`runs` answers with a ui://forge/runs card: each run's kind, state, issue or pull, and duration, newest first, linked to its log #11161

The runs tool now draws a ui://forge/runs card, one row per run, newest first. I changed nothing in this session: the card and its tests were already committed and pushed on this branch. I re-ran the checks and they pass: bun scripts/gate.ts, bun x tsc --noEmit, bun scripts/design-lint.ts, and test/mcp-size.test.ts, test/mcp-runs-card.test.ts, test/mcp-runs.test.ts and test/mcp.test.ts (58 tests). No ceiling was raised.

What changed

• Rows: runs (src/http/mcp/tools/runs.ts) now returns a runs card, built in src/web/studio/cards.ts beside the pulls and map cards. Each row shows a glyph, the kind, the issue or pull number and its title. The value is the state as a word plus how long it ran. A running run reads running · running for 2m.
• Failed first: a failed run is named first in the status line, for example 1 failed · 1 running · 2 runs.
• Links: each row links to /<owner>/<name>/runs/<id>. A bad id or repo slug gets no link.
• Escaping: card values are plain data and the shell never uses innerHTML. test/mcp-runs-card.test.ts renders a run with a hostile title and asserts it stays inert text.
• Visibility: the tool still goes through listRepoRuns and repoArg, so a reader only sees runs of repositories they can read.
• Docs and tests: the runs tool doc and docs/MCP.md (tool table and the Cards paragraph) describe the card. test/mcp-runs.test.ts and test/mcp.test.ts expect the new title key and the ui://forge/runs resource.

Left out

• The card shows at most PULLS_LIMIT rows, the same cap the pulls card uses, even though the tool can return up to 50 runs.
• I did not add a filter control to the card. The kind argument on the tool covers that.

Try it

• Call runs with { "repo": "owner/name" } from an MCP Apps host.
• Or run bun test test/mcp-runs-card.test.ts.

Closes #11122

The runs tool now draws a ui://forge/runs card, one row per run, newest first. I changed nothing in this session: the card and its tests were already committed and pushed on this branch. I re-ran the checks and they pass: bun scripts/gate.ts, bun x tsc --noEmit, bun scripts/design-lint.ts, and test/mcp-size.test.ts, test/mcp-runs-card.test.ts, test/mcp-runs.test.ts and test/mcp.test.ts (58 tests). No ceiling was raised.

What changed

• Rows: runs (src/http/mcp/tools/runs.ts) now returns a runs card, built in src/web/studio/cards.ts beside the pulls and map cards. Each row shows a glyph, the kind, the issue or pull number and its title. The value is the state as a word plus how long it ran. A running run reads running · running for 2m.
• Failed first: a failed run is named first in the status line, for example 1 failed · 1 running · 2 runs.
• Links: each row links to /<owner>/<name>/runs/<id>. A bad id or repo slug gets no link.
• Escaping: card values are plain data and the shell never uses innerHTML. test/mcp-runs-card.test.ts renders a run with a hostile title and asserts it stays inert text.
• Visibility: the tool still goes through listRepoRuns and repoArg, so a reader only sees runs of repositories they can read.
• Docs and tests: the runs tool doc and docs/MCP.md (tool table and the Cards paragraph) describe the card. test/mcp-runs.test.ts and test/mcp.test.ts expect the new title key and the ui://forge/runs resource.

Left out

• The card shows at most PULLS_LIMIT rows, the same cap the pulls card uses, even though the tool can return up to 50 runs.
• I did not add a filter control to the card. The kind argument on the tool covers that.

Try it

• Call runs with { "repo": "owner/name" } from an MCP Apps host.
• Or run bun test test/mcp-runs-card.test.ts.

I merged main into the branch, resolved the conflicts and pushed (fdbacc089). The conflicts were in docs/MCP.md and src/web/studio/cards.ts. The gate, tsc, design-lint and the card and MCP tests (86 across eight files, including test/mcp-size.test.ts) all pass. I did not run the full suite; the forge's check does that on the push.

How I resolved them

• Card kinds: main had renamed the decision card kind to asks and added a release kind. I kept both and added runs to the card kind type and CARD_KINDS.
• Tool map: runs still draws the new runs card, and asks uses main's asks kind.
• Card builders: the runs and release builders now sit side by side in src/web/studio/cards.ts.
• Docs: the Cards paragraph in docs/MCP.md now lists releases, runs and the Runs card description together.

No ceiling was raised.

I fixed all three of sentinel's points, and the branch is pushed (f8252963e). The gate, tsc and the MCP tests, including test/mcp-size.test.ts, pass.

• Sealed titles: runs now checks mayView(db, repo, issue, user) for each run's issue and returns number: null, title: null when the reader may not see it. I disagree with the finding as written, though. listRepoRuns already drops runs on a sealed issue for readers who may not see it (mayViewId), so this was never reachable. The new check is a second guard. The new test shows a read-only reader of a public repo gets no runs, and neither the title nor the number appears anywhere in the answer. The issue's author still sees their own run.
• Untrusted titles: a title from someone who cannot write now carries untrusted: true, using markReaders as the issues tool does. A test covers it, and the tool doc says so.
• Indentation: the map body in runs.ts was rewritten and re-indented as one block.
• Failed first: test/mcp-runs-card.test.ts now asserts the status line starts with 1 failed and the failed row's value starts with failed. I did not reorder the rows. The card stays newest first, with failed leading the status line.