[timeleft--]

Addresses Q1 of iaser-ai#335.

What changes

Card address. The card is served at ui://views/answer-card.v1.html, which names its compatibility line, not a hash of its bytes. Every compatible release keeps this address, including a release stamp or a card change. So a host that still holds a tool descriptor from before a deploy reads a card the server still serves. VIEW_MAJOR in src/server/view-artifact.ts goes up only for a change an already-loaded card can't handle.

SAGA 1.0 release candidate. The SDK is packed from its source at commit 8b4d987 and committed under apps/mcp/candidates/ until 1.0.0-rc.1 is on the npm registry.

• The manifest pins the exact candidate version, and the root pnpm-workspace.yaml overrides the five siblings to those tarballs.
• @modelcontextprotocol/server and client move to 2.4.0, hono to 4.13.16 and @hono/node-server to 2.1.5. The root overrides for dompurify and @hono/node-server follow the SDK's documentation; apps/mcp is their only consumer.
• The Dev release gate accepts the candidate. Beta and Production refuse it, so neither deploys until the release is published.
• The card's SDK identity names the candidate by its source commit and tarball digests.

Less SAGA glue. SAGA's createMcpFetchApp now serves /mcp: Host and Origin admission, body bounds, the request budget, lifetime attachment and the official stateless handler. Ansari keeps its own routes (/health and the Dev-only diagnostics beacon) and its logs. The server factory receives the request lifetime instead of building budgets, and the budget stays at the tool's 60 s bound. mcp.error carries SAGA's fixed failure kind.

Tests

• Retired, because they only re-checked what SAGA tests itself: request-timer release when a server closes (two fixtures), and the malformed and oversize body refusals.
• Kept, because they check Ansari's own wiring: the 60 s budget, deployment allowlists replacing localhost on both targets, the upstream abort on timeout, and the Node listener's SIGTERM shutdown.
• The packaging test checks a candidate against the committed tarball bytes, and a registry release against its npm integrity.
• Address tests expect the v1 address. A release changes the card's bytes, never its address.

Also: app version 0.17.0, the SAGA credit links to the SDK repository, and the README and walkthrough describe the above.

Verification

• pnpm mcp verify passed: lint, typecheck, 145 tests and 78 browser checks.
• Changing one byte of a committed tarball fails the lockfile-integrity and card-identity checks; restoring it passes them.
• bash apps/mcp/scripts/verify-container.sh passed. The root-context Docker build installs from the committed tarballs, and the container serves ui://views/answer-card.v1.html.
• pnpm check:release --target dev passes. --target beta refuses the candidate.

Before upstream

When 1.0.0-rc.1 is on the npm registry:

1. Replace the candidate pin and the five SAGA overrides with the registry version.
2. Delete apps/mcp/candidates/.
3. Add the registry integrity to test/packaging.test.ts.

🤖 Generated with Claude Code

[yia-mw-agent]

Helmet-Review-Request: #12 (comment)

Reviewed the full merge-base-to-head diff (base develop 7c605bb = merge-base, 28 files) at 21cc907, which was also the live head at start and before submitting. No material defects.

Checks run on that head:

• pnpm install --frozen-lockfile --filter ansari-mcp... succeeds, so the lockfile and the five root overrides agree.
• pnpm mcp lint, typecheck and test pass (145 of 145).
• pnpm test:browser passes (78 checks, no errors).
• check-release --target dev passes; --target beta and --target production refuse the candidate as the PR says, so neither can deploy it before the registry release.
• The five committed tarballs all report 0.0.0-candidate.g8b4d987dd233. A scan found no machine paths or credential markers in them. The packaging test checks the lockfile integrity against the committed bytes.
• The MCP application and frontend CI jobs pass; the api and auth jobs were still pending when I looked, and this PR does not touch them.

What I checked in the code:

• createMcpFetchApp replaces the hand-built Host/Origin, body and budget wiring. requestMs still defaults to the 60 s HANDLER_BOUND_MS, so the SDK's 10 s default does not clamp the tool.
• The lifetime is passed into createAnsariMcpServer. The retired fixtures only re-tested behavior the SDK now owns, and the Ansari-specific checks stay: the 60 s budget, the allowlists replacing localhost, upstream abort on timeout, and SIGTERM shutdown.
• VIEW_MAJOR = 1 gives ui://views/answer-card.v1.html on both the closed and diagnostic cards. The tests and the Docker smoke script now expect the stable address.
• sdk-identity.mjs, the build stamp and the packaging test compute the same candidate identity, and the check-release regex accepts the apps/mcp/candidates path.

Non-blocking:

• apps/mcp/README.md line 104 (MCP_VIEW_DIAGNOSTICS_ORIGIN row) says a diagnostics origin "changes the card's bytes, address and permissions", and line 139 says "changes bytes, address and permissions". The address no longer changes (view-artifact.ts and diagnostics.test.ts say so). Dropping "address" from both would make the docs consistent.
• Before upstream, replace the candidate pin as the PR describes. Until then the tarballs are the only record of what the SDK candidate contained.