refactor(explorer): remove the task runner's embedded-host mode, which never runs outside Scale's hub

[polakamtejas]

The explorer UI was forked from the hosted AgentEnvHub frontend. Its task runner was built to run inside an iframe in a host pipeline, and the parent page owned the task definition. So the runner:

• created the task from the parent's config (POST /api/v1/tasks)
• re-created it before Run or grading when the parent's config had drifted
• rotated it to a new id to "start over"
• graded attempts the parent handed down
• reported runs, starred ("preferred") runs and materialized fields back to the parent

None of that runs here. The host hook, lib/external-app, is a stub that always reports "not embedded". The step editor that saved through POST /api/v1/tasks was already unreachable (its own comment said so). And the explorer backend has no POST /api/v1/tasks route.

Removed

• lib/external-app/ and both uses (task-runner-page.tsx, universes-page.tsx).
• Every window.parent !== window check:
  • the runner's embed branches
  • the Universes Hub's two-step picker and "Use this universe" bar
  • the instance viewer's hidden Task Run Context tab (and its showRunContext prop)
  • the app shell's hidden sidebar
• From the runner:
  • createTask and both POST /api/v1/tasks calls
  • the task_config startup branch, applyLatestConfig, the drift hash and banner
  • "Start over", gradeAllAttempts, the preferred-run stars, the model picker, the run cap, runs-per-click and auto_run
  • sendResult/sendResults and the trajectory/step warming that fed them
  • the unreachable step editor

Why behavior doesn't change

• Every removed branch was gated on something only the host could set: receivedInputs, receivedOutput, window.parent !== window, or editing, which was never set true. Each was evaluated as the standalone value and simplified.
• The runner still loads the task from /task-runner/<id>, runs it, cancels a pending or running run, polls, and shows each instance. It still initialises once per mount, as before.
• handleRun loses its count fan-out, which only runs_per_click used. A single run launches, fails and reports exactly as before.

Lines removed: task-runner-page.tsx −1779/+77 (2403 → 701 lines), universes-page.tsx −125/+2, task-instance-viewer.tsx −25/+5, lib/external-app/index.ts −61, [[...slug]].tsx −10 (+89/−99 as shown, because the sidebar block dedents one level).

Testing

• New ui/src/lib/task-runner-run.smoke.ts (31 checks) covers the run lifecycle, which the second commit moves unchanged from the page component into lib/task-runner-run.ts so the node-only smoke runner can reach it:

  • the run request
  • a launch that starts and returns its workflow id
  • a refused launch surfacing the backend's detail instead of leaving the page on "Starting…"
  • the started run being picked up and selected once it appears in the instance poll
  • polling continuing until it does, and stopping once every run is finished
  • cancel with both ids encoded

  Breaking the "keep polling while a started run hasn't appeared" rule, or dropping the backend detail, each fails it.

• npx tsc --noEmit, npm run test:smoke (12 passed; the 2 skips need trajectory fixtures) and npm run build:static pass. No useExternalApp, isEmbedded, window.parent, sendResult, parentTaskConfig or api/v1/tasks POST left in src/.

• Browser, against agent-env up:

  • Opened /task-runner/hello-flat, clicked Run: a new instance appeared, completed at 100%, and was auto-selected.
  • A completed instance renders the same as on main (Verifier Results default, Task Run Context tab present).
  • The task detail page and its instance view match main.
  • The Universes Hub lists universes, and a card click navigates to the detail page.

Not changed here, noted for follow-ups:

• The runner's evaluator-return flow (handleEvaluatorSaved) and its onEditEvaluator / onRerunFromStep props were already unreachable before this PR.
• ?projectId= is still forwarded on Run when present in the page URL.

🤖 Generated with Claude Code

Confidence Score: 5/5

This PR appears safe to merge; no outstanding issue was identified in its changes.

What we checked:

• A refused run leaves “Starting…” on screen: A failed launch clears the pending state and shows the server’s reason. The smoke test checks that the reason is preserved.

Summary

The Explorer drops iframe-host task flows that depended on parent-page configuration, while keeping the standalone task-run lifecycle. It also removes iframe-only page controls and moves run-request and polling logic into a helper with smoke coverage.

• The runner loads a task by URL ID and no longer creates or reports tasks through a host page.
• A run still starts, gets polled, and appears in the instance viewer; the lifecycle logic now has a smoke test.
• Universe navigation, the instance context tab, and the app sidebar use the standalone Explorer flow.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Open task runner] --> B[Load task]
  B --> C[Start run]
  C --> D[Poll instances]
  D --> E[Select new instance]
  C --> F[Show launch error]
  D --> G[Cancel run]

Loading

_{Reviews (4) · Last reviewed commit: "test(explorer): cover the task runner's ..."}