The problem
A long agent run scrolls past in a terminal. It is hard to tell which task it is on, what each subagent is doing and how much context is left, and in long replies important details get missed.
Next: What I didWhat I did
I built it as a Claude Code plugin with two main parts:
The goal, every task with its own bar, and a card for the main thread and each subagent: what it is doing, how full its context is and how much of the session is its own.
Optional. A small second model reads each reply, marks up to four phrases that matter, and adds a line when something was left out, such as a command that failed.
The plugin makes no network call of its own and writes no file. Marks and the board live in the session and go with it.
/progress and /highlight are the commands. The highlighter costs one extra model call per turn, so it can be turned off in /config.
Where it stands
Published and installable with /plugin install sentinel --marketplace wellavelino/sentinel. MIT licensed.
It depends on Claude Code’s function-hook mods, an early-access API that may change between releases; it was written against 2.1.296.
Next: What I'd do nextWhat I'd do next
Track the mod API. While the API is early access, keeping up with releases comes first.
Get it in front of users. It was built for my own runs; feedback from others would show what is missing.