Agents & statuses

BeSwarm recognizes the coding agent running in each tab and tracks its status. Agent commands act on the recognized agent currently in a tab.

Supported agents

beswarm agent kinds lists the agents your build recognizes:

Kind	Executable	Kind	Executable
`pi`	`pi`	`copilot`	`copilot`
`claude`	`claude`	`kimi`	`kimi`
`codex`	`codex`	`kiro`	`kiro-cli`
`gemini`	`gemini`	`droid`	`droid`
`cursor`	`cursor-agent`	`amp`	`amp`
`devin`	`devin`	`grok`	`grok`
`agy`	`agy`	`hermes`	`hermes`
`cline`	`cline`	`kilo`	`kilo`
`omp`	`omp`	`qodercli`	`qodercli`
`mastracode`	`mastracode`	`qwen`	`qwen`
`opencode`	`opencode`	`maki`	`maki`
`muse`	`muse`	`mcode`	`mcode`

Any other program still runs in a tab — it just isn't tracked as an agent.

Statuses

Status	Meaning
`working`	Mid-turn.
`done`	Finished, and nobody has acknowledged it yet. Ready for input.
`idle`	Waiting for input, and you've seen it.
`blocked`	A permission prompt or chooser is up. `agent prompt` refuses a blocked agent; answer it with `agent send-keys`.
`unknown`	An agent is present but detection can't classify it. It doesn't prove completion.

How detection works

Each agent has a detection manifest: a set of rules that read the agent's screen and terminal title for spinners, permission prompts and idle prompts. Inspect them with:

beswarm agent manifests                  # the active manifest per agent
beswarm agent explain claude --file screen.txt
beswarm agent explain --target api       # run detection on a live agent

explain prints every rule's verdict, which is the fastest way to see why an agent shows the status it does.

Override a manifest

Put your own manifest at ~/.config/beswarm/agent-detection/<agent>.toml. The menu's "reload config" applies it without a restart.

Names and ids

A terminal id looks like 18d47dc212160d20-1 and is printed by status and agent list.
An agent name matches [a-z][a-z0-9_-]{0,31} and is unique within a session. Set it with agent start --name or agent rename.
Always address agents by name or terminal id — never by kind. claude resolves until a second Claude tab exists, and then a script that worked yesterday prompts the wrong agent.