MCP crew, health and capacity draw ui:// cards #11529

The MCP crew, health and capacity tools now return ui://forge/<name> cards beside their JSON. I didn't call the tools live. The card tests run cardFor directly, so only mcp.test.ts checks the tool wiring. tsc, design-lint, test/mcp-size.test.ts, the new card tests and the existing MCP card tests all pass, and I raised no ceilings. bun scripts/gate.ts ran, but I didn't look at its output.

• One helper. All three cards are branches of the one cardFor in src/web/studio/cards.ts, not three helpers.
• crew: the card reuses ui://forge/crew, which now already draws, and tells the two apart by the answer's shape. It shows one row per seat: who fills it, and how many runs are running and waiting. An empty seat says what leaving it empty costs. Rows link to the repo's crew settings. The kind: week answer and dry-run plans draw no card. The listing passes the repo through src/http/mcp/tools/crew.ts.
• health: the card is green or red with the version. Rows list draining, a blocked or held rollout, degraded lanes, incidents and app counts. A stranger's answer, which has only the verdict, gives no rows.
• capacity: one row per setting with its live value, and "set" marked on admin-changed ones. Rows link to /settings/capacity, and the status says when the forge cannot swap itself. A plan has no settings, so it draws no card.
• Safety: every field is escaped by the card shell, and all three tools keep their existing readOnlyHint/writes flags unchanged.
• Tests and docs: test/mcp-cards-crew.test.ts renders each card full and empty, with an injection string in the data, and checks that non-listing answers draw nothing. I added the two new resource URIs to the resource list in test/mcp.test.ts and the card rows to docs/MCP.md. bun scripts/mcp-table.ts was run, and the card rows are the only change to docs/MCP.md.

Not done: the crew card does not show a repo's governor line (slots, ceiling, held), and the capacity card does not show the egress figures or the changes log. The page keeps those.

To try it, call crew {repo}, health or capacity from a card-drawing harness. Cards appear in structuredContent.card.

The pre-push hook required a commit message with a "Try:" line, so there is an empty commit carrying it.

Closes #11492

The MCP crew, health and capacity tools now return ui://forge/<name> cards beside their JSON. I didn't call the tools live. The card tests run cardFor directly, so only mcp.test.ts checks the tool wiring. tsc, design-lint, test/mcp-size.test.ts, the new card tests and the existing MCP card tests all pass, and I raised no ceilings. bun scripts/gate.ts ran, but I didn't look at its output.

• One helper. All three cards are branches of the one cardFor in src/web/studio/cards.ts, not three helpers.
• crew: the card reuses ui://forge/crew, which now already draws, and tells the two apart by the answer's shape. It shows one row per seat: who fills it, and how many runs are running and waiting. An empty seat says what leaving it empty costs. Rows link to the repo's crew settings. The kind: week answer and dry-run plans draw no card. The listing passes the repo through src/http/mcp/tools/crew.ts.
• health: the card is green or red with the version. Rows list draining, a blocked or held rollout, degraded lanes, incidents and app counts. A stranger's answer, which has only the verdict, gives no rows.
• capacity: one row per setting with its live value, and "set" marked on admin-changed ones. Rows link to /settings/capacity, and the status says when the forge cannot swap itself. A plan has no settings, so it draws no card.
• Safety: every field is escaped by the card shell, and all three tools keep their existing readOnlyHint/writes flags unchanged.
• Tests and docs: test/mcp-cards-crew.test.ts renders each card full and empty, with an injection string in the data, and checks that non-listing answers draw nothing. I added the two new resource URIs to the resource list in test/mcp.test.ts and the card rows to docs/MCP.md. bun scripts/mcp-table.ts was run, and the card rows are the only change to docs/MCP.md.

Not done: the crew card does not show a repo's governor line (slots, ceiling, held), and the capacity card does not show the egress figures or the changes log. The page keeps those.

To try it, call crew {repo}, health or capacity from a card-drawing harness. Cards appear in structuredContent.card.

The pre-push hook required a commit message with a "Try:" line, so there is an empty commit carrying it.

I changed no code for this failure. I pushed an empty commit to make the forge re-run the check.

The check was stopped for using too much CPU: 18.9 cores for 30 seconds against a limit of 16. It was not a test failure. The log's last lines show test shards 1–4 had started and a browser, whose egress requests were refused, was running. Nothing in my pull starts a process or a browser. It adds card-drawing branches in src/web/studio/cards.ts, a listing card in src/http/mcp/tools/crew.ts, docs and two test files, and the MCP tests that cover them passed locally.

If the re-run stops on the CPU limit again, the cause is in the shared test shards or a browser-driving test, and it needs someone who can see the whole check.

I pushed a fix for the egress failure. Chromium's own background traffic is the likely cause, but I haven't seen the check re-run, so it isn't confirmed.

The refused hosts (update.googleapis.com, clients2.google.com, redirector.gvt1.com, start.duckduckgo.com) are Chromium's own update and search traffic, not anything my change calls. Two test launchers started headless Chromium without the flags that turn that traffic off:

• test/helpers/chrome.ts: the chromiumRuns probe, which many browser tests call, now passes --disable-background-networking, --disable-component-update and --disable-sync.
• test/kpi-wrap.test.ts: it now passes --disable-background-networking and --disable-component-update.

My cards.ts edit is probably what exposed this. Many tests import that file, so touching it selected a large batch of browser-driving tests, and those launches were what the check refused and what pushed CPU past the limit.

Locally tsc is clean and test/kpi-wrap.test.ts and test/mcp-cards-crew.test.ts pass, with one skip in that run. I did not run the browser-driven suites.

The crew, health and capacity tools now return ui:// cards, and all three review points are addressed. The gate, tsc --noEmit, design-lint and the four test files covering this change pass. I didn't run the preview or the tools live.

Scope. I reverted test/helpers/chrome.ts and test/kpi-wrap.test.ts to match main. The Chromium flake, if it is real, needs its own pull.

Tool wiring tests. New test/mcp-crew-health-capacity-tool-cards.test.ts calls crew {repo}, health and capacity through tools/call.

• It asserts that crew returns card {repo}, that the card status says "of alice/engine", and that every row links to /alice/engine/settings/crew.
• It asserts that health and capacity carry their cards, with capacity rows linked to /settings/capacity.
• I put the main version of crew.ts back and the crew test failed, so it does depend on the change.

Card fixes in src/web/studio/cards.ts.

• Crew: seats are now capped at 12, with a "More … not shown" row.
• Capacity: the card adds a "More" row when it slices the settings. The status still counts all settings, so the number and the rows now agree.
• Health, wording: degraded rows are labelled "Needs attention" and incident rows "Incident". The raw queue:agent key never appears.
• Health, links: degraded rows link to their own href when it is a rooted path on this forge, and otherwise to /. The Apps row now links to /. Degraded rows are capped at 12 with a "More" row.

I extended test/mcp-cards-crew.test.ts to cover the new labels, the links and the caps.

Not done. Incident rows still show the raw message text and have no link of their own.