A routine can be configured before it is run (fix #1507) - #1634
Conversation
Run now spends an agent on settings that are nowhere on the Routine work card: the model and where it runs come from the Global options, so the only way to change them was to leave, edit the preferences, and come back. Each Run now becomes a split button. Its chevron opens one item, "Configure first, then run", which hands the routine's prompt to the picked project's launcher — where the model, the location, and the prompt itself can all be changed before anything is spent. The carry is the draft stash a hot ticket already uses (#1066/#1139) rather than a second mechanism: the launcher takes it once as it mounts, so a reload does not re-seed it. A split button rather than the hover-revealed control the issue floats: a control that only exists under a mouse is reachable by neither keyboard nor touch, and the chevron keeps the common click at one click. The drain row's two halves really do different jobs — its Run now fires a fan-out sweep and the launcher can only ever send one agent — so its menu item says so instead of looking like the same action.
|
🤖 automated Merging this — but one decision in it is yours, and it is cheap to reverse.
The issue floated "hovering a prompt button shows a secondary action … maybe as a dropdown?". I built the dropdown half and made its trigger always visible: a control that exists only under a mouse is reachable by neither keyboard nor touch. That is the one thing worth overruling if you disagree — it is a chevron, so switching it is a small follow-up, not a revert.
Second, smaller one: "Spin up agents working on the AI queue" is the only row whose Next is #1506 — tooltips on these same buttons, where "what model / where it runs" gets said on the card itself instead of only after you arrive at the launcher. |
Run now spends an agent on settings that are nowhere on the Routine work card: the model and where it runs come from the Global options, so the only way to change them was to leave, edit the preferences, and come back. Each Run now becomes a split button. Its chevron opens one item, "Configure first, then run", which hands the routine's prompt to the picked project's launcher — where the model, the location, and the prompt itself can all be changed before anything is spent. The carry is the draft stash a hot ticket already uses (#1066/#1139) rather than a second mechanism: the launcher takes it once as it mounts, so a reload does not re-seed it. A split button rather than the hover-revealed control the issue floats: a control that only exists under a mouse is reachable by neither keyboard nor touch, and the chevron keeps the common click at one click. The drain row's two halves really do different jobs — its Run now fires a fan-out sweep and the launcher can only ever send one agent — so its menu item says so instead of looking like the same action.
1 participant
Closes #1507.
The problem in one line:
Run nowon the Routine work card spends an agent on settings you cannot see from that card. The model and where it runs come from the Global options, so changing them meant leaving, editing preferences, and coming back.What it looks like
Each
Run nowis now a split button. The chevron opens one item:Clicking it hands the routine's prompt to the picked project's launcher, where the model, the location, and the prompt itself can all be changed before anything is spent.
Two decisions worth your eye
A split button, not a hover-revealed control. The issue floats "hovering a prompt button shows a secondary action ... maybe as a dropdown?" — I went with the dropdown half and made its trigger always visible. A control that exists only under a mouse is reachable by neither keyboard nor touch, and the chevron keeps the common click at one click. Say the word if you'd rather it were hover-gated.
The drain row says what it costs. "Spin up agents working on the AI queue" is the one row whose
Run nowfires a fan-out sweep, and the launcher can only ever send one agent — so its two halves really do different jobs. Its menu item reads "Opens the launcher with this prompt — one agent, not the fan-out." rather than looking like the same action.How it's built
No new mechanism: the prompt travels on the draft stash a hot ticket already uses (#1066/#1139), which the launcher takes once as it mounts. Verified in the browser that a reload leaves the composer empty rather than re-seeding.
RoutineWorkgains one prop,onSelectProject— the same oneDashboardPagealready passes to the Human Queue and the hot tickets.Verification
Dogfooded on the real daemon, not only jsdom — the lesson from #1500 was that a component can render correctly and still be mounted where nobody sees it:
gemstacklauncher with the routine's whole prompt in the box, model picker showing Opus, no agent startedTests: 5 new, each proven by breaking the code it guards. Dropping the stash, pointing at the first project instead of the picked one, flattening the two descriptions to one, inverting the drain check, and gating the menu item on
busyeach failed exactly the intended test and nothing else. Full suite green: 1536 node + 788 dashboard.RoutineWork.SPEC.md, its test SPEC, andFEATURES-SPEC.mdall updated.Next up is #1506 (tooltips on these same buttons), which is where "what model / where it runs" gets said on the card itself rather than only after you arrive at the launcher.