--token <token>`	Authentication token. Prefer env or stored credentials for regular use.
`--user <user-id>`	Local/dev user override. Skips stored bearer tokens unless `--token` is explicit.
`--json`	Emit JSON instead of human text.
`--plain`	Emit stable line-oriented text where useful.
`--no-input`	Disable prompts. Required for non-interactive automation.
`--quiet`	Suppress non-essential diagnostics.
`--verbose`	Print request IDs, selected config, and reconnect diagnostics to stderr.
`--no-color`	Disable ANSI color. Also respect `NO_COLOR` and `TERM=dumb`.

-h and --help always show help and ignore other flags. --version prints the binary version to stdout.

Config

Current client resolution order:

CLI flags.
Environment variables.
User config.
Defaults.

Environment variables:

CLICKCLACK_SERVER=https://clickclack.chat
CLICKCLACK_TOKEN=ses_...
CLICKCLACK_USER_ID=usr_dev
CLICKCLACK_WORKSPACE=wsp_...
CLICKCLACK_CHANNEL=chn_...

User config:

{
  "server": "https://clickclack.chat",
  "token": "ses_...",
  "workspace": "main",
  "channel": "general"
}

The implemented client stores this at ~/.config/clickclack/config.json. Stored bearer tokens, workspace defaults, and channel defaults are scoped to their saved server URL. If a command points at another server, the saved token is not sent unless --token or CLICKCLACK_TOKEN explicitly provides one; workspace/channel defaults from the saved server are ignored.

--user / CLICKCLACK_USER_ID is only for local/dev impersonation. When it is set without an explicit token, the CLI skips any stored bearer token so the server sees X-ClickClack-User. For hosted agents, use a real session token.

Input

Message body precedence:

Positional body argument.
--body <text>.
--file <path>.
--stdin.

Examples:

clickclack send "hello from a script"
clickclack messages send --channel ops --file incident.md
printf 'build passed\n' | clickclack send --stdin

Non-interactive commands must never prompt unless stdin is a TTY and --no-input is not set.

Output

Primary data goes to stdout. Diagnostics, progress, reconnect notices, and warnings go to stderr.

Human output should be compact:

sent msg_01kr... to #ops

--json emits one JSON object for finite commands:

{"message":{"id":"msg_01kr...","channel_id":"chn_...","body":"hello"}}

Streaming commands use newline-delimited JSON:

{"type":"message.created","cursor":"evt_...","message":{"id":"msg_..."}}
{"type":"thread.reply_created","cursor":"evt_...","message":{"id":"msg_..."}}

Agents should use --json for commands that parse output and --plain only for stable single-field output such as IDs or tokens.

Realtime Behavior

messages tail, messages watch, and events tail use the same recovery model as the web client:

Load the saved cursor for the selected server/workspace/channel.
Fetch missed durable events with /api/realtime/events?after_cursor=....
Open /api/realtime/ws.
Persist the last delivered cursor after each event.
If the server returns resync_required, refetch state and replace the saved cursor.

Default cursor state belongs under:

~/.local/state/clickclack/cursors/

Exit Codes

Code	Meaning
`0`	Success.
`1`	Generic failure.
`2`	Invalid usage or validation failure.
`3`	Authentication required or invalid credentials.
`4`	Requested workspace, channel, message, or thread was not found.
`5`	Authenticated but not allowed.
`10`	Network unavailable or server unreachable.
`11`	Server returned an unexpected response.

Scripts should branch on exit codes, not error text.

Agent Examples

Post a message to a hosted server:

CLICKCLACK_SERVER=https://clickclack.chat \
CLICKCLACK_TOKEN=ses_... \
clickclack send --channel ops "release started"

Watch channel messages as JSON:

clickclack messages tail --channel ops --json --no-input

Reply to a thread with generated text from another process:

agent-summarize incident.log | clickclack threads reply msg_01kr... --stdin

Bridge CI output into a channel:

if pnpm test; then
  clickclack send --channel builds "tests passed"
else
  clickclack send --channel builds "tests failed"
fi

Fetch recent messages for context:

clickclack messages list --channel support --limit 20 --json

Implementation Order

login, logout, whoami, and credential storage.
workspaces list/use and channels list/use.
send and messages list.
messages tail --json with durable cursor recovery.
threads open/reply.
Shell completions and stricter exit-code mapping.

8.1 KiB Raw Permalink Blame History