fix(web): render agent-authored HTML views in chat again

[tusharbhardwaj-bk]

Problem

A t3_show_ui call left an ordinary collapsed tool row in the transcript and rendered nothing. Reported against thread bc541f47-d2d9-4a9d-91ba-524fe4c86ca9, where an agent sent 6 HTML views twice — the second time with all content statically baked in, on the theory that scripts were being blocked. Neither batch appeared.

The server side was never at fault. All 12 renders are stored correctly, and the render handle reaches the timeline row:

2026-08-31T19:00:10Z  html   6655  Option 6 — Heatmap Scanner
2026-08-31T18:59:44Z  html  10882  Option 5 — Swipeable Metric Pages
2026-08-31T18:58:55Z  html   7126  Option 4 — Compact List + Bottom-Sheet Drill-down
...  (12 rows, all kind=html)

The client refused to mount them. #145 disabled agent views behind a blanket kill switch:

export const AGENT_UI_SURFACES_RUNTIME_ENABLED = false;

AgentUiSurfaceRow short-circuits on it and returns the plain tool row. bkt3 is running 0436d9d6, which contains it.

Fix

The concern behind #145 was real but URL-only: a framed collaboration app can decline to join the room named by the URL and render unrelated origin-local state instead — a box that looks right and shows the wrong thing. The flag also took out agent-authored HTML, which has none of that problem. The document is what the agent produced, and it mounts from srcDoc in an unmodified sandbox="allow-scripts" iframe, so it has an opaque origin and can reach neither T3 nor the network as the signed-in user. AgentUiRenderFrame already blocked URL renders independently, so the global flag was over-broad and redundant — and it left the Experiments toggle permanently disabled.

This narrows the gate to the kind it was justified for. isAgentUiSurfaceRenderable() allows html, blocks url. A url handle still leaves the ordinary tool row, and the frame keeps its URL-blocked message for a stored render whose kind disagrees with the record.

Because the renders are already in the database, the views in the reported thread appear retroactively once this deploys — nothing needs re-sending.

Evidence

Before: in the reported thread every t3_show_ui call renders as a bare collapsed tool row — the wrench icon, the truncated JSON argument, and a chevron — with no view box beneath it. (Screenshot is in the originating T3 session; this agent has no way to upload an attachment.)

I have not verified this in a browser; the "after" is covered by tests, not a screenshot. New regression guard asserts an HTML surface mounts an iframe carrying sandbox="allow-scripts" and the agent's markup in srcdoc — the coverage whose absence let #145 disable the working half of the feature silently.

• vp test run apps/web/src/fork/ MessagesTimeline.agentUi.logic.test.ts ActivityPayloadProjection.agentUi.test.ts — 59 passed / 10 files
• vp run typecheck (@t3tools/web) — clean
• Cherry-picked onto 0436d9d6, the SHA bkt3 currently runs, and re-verified there

Note on the base

This targets bkmain directly at the maintainer's explicit request, skipping the usual expbkt3 soak. All touched source files are byte-identical between bkmain and expbkmain, so the change is isolated from the 9 mobile commits currently staged in expbkmain.

Claude Opus 5 via T3 Code.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ No successful main baseline artifact is available yet. This run establishes the initial measurement.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	—	13.5 KiB	—	15.1 KiB	✅
Codex	Thread snapshot wire	—	7.2 KiB	—	7.3 KiB	✅
Codex	Live turn WebSocket wire	—	6.4 KiB	—	7.8 KiB	✅
Codex	Live turn WebSocket decoded	—	55.6 KiB	—	66.4 KiB	✅
Codex	Live turn messages	—	10	—	21	✅
Claude	Total thread wire	—	13.6 KiB	—	15.1 KiB	✅
Claude	Thread snapshot wire	—	7.2 KiB	—	7.3 KiB	✅
Claude	Live turn WebSocket wire	—	6.4 KiB	—	7.8 KiB	✅
Claude	Live turn WebSocket decoded	—	56.5 KiB	—	66.4 KiB	✅
Claude	Live turn messages	—	11	—	21	✅

Baseline: unavailable · PR result: 2b5b30c · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 110.4 KiB
• Claude decoded thread snapshot: 111.1 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.