A `review` request and an `answer` request are deliberately never units — neither produces an artifact, so counting them would move a goal's pie every time somebody asked a question — and every cross-goal list was a fold over units. So the most ordinary shape of "an agent is working right now" (the review request auto-review writes under the daemon's own reviewer identity, and any open one an agent has claimed) was the one shape "With an agent" could not draw, and an open answer request appeared nowhere but its thread. One row type (`AskContext`), two folds (`reviewAsks`, `answerAsks`), one union module (`asks.ts`), and `AnswerRow` generalised into `AskRow`. An ask is drawn in exactly one of three lists — with an agent, awaiting input, or owed by a person — decided by one predicate pair (`askWithAgent` / `askWithPerson`), so no row and no rail badge is counted twice. A review ask links to the drawer at the exact version it pins, project-scoped ones included; an answer ask links to the thread. Per-goal badges count asks too, and the state ladder for an open request is now one exported function in core (`openRequestState`) rather than a closure inside `timeline()`. Units are untouched: no timeline, tally or pie moved. Co-authored-by: claudebot.disnetdev.com (did:plc:n6ku5xddiuguwze3f356evla) <claudebot.disnetdev.com@noreply.radial>