feat: Display eval prompt in dashboard task detail (#526)

[Shayne Boyer (spboyer)]

Closes #526

Summary

Adds the eval prompt used to run each task into the dashboard, alongside graders and outcome details.

⚠️ Draft — developed in parallel with a GPT-5.5 approach. Please review both before merging.

Approach (Claude Opus 4.7)

End-to-end: persist the prompt in results JSON, expose it through the webapi, render an inline panel in the expanded task row on the dashboard.

Backend

• internal/models/outcome.go — TestOutcome now has optional Prompt, PromptFile, FollowUps fields (all omitempty, backward compatible with legacy results.json files).
• internal/orchestration/runner.go — populates the new fields from TestCase.Stimulus at the three TestOutcome construction sites: the main runTestUncached return path, and both the sequential and concurrent before_task hook-failure branches.
• internal/webapi/types.go + internal/webapi/store.go — TaskResult gains prompt / promptFile / followUps fields, and outcomeToDetail forwards them.
• internal/webapi/additional_test.go — new TestOutcomeToDetailMapsPrompt asserts prompt round-trips through the API and that legacy tasks without a prompt still work.

Frontend

• web/src/api/client.ts — TaskResult interface extended with the optional prompt fields.
• web/src/components/RunDetail.tsx — new PromptPanel component rendered above graders in the expanded task row:
  • Monospace <pre> with wrapped/scrollable content, dark theme.
  • Copy to clipboard button (uses navigator.clipboard, flips to a Check icon for ~1.5s).
  • Shows the prompt_file source path when the prompt was loaded from a file.
  • Collapsible Follow-up prompts section listing each follow-up.
  • Graceful "Prompt not recorded" hint for older results.

Docs

• README.md — added the prompt panel to the dashboard views bullet list.
• site/src/content/docs/guides/dashboard.mdx — expanded the "Run Details" list with a description of the prompt panel.

Verification

• go fmt ./... && go test ./... — all packages pass.
• golangci-lint run — clean on our code (two flatted/golang/pkg govet warnings live in web/node_modules/, unrelated to this change).
• cd web && npm run build — TypeScript compiles, Vite build succeeds.
• cd web && npm run lint — clean.

Design choices

• Inline panel, not a new tab. Keeps the prompt colocated with its graders and outcome; matches the existing expand-a-row pattern.
• Plain-text <pre>, no syntax highlighting. YAML and Markdown prompts stay readable without adding a highlighter dependency; a future PR can layer highlighting on top.
• omitempty + optional TS field. Older results.json files just render without the panel — no schema-version bump needed.

Parallel approaches

This is one of two agent-generated implementations for #526. See the sibling GPT-5.5 PR for an alternative before merging.

[Copilot]

Pull request overview

Adds first-class support for capturing and surfacing the resolved eval prompt (plus prompt source path and follow-up prompts) in run results and the dashboard, addressing issue #526 by letting the UI display what instruction(s) the agent actually received.

Changes:

• Extend the Go results model (TestOutcome) and runner to record prompt, prompt_file, and follow_ups per task.
• Expose these new fields through the web API layer and validate mapping via a new unit test.
• Update the web dashboard (and TS API types) to render a prompt panel with copy-to-clipboard and follow-up prompt display.

Show a summary per file

File	Description
internal/models/outcome.go	Adds prompt/prompt_file/follow_ups fields to the persisted task outcome model.
internal/orchestration/runner.go	Attempts to populate the new prompt-related fields when producing TestOutcome.
internal/webapi/types.go	Extends API TaskResult response shape to include prompt data.
internal/webapi/store.go	Maps the new model fields into the API response payload.
internal/webapi/additional_test.go	Adds a unit test to ensure prompt fields are mapped into API output.
web/src/api/client.ts	Updates the TypeScript TaskResult interface with prompt fields.
web/src/components/RunDetail.tsx	Adds a new prompt panel UI within expanded task rows (copy + follow-ups).
web/dist/index.html	Updates hashed bundle references in the tracked index.html.

Review details

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

• Files reviewed: 9/10 changed files
• Comments generated: 1
• Review effort level: Lite

[Copilot]

Review details

Suppressed comments (6)

Previously missed (1) — in code that hasn't changed since the last review.

web/src/components/RunDetail.tsx:183

• This table contains expanded grader rows with 6 cells (GraderRow), so the table effectively has 6 columns. Using colSpan={5} here leaves an extra blank column and can cause the prompt panel background/borders to misalign with the grader detail rows.

This issue also appears in the following locations of the same file:

• line 192
• line 202

        <td colSpan={5} className="px-12 py-3 text-xs italic text-zinc-500">

internal/orchestration/runner.go:845

• PromptFile is being populated from tc.Stimulus.MessageFile, but TaskStimulus.resolvePromptFile clears MessageFile after loading the file (internal/models/testcase.go:566). As a result, tasks that used prompt_file will always end up with an empty PromptFile in results.json, so the dashboard can’t show the source path.

				outcomes = append(outcomes, models.TestOutcome{
					TestID:      tc.TestID,
					DisplayName: tc.DisplayName,
					Golden:      tc.Golden,
					Status:      models.StatusFailed,
					Prompt:      tc.Stimulus.Message,
					PromptFile:  tc.Stimulus.MessageFile,
					FollowUps:   tc.Stimulus.FollowUps,

internal/orchestration/runner.go:936

• PromptFile is being populated from test.Stimulus.MessageFile, but TaskStimulus.resolvePromptFile clears MessageFile after loading the file (internal/models/testcase.go:566). This means PromptFile will be empty for prompt_file tasks, so the API/UI won’t be able to display the source path.

					resultChan <- result{index: idx, outcome: models.TestOutcome{
						TestID:      test.TestID,
						DisplayName: test.DisplayName,
						Golden:      test.Golden,
						Status:      models.StatusFailed,
						Prompt:      test.Stimulus.Message,
						PromptFile:  test.Stimulus.MessageFile,
						FollowUps:   test.Stimulus.FollowUps,

internal/orchestration/runner.go:1121

• PromptFile is being populated from tc.Stimulus.MessageFile, but prompt_file resolution clears MessageFile (internal/models/testcase.go:566). That makes PromptFile always empty for prompt_file tasks, so this won’t actually persist the source path into results.json.

		Group:       r.resolveGroup(),
		Golden:      tc.Golden,
		Status:      status,
		Prompt:      tc.Stimulus.Message,
		PromptFile:  tc.Stimulus.MessageFile,
		FollowUps:   tc.Stimulus.FollowUps,

web/src/components/RunDetail.tsx:196

• navigator.clipboard.writeText can throw or reject (e.g., insecure context, permissions denied). Right now this would create an unhandled rejection and leave the UI in an inconsistent state. Please guard for clipboard availability and add a .catch handler.

  const handleCopy = () => {
    if (!copyable) return;
    void navigator.clipboard.writeText(copyable).then(() => {
      setCopied(true);
      window.setTimeout(() => setCopied(false), 1500);

web/src/components/RunDetail.tsx:202

• This prompt panel should span the full width of the expanded detail area. Because GraderRow renders 6 cells, the effective column count is 6; colSpan={5} can leave a trailing blank column and misaligned borders.

      <td colSpan={5} className="px-12 py-3">

• Files reviewed: 9/10 changed files
• Comments generated: 0 new
• Review effort level: Lite

[Shayne Boyer (spboyer)]

Closing as superseded by #542, the selected implementation for #526. This branch overlaps the same feature and additionally has unresolved prompt-file provenance, cache hydration, and regrade preservation concerns, so the two approaches should not both merge. No branch was deleted.