# Board attention status correction

Jason accepted discovery/reply tests 1-4, then reported Researcher still showing
waiting after its completed reply. He approved separating idle from explicit
requests for human input. This is a #1503 acceptance defect, not new chat scope.

Darkwing authors; internal Filbert independently reviews. Baseline observed on
refactor: d9745a451004699256af32c2145df40df9f5b4a5. Preserve existing shared edits
and Dewey's UI work. Source scope: control-board scanner, tests and README,
AGENTS.md status guidance, the downstream packages/webui/tests/fixture.mjs waiting
case and scoped tracking records. Dewey confirmed no overlapping ownership. No home-directory or
fleet changes, agent restarts, new public JSON fields, approval/control authority,
text classification service or unrelated UI changes.

## Rule and evidence

Completed normal assistant replies are idle. A completed assistant reply requests
human attention only when its first nonblank text line is exactly prefixed
`Input needed: ` followed by a nonempty request. This readable status convention
is explicit, not a guess based on questions, sentiment, or turn completion.
Quoted examples, user/tool messages and thinking blocks do not signal attention.
Tool activity and errors retain their existing precedence. The latest message
wins; a later ordinary completed reply clears the previous request.

The signal is advisory display state, never permission to perform an action.
Agents use it only for Jason's decision/input, not a wait on another agent or a
routine completion. Existing conversations without the signal become idle after
normal completion. Earlier unstated human blockers cannot be inferred safely.

Seen acknowledges a specific waiting/error event. It hides that event from the
attention list but does not resolve a genuine input request. New explicit requests
reappear; ordinary completions do not. No change to existing API enum or fields.

Acceptance: reproduce BOARD_REPLY_OK incorrectly becoming waiting, test completed
replies/explicit requests/quoted and stale markers/tool activity/errors/Seen,
run scanner/server and relevant downstream regressions, and obtain exact-candidate
independent review. Verify the corrected classification against Researcher's
actual session without editing it. Running-board replacement and operator retest
must be reported separately from source tests; no unattended service restart.

Before implementation: new attention tests reproduce two failures, including
BOARD_REPLY_OK being classified as waiting. One precedence test already passes.