agent-watch

A Claude Code mod. While subagents run, the band above the prompt shows one line per agent:

Sonnet · Build snap screenshot mod · 6 min · now: Bash claude plugin test

• Label: the agent's model (Sonnet, Opus, Haiku, Fable), else its subagent type.
• Minutes update every 15 s.
• now: is the agent's latest tool call (file name, first 40 characters of a Bash command, or pattern), cut to the band width.
• When an agent ends, its line shows done (green) for 30 s, or stopped (yellow) if it failed or was stopped, with one toast Agent finished: <description>.
• With no agents the mod draws nothing. Other mods' rows stay below ours.

Engine data used

• Start: agent.spawn. next(e) answers { model, agentId }; the input holds description and subagentType.
• Progress: tool.call with agentId (set only inside a subagent loop). Hooks always pass through.
• Finish: turn.complete with agentId (reason answer is done, anything else stopped), and every 15 s $.agent.list() status (completed, failed, killed) as a fallback.
• State: agent-watch.agents in $.state, so a hot reload keeps the rows.

Limits

• The engine does not tell a background agent's end apart from a foreground one; both end through turn.complete or $.agent.list().
• Teammates (type: teammate) never pass agent.spawn, so they are not shown.

Develop

claude plugin validate .
claude plugin test .
npx tsc -p .