feat(web): AgentPrism trace viewer for eval runs

[JesseVent]

Summary

Adds a trace viewer to the web results app: every result row gets a "View span tree" link that opens the full span tree for that run — system/user messages, every tool call with its input/output, assistant turns, and check results — with per-span duration, token usage, and status.

What's added

Core (packages/core)

• trace-viewer.ts — adapter that turns an EvalResult (+ its raw transcript) into an AgentPrism span tree (evalResultToTraceSpans), plus tests (trace-viewer.test.ts).
• transcript/types.ts / parsers/adapt.ts — a TokenUsage type and usage?/ts? fields threaded through TranscriptEvent/TranscriptPart, so per-turn token counts and timestamps survive the transcript → trace pipeline instead of only being available as run-level aggregates.
• agents/{claude-code,codex,opencode}/parser.ts — each agent parser now extracts real per-turn usage (extractUsage/extractTurnUsage) from its own transcript format.
• index.ts — re-exports the above, plus per-step usage/timestamp capture in the AI SDK agent path.

Export (apps/framework/scripts/export-results.ts)

• Alongside the aggregate eval-results.json, writes one apps/web/src/data/traces/<evalId>.json per eval via evalResultToTraceSpans, preferring a run that actually recorded a transcript over one that didn't (so the viewer shows a real span tree, not an empty no-skills run). Lazy/one-per-file so the aggregate bundle stays lean — the web app fetches a trace only when its row is opened. New --no-traces flag skips writing them.

Web (apps/web)

• components/agent-prism/** — the AgentPrism component set (span cards, tree view, details view with Attributes/In-Out/Raw tabs, theme tokens) vendored into the repo.
• components/trace-panel.tsx — the integration point: fetches a trace's JSON lazily, revives Date fields (JSON round-trips them as strings, which breaks AgentPrism's duration math), and renders TraceViewer full-screen with an Escape-to-close overlay.
• components/results/eval-details.tsx — adds the "View span tree" link.
• Tool call input/output defaults to a human-readable "Plain" view with real line breaks (multi-line shell commands/output are unreadable as JSON's escaped \n form); "JSON" stays available as a tab for exact, copy-pasteable output.
• Trimmed the vendored TraceViewer to fit how we actually use it: it's built for a multi-trace list (pick between several loaded traces), but we only ever open one trace at a time, so the always-one-item trace-list sidebar (and, on mobile, its matching "back to list" dead-end button) is hidden — the span tree gets the full width instead.

How to try it

pnpm install
pnpm --filter web dev

Open the results table, click any row, click "View span tree".

Agent Prism Trace Visualiser

## Notes

• Rebased cleanly onto current main: kept main's regenerated eval-results.json/regression-eval-results.json instead of the stale snapshot this originally carried, and dropped everything that had leaked in from an unrelated, not-yet-merged skill-trigger-suite branch (trigger trace fixtures, EXPERIMENT_SUITES/url-state.ts entries, a README section) — this PR is the viewer only.
• pnpm --filter core test (131/131), pnpm --filter web run typecheck, and pnpm --filter web run build all pass. One pre-existing eval-results.test.ts data-consistency assertion fails on this branch — confirmed it fails identically on plain main today, unrelated to this change.

[vercel]

@JesseVent is attempting to deploy a commit to the Supabase Team on Vercel.

A member of the Team first needs to authorize it.