Match Claude Science compute and artifact workflows

docs/notes/claude-science-ui-behavior-audit.md

@@ -42,7 +42,7 @@ This is a behavior index, not a request to reproduce Claude branding. It records
 
 | Surface              | Claude Science behavior                                                                  | OpenScience contract                                                                                                               |
 | -------------------- | ---------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------- |
-| Default presentation | A step opens into a language/environment header, source, and separate output control.    | Notebook and R cards auto-open so code is never hidden behind Show steps.                                                          |
+| Default presentation | A step opens into a language/environment header, source, and separate output control.    | Notebook, artifact, and remote-compute cards start collapsed so completed work stays compact and expands on demand.                |
 | Source height        | Long source remains contained inside the operation instead of dominating the transcript. | Show exactly five code lines in a vertically and horizontally scrollable source window; retain the complete source in that window. |
 | Output               | Text output is visually separated from source.                                           | Text output is open by default and independently scrollable.                                                                       |
 | Figures              | Figures appear immediately after the cell that produced them.                            | Inline notebook images stay visible even when text output is collapsed.                                                            |
@@ -51,15 +51,15 @@ This is a behavior index, not a request to reproduce Claude branding. It records
 
 ## 5. Compute
 
-| Surface         | Claude Science behavior                                                                                      | OpenScience contract                                                                                                                               |
-| --------------- | ------------------------------------------------------------------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------- |
-| Host strip      | Memory, CPU, live-kernel count, and running count are always visible.                                        | Host totals combine local kernels, shell commands, and remote jobs. Unknown metrics render as unavailable, never fabricated zeroes.                |
-| Project ledger  | Work is grouped by owning session with a current-session marker.                                             | Compute aggregates every session in the project and does not reset when the selected session changes.                                              |
-| Kernel row      | Language, state, age/cell count, activity label, RSS, CPU, and stop action form one dense row.               | Live rows show named kernel, state/recovery text, uptime, RSS, cores, and Stop.                                                                    |
-| Job row         | Long-running/background work stays visible independently of chat scroll.                                     | Shell commands and Modal/GPU jobs are first-class rows with command/target, resources, duration, status, output, artifacts, cleanup, and cancel.   |
-| Completed work  | Claude commonly leaves idle kernels visible.                                                                 | OpenScience intentionally stops finished kernels, then keeps up to five recent local completion rows so the trail remains without wasting compute. |
-| Manual creation | Claude exposes environment setup as part of agent work, not a user kernel launcher in the completed session. | Do not expose manual kernel creation. Kernels start only when an agent executes work.                                                              |
-| Cleanup         | Claude exposes stop/kill per kernel but may leave kernels idle.                                              | The research agent must stop every named kernel after outputs and artifacts are verified. Remote cleanup warnings remain visible.                  |
+| Surface         | Claude Science behavior                                                                                      | OpenScience contract                                                                                                                             |
+| --------------- | ------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------ |
+| Host strip      | Memory, CPU, live-kernel count, and running count are always visible.                                        | Host totals combine local kernels, shell commands, and remote jobs. Unknown metrics render as unavailable, never fabricated zeroes.              |
+| Project ledger  | Work is grouped by owning session with a current-session marker.                                             | Compute aggregates every session in the project and does not reset when the selected session changes.                                            |
+| Kernel row      | Language, state, age/cell count, activity label, RSS, CPU, and stop action form one dense row.               | Live rows show named kernel, state/recovery text, uptime, RSS, cores, and Stop.                                                                  |
+| Job row         | Long-running/background work stays visible independently of chat scroll.                                     | Shell commands and Modal/GPU jobs are first-class rows with command/target, resources, duration, status, output, artifacts, cleanup, and cancel. |
+| Completed work  | Claude commonly leaves idle kernels visible.                                                                 | OpenScience removes completed, stopped, and killed local kernels from Compute; their source, results, and artifacts remain in chat and Files.    |
+| Manual creation | Claude exposes environment setup as part of agent work, not a user kernel launcher in the completed session. | Do not expose manual kernel creation. Kernels start only when an agent executes work.                                                            |
+| Cleanup         | Claude exposes stop/kill per kernel but may leave kernels idle.                                              | The research agent must stop every named kernel after outputs and artifacts are verified. Remote cleanup warnings remain visible.                |
 
 ## 6. Files and artifacts
 
@@ -98,7 +98,7 @@ OpenScience's research prompt now requires at least two decision-useful figures
 
 ## 9. Intentional OpenScience differences
 
-- Finished kernels are stopped automatically rather than left idle. Recent completion rows preserve visibility without retaining memory.
+- Finished kernels are stopped automatically rather than left idle, and disappear from Compute once they are no longer live.
 - Manual kernel creation is removed. The execution ledger describes real work; it is not a launcher.
 - Compute also includes shell subprocesses and Modal/GPU jobs, which the reference surface did not expose in this exact local run.
 - Project-wide Files and Compute remain stable while sessions switch, matching the requested cross-session workspace model.
@@ -108,13 +108,13 @@ OpenScience's research prompt now requires at least two decision-useful figures
 - [x] Exact prompt starts exactly four named managed kernels.
 - [x] Four kernel rows are visible in Compute during execution.
 - [x] Python source and output remain visible in chat while working and after completion.
-- [x] Source is auto-open but capped to a five-line scroll window.
+- [x] Computed cards start collapsed; expanded source is capped to a five-line scroll window.
 - [x] Figures display inline beside their producing cells.
 - [x] Saved report, tables, and figures auto-appear in Files.
 - [x] Artifact titles are meaningful and previews open beside chat.
 - [x] Failed analysis is visible and can be retried without losing history.
 - [x] Every named kernel stops after result verification.
-- [x] Recent completed local work remains visible in Compute.
+- [x] Completed, stopped, and killed local kernels disappear from Compute.
 - [x] Modal/GPU jobs have live and recent-result rows with resources, logs, artifacts, cancel, and cleanup state.
 - [x] The right workspace remains project-scoped across session changes.
 - [x] Compute and artifact cards adapt at narrow container widths.

frontend/ui/src/components/message-part-artifact.test.ts

@@ -6,6 +6,10 @@ const source = () => readFileSync(fileURLToPath(new URL("./message-part.tsx", im
 
 test("saved workspace artifacts render previewable, openable results", () => {
   const part = source()
+  const artifact = part.slice(
+    part.indexOf("function SavedArtifactTool"),
+    part.indexOf('ToolRegistry.register({\n  name: "artifact"'),
+  )
 
   expect(part).toContain('name: "artifact"')
   expect(part).toContain('data-component="saved-artifact-tool"')
@@ -16,12 +20,18 @@ test("saved workspace artifacts render previewable, openable results", () => {
   expect(part).toContain("data.openFile?.(artifact().path)")
   expect(part).toContain("Open beside chat")
   expect(part).toContain("<summary>Show save receipt</summary>")
+  expect(artifact).not.toContain("defaultOpen")
 })
 
-test("Modal and compute job results use a dedicated visible renderer", () => {
+test("Modal and compute job results use a dedicated compact renderer", () => {
   const part = source()
+  const remote = part.slice(
+    part.indexOf("function RemoteComputeTool"),
+    part.indexOf('ToolRegistry.register({ name: "modal"'),
+  )
 
   expect(part).toContain('name: "modal"')
   expect(part).toContain('name: "compute_job"')
   expect(part).toContain('title: props.tool === "modal" ? "Modal compute" : "Remote compute result"')
+  expect(remote).not.toContain("defaultOpen")
 })

frontend/ui/src/components/message-part-notebook.test.ts

@@ -5,8 +5,12 @@ import { fileURLToPath } from "node:url"
 const source = () => readFileSync(fileURLToPath(new URL("./message-part.tsx", import.meta.url)), "utf8")
 const styles = () => readFileSync(fileURLToPath(new URL("./message-part.css", import.meta.url)), "utf8")
 
-test("notebook tools open source, text output, and figures by default", () => {
+test("notebook tools keep complete source, output, and figures behind a compact summary", () => {
   const part = source()
+  const kernel = part.slice(
+    part.indexOf("function KernelTool"),
+    part.indexOf('ToolRegistry.register({\n  name: "notebook"'),
+  )
 
   expect(part).toContain('name: "notebook"')
   expect(part).toContain('name: "rkernel"')
@@ -20,6 +24,7 @@ test("notebook tools open source, text output, and figures by default", () => {
   expect(part).toContain('props.input.action === "stop"')
   expect(part).toContain('trigger={{ title: "Kernel stopped"')
   expect(part).toContain('title: props.status === "completed" ? "Computed" : "Computing"')
+  expect(kernel).not.toContain("defaultOpen")
   expect(styles()).toContain("max-height: calc(5 * 1.55em + 20px)")
   expect(styles()).toContain("overflow: auto")
 })

frontend/ui/src/components/message-part.tsx

@@ -830,7 +830,6 @@ function KernelTool(props: ToolProps & { language: "python" | "r"; label: "Pytho
   return (
     <BasicTool
       {...props}
-      defaultOpen
       icon="code"
       trigger={{
         title: props.status === "completed" ? "Computed" : "Computing",
@@ -906,7 +905,6 @@ function SavedArtifactTool(props: ToolProps) {
   return (
     <BasicTool
       {...props}
-      defaultOpen
       icon="archive"
       trigger={{
         title: saved() ? "Saved artifact" : props.title || "Artifact",
@@ -985,7 +983,6 @@ function RemoteComputeTool(props: ToolProps) {
   return (
     <BasicTool
       {...props}
-      defaultOpen
       icon="console"
       trigger={{
         title: props.tool === "modal" ? "Modal compute" : "Remote compute result",

frontend/workspace/src/atlas/ComputeSurface.css

@@ -312,31 +312,6 @@
   border-top: 8px solid var(--color-bg-subtle);
 }
 
-.compute-surface .kernel-history-card {
-  grid-template-columns: minmax(160px, 1fr) 64px minmax(112px, auto);
-}
-
-.compute-surface .kernel-history-card__result {
-  display: flex;
-  min-width: 0;
-  flex-direction: column;
-  gap: 2px;
-}
-
-.compute-surface .kernel-history-card__result strong {
-  color: var(--color-text);
-  font-size: 11px;
-  font-weight: 500;
-}
-
-.compute-surface .kernel-history-card__result small {
-  overflow: hidden;
-  color: var(--color-text-faint);
-  font-size: 10px;
-  text-overflow: ellipsis;
-  white-space: nowrap;
-}
-
 .compute-surface .remote-job-card {
   grid-template-columns: minmax(160px, 1fr) 64px minmax(118px, auto) auto;
 }
@@ -441,14 +416,6 @@
     grid-template-columns: minmax(0, 1fr) auto;
   }
 
-  .compute-surface .kernel-history-card {
-    grid-template-columns: minmax(0, 1fr) auto;
-  }
-
-  .compute-surface .kernel-history-card__result {
-    padding-left: 39px;
-  }
-
   .compute-surface .remote-job-card__result {
     padding-left: 39px;
   }
@@ -466,10 +433,6 @@
   .compute-surface .remote-job-card__result {
     grid-column: 1 / -1;
   }
-
-  .compute-surface .kernel-history-card__result {
-    grid-column: 1 / -1;
-  }
 }
 
 @media (prefers-reduced-motion: reduce) {

frontend/workspace/src/atlas/KernelCard.tsx

@@ -73,43 +73,6 @@ export function KernelCard(props: {
   )
 }
 
-const ago = (value: number | null) => {
-  if (!value) return "—"
-  const seconds = Math.max(0, Math.floor((Date.now() - value) / 1_000))
-  if (seconds < 60) return `${seconds}s ago`
-  const minutes = Math.floor(seconds / 60)
-  if (minutes < 60) return `${minutes}m ago`
-  const hours = Math.floor(minutes / 60)
-  if (hours < 24) return `${hours}h ago`
-  return `${Math.floor(hours / 24)}d ago`
-}
-
-export function KernelResultCard(props: { kernel: KernelStatus }): JSX.Element {
-  return (
-    <article class="kernel-card kernel-history-card" data-kernel-id={props.kernel.id} data-state={props.kernel.state}>
-      <div class="kernel-card__main">
-        <span class="kernel-card__language" aria-hidden="true">
-          {props.kernel.language === "python" ? "Py" : props.kernel.language === "r" ? "R" : "›_"}
-        </span>
-        <div class="kernel-card__copy">
-          <strong title={kernelLabel(props.kernel)}>{kernelLabel(props.kernel)}</strong>
-          <span>
-            <i data-tone={props.kernel.state === "crashed" ? "danger" : "muted"} aria-hidden="true" />
-            {kernelStateLabel(props.kernel.state)} · output preserved in chat and Files
-          </span>
-        </div>
-      </div>
-      <span class="kernel-card__uptime" aria-label={`Finished ${ago(props.kernel.last_activity_at)}`}>
-        {ago(props.kernel.last_activity_at)}
-      </span>
-      <span class="kernel-history-card__result">
-        <strong>{props.kernel.state === "crashed" ? "Needs review" : "Complete"}</strong>
-        <small>workspace cleared</small>
-      </span>
-    </article>
-  )
-}
-
 function Metric(props: { label: string; value: string }): JSX.Element {
   return (
     <span class="kernel-card__metric">

frontend/workspace/src/atlas/KernelPanel.test.ts

@@ -21,9 +21,11 @@ describe("live compute inventory", () => {
     expect(panel).toContain("<CommandCard")
     expect(panel).toContain("jobApi.list()")
     expect(panel).toContain("<RemoteJobCard")
-    expect(panel).toContain("<KernelResultCard")
-    expect(panel).toContain('aria-label="Recent local results"')
+    expect(panel).not.toContain("KernelResultCard")
+    expect(panel).not.toContain("recentLocal")
+    expect(panel).not.toContain('aria-label="Recent local results"')
     expect(panel).toContain('aria-label="Recent remote results"')
+    expect(panel).toContain('kernels.filter((kernel) => kernel.active || kernel.state === "starting")')
   })
 
   test("does not expose any manual kernel creation or restart path", () => {

frontend/workspace/src/atlas/KernelPanel.tsx

@@ -3,7 +3,7 @@ import { createStore } from "solid-js/store"
 import { useParams } from "@solidjs/router"
 import { IconCpu } from "@/atlas/shared/Icon"
 import { identify } from "@/atlas/poll-identity"
-import { KernelCard, KernelResultCard, type KernelAction } from "@/atlas/KernelCard"
+import { KernelCard, type KernelAction } from "@/atlas/KernelCard"
 import { CommandCard } from "@/atlas/CommandCard"
 import type { Job, Status } from "@/atlas/ComputeJobsAPI"
 import { RemoteJobCard, jobLive } from "@/atlas/RemoteJobCard"
@@ -120,12 +120,6 @@ export function KernelPanel(props: KernelPanelProps = {}): JSX.Element {
       .filter((job) => job.target.kind === "modal" && terminal.has(job.status) && !jobLive(job))
       .slice(0, 5),
   )
-  const recentLocal = createMemo(() =>
-    [...kernels]
-      .filter((kernel) => !kernel.active && kernel.state !== "lazy" && kernel.last_activity_at !== null)
-      .sort((a, b) => (b.last_activity_at ?? 0) - (a.last_activity_at ?? 0))
-      .slice(0, 5),
-  )
   const groups = createMemo(() =>
     [...grouped().keys()].sort((a, b) => {
       const current = Number(route() === b) - Number(route() === a)
@@ -221,7 +215,7 @@ export function KernelPanel(props: KernelPanelProps = {}): JSX.Element {
         </Show>
 
         <Show
-          when={groups().length > 0 || recentLocal().length > 0 || recentRemote().length > 0}
+          when={groups().length > 0 || recentRemote().length > 0}
           fallback={
             <div class="kernel-panel__empty">
               <span aria-hidden="true">
@@ -283,20 +277,6 @@ export function KernelPanel(props: KernelPanelProps = {}): JSX.Element {
                 </section>
               )}
             </For>
-            <Show when={recentLocal().length > 0}>
-              <section class="remote-results" aria-label="Recent local results">
-                <header class="kernel-session__header">
-                  <div class="kernel-session__identity">
-                    <span aria-hidden="true">↳</span>
-                    <strong>Recent local results</strong>
-                  </div>
-                  <span>{recentLocal().length} completed</span>
-                </header>
-                <div class="kernel-panel__list">
-                  <For each={recentLocal()}>{(kernel) => <KernelResultCard kernel={kernel} />}</For>
-                </div>
-              </section>
-            </Show>
             <Show when={recentRemote().length > 0}>
               <section class="remote-results" aria-label="Recent remote results">
                 <header class="kernel-session__header">