Improve the Forge MCP UI as it appears inside ChatGPT. #11959

Card payloads for forge_view, pulls, issues, issue, pull, runs and search are tightened; nothing was tried in the real ChatGPT connector. This sandbox can't reach it, so the cards are checked against the payload model and the existing MCP test suites (the 1073 mcp*/card* tests pass). bun test test/mcp-size.test.ts, bun x tsc --noEmit, bun scripts/gate.ts and bun scripts/design-lint.ts are green. I did not run the whole suite. The keep_plan tool wasn't available to me, so the plan on the issue was not kept.

Who controls what (now written into docs/MCP.md):

• ChatGPT controls:

- the frame around a card: its size, border and theme variables; - whether it renders ui:// pages at all; - the tool picker's name and description. The generic "Find issues, pulls…" line is the tool's description; it is written for the model and held to the size ceiling, so I left it alone.

• Forge controls:

- the structuredContent.card payload: title, status, rows of label/value/link/depth/actions; - the one page that draws it (CARD_SHELL).

I don't promise fonts, per-row colour or anything beyond title, status, rows, links and buttons.

Card vocabulary (also in docs/MCP.md):

• Title: says what the card is, cut at 90 characters on a word with …. The full text stays in the tool's text answer.
• Status: one line with the counts every row shares.
• Row label: the identity, #n title, linked when it has a page.
• Row value: only what differs between rows.

What changed:

• Titles: a clip() helper shortens titles on the pulls, issues, issue, pull, runs and search cards.
• Shell CSS: rows get a divider and a muted detail line. Text wraps instead of clipping.
• Pulls:

- the status now reads 3 open · 1 red · 2 green, counted over all returned pulls, not just the ones shown; - rows no longer repeat open; - they read check green · approved or no check · no review.

• Issues:

- a state shared by every row moves to the status line (2 open issues); - a mixed list shows 2 issues · 1 open · 1 closed and keeps the state per row.

• Issue: the title is now the card title, so the duplicate #n row is gone. A single "On the forge" row carries the link.
• forge_view: new card, ui://forge/forge. The status reads healthy · 3 running · 1 queued. It lists the apps with work going, busiest first, then counts all apps and what waits on you. docs/MCP.md is regenerated for the tool table.
• Unchanged: the health, apps, crew and runs cards, apart from the shell styling and the clipped runs titles. The issue asked for them, but they were already compact (status first, short rows, plain-word states), so I changed nothing more.
• Tests:

- test/mcp-card-vocabulary.test.ts is new and covers clipping, missing and long values, an HTML title passed through unescaped for the shell's textContent to handle, and forge_view output; - I updated the existing expectations in test/mcp-cards.test.ts, test/mcp-pulls-card.test.ts, test/mcp-search-issues-card.test.ts and test/mcp.test.ts. They pinned the old row wording, so the tests were out of date, not the code; - no data in the text answers changed.

Left out: colour or badges per row, since the host doesn't promise them; new tools (none added); the tool descriptions.

To try: call forge_view, pulls and issues through /mcp, or in a host that renders ui:// cards, and read structuredContent.card.

🤖 Generated with Claude Code

Closes #11956

Card payloads for forge_view, pulls, issues, issue, pull, runs and search are tightened; nothing was tried in the real ChatGPT connector. This sandbox can't reach it, so the cards are checked against the payload model and the existing MCP test suites (the 1073 mcp*/card* tests pass). bun test test/mcp-size.test.ts, bun x tsc --noEmit, bun scripts/gate.ts and bun scripts/design-lint.ts are green. I did not run the whole suite. The keep_plan tool wasn't available to me, so the plan on the issue was not kept.

Who controls what (now written into docs/MCP.md):

• ChatGPT controls:

- the frame around a card: its size, border and theme variables; - whether it renders ui:// pages at all; - the tool picker's name and description. The generic "Find issues, pulls…" line is the tool's description; it is written for the model and held to the size ceiling, so I left it alone.

• Forge controls:

- the structuredContent.card payload: title, status, rows of label/value/link/depth/actions; - the one page that draws it (CARD_SHELL).

I don't promise fonts, per-row colour or anything beyond title, status, rows, links and buttons.

Card vocabulary (also in docs/MCP.md):

• Title: says what the card is, cut at 90 characters on a word with …. The full text stays in the tool's text answer.
• Status: one line with the counts every row shares.
• Row label: the identity, #n title, linked when it has a page.
• Row value: only what differs between rows.

What changed:

• Titles: a clip() helper shortens titles on the pulls, issues, issue, pull, runs and search cards.
• Shell CSS: rows get a divider and a muted detail line. Text wraps instead of clipping.
• Pulls:

- the status now reads 3 open · 1 red · 2 green, counted over all returned pulls, not just the ones shown; - rows no longer repeat open; - they read check green · approved or no check · no review.

• Issues:

- a state shared by every row moves to the status line (2 open issues); - a mixed list shows 2 issues · 1 open · 1 closed and keeps the state per row.

• Issue: the title is now the card title, so the duplicate #n row is gone. A single "On the forge" row carries the link.
• forge_view: new card, ui://forge/forge. The status reads healthy · 3 running · 1 queued. It lists the apps with work going, busiest first, then counts all apps and what waits on you. docs/MCP.md is regenerated for the tool table.
• Unchanged: the health, apps, crew and runs cards, apart from the shell styling and the clipped runs titles. The issue asked for them, but they were already compact (status first, short rows, plain-word states), so I changed nothing more.
• Tests:

- test/mcp-card-vocabulary.test.ts is new and covers clipping, missing and long values, an HTML title passed through unescaped for the shell's textContent to handle, and forge_view output; - I updated the existing expectations in test/mcp-cards.test.ts, test/mcp-pulls-card.test.ts, test/mcp-search-issues-card.test.ts and test/mcp.test.ts. They pinned the old row wording, so the tests were out of date, not the code; - no data in the text answers changed.

Left out: colour or badges per row, since the host doesn't promise them; new tools (none added); the tool descriptions.

To try: call forge_view, pulls and issues through /mcp, or in a host that renders ui:// cards, and read structuredContent.card.

🤖 Generated with Claude Code

I fixed the real failure: bun scripts/gate.ts rejected three comment lines in src/web/studio/cards.ts that ran past 100 columns. I rewrapped them and pushed (3dbf19b56). The gate, tsc --noEmit and the card tests now pass; I did not rerun the MCP size test or the wider card suites after the rewrap, which only touched comments.

The preview deploy itself was not broken. The log ends with ready and done, and /healthz on the preview answers 200. "No tests declared, none run" only says forge.json has no app.test entry. This repository declares its tests under check, and the forge runs that on each push, so there was nothing to change in forge.json.

The pull description from my earlier message still stands unchanged.

The failing check was a stale test, not a bug in the app. I fixed the test and pushed.

An earlier commit moved the issue card's hostile text from the first row into the card title ("#1 <img …>"). test/xss-everywhere.test.ts still read rows[0].value, which is now alice/game #1. The test now asserts that card.title contains the hostile string. It still checks that nothing is HTML-escaped in the card data and that the card shell draws with textContent.

test/xss-everywhere.test.ts passes locally (13 of 13) and tsc --noEmit is clean. I didn't run the rest of the suite, so the forge's check on this push is the full verdict.