agenter-attention

name: agenter-attention description: Inspect and settle attention work. Use this when you need to understand or update attention debt.

Use this skill when you need to inspect unresolved attention work or record that the work has truly progressed.

Quick start:

If the current task already names the concrete file, room, terminal, or URL work to do next, do that real work before browsing attention.
Run attention list only when you need the active contextId or need to confirm which attention work still remains.
Run attention query with JSON stdin only if attention list is still not enough.
When the real work is finished through root_bash, keep command=attention commit and put the JSON object in the tool-level stdin field.

Key laws:

score > 0 means the attention work still exists.
done: true is the normal way to resolve the active scores for a context.
Lower scores only after the real work has actually happened.
For attention query and attention commit through root_bash, default to the minimal command plus JSON in the tool-level stdin; only use argv JSON when you are already inside a plain shell and the payload is trivially short.
Do not synthesize stdin with shell glue such as echo '{...}' | attention commit when you control root_bash; pass the JSON via the tool's stdin field instead.
If you do choose the plain-shell fallback, pass the JSON directly as one argv object like attention commit '{...}'; do not add a redundant echo.
If attention query --help or attention commit --help marks compact as Suggested or Available, --compact is an optional positional mode. If the positional array becomes unclear, switch back to standard object JSON immediately.
For a simple single-room delivery task, do not interrupt the first action just to inspect attention if the current room already states the work clearly.
If you already sent the durable result, verified the file, or confirmed the external side effect, it is usually reasonable to settle the attention in the same round instead of leaving solved work active.
If the work is still waiting on another participant, relay room, or external source, the context is not complete yet and should stay open.
For relay work, done: true is usually appropriate only after the origin room already received the final answer, not merely a progress acknowledgement that you are asking someone else.

Common endings:

Delivery finished and the user already got the result:

root_bash.command: attention commit
root_bash.stdin: {"contextId":"ctx-...","summary":"Delivered the required reply and verified the result.","done":true}

Plain-shell fallback only when you are already inside a shell and the payload is trivially short: attention commit '{"contextId":"ctx-...","summary":"Delivered the required reply and verified the result.","done":true}'
You need to keep the work open because evidence is still missing: leave the context unresolved, gather evidence, and only commit once the real state changed.

References:

references/settlement.md: how to decide whether attention is still active and how to use commit