Weekly research scan [2026-08-03]

[jgoldfoot]

Automated weekly research scan. Draft — do not merge without review.

Window: 2026-07-21 → 2026-08-03 (prior scan #146 closed 2026-07-28).

Findings

#	Category	Claim	Primary source
1	[STALE]	MCP 2026-07-28 was cited as a release candidate locked 21 May 2026. It was officially released on 28 July 2026.	MCP blog, 2026-07-28
2	[CITATION]	Whitepaper changelog link pointed at /specification/draft/changelog, which now tracks the next revision.	2026-07-28 changelog
3	[NEW]	server/discover — servers MUST implement this RPC to advertise protocol versions, capabilities, identity, and extensions. Replaces the removed initialize handshake.	Changelog, major change 3
4	[NEW]	Multi Round-Trip Requests (MRTR) replaces server-initiated requests (roots/list, sampling/createMessage, elicitation/create) with InputRequiredResult + client retry.	Changelog, major change 7
5	[STALE]	docs/implementation-guide.md Pattern 11 teaches Sampling and Roots with no caveat; both are now formally Deprecated (12-month window). The whitepaper already recorded this — the guide had not.	Changelog, Deprecated 1 (SEP-2577)
6	[CITATION]	WebMCP was cited as "Draft Community Group Report (20 July 2026)". The document is a continuously updated editor's draft whose ReSpec header re-dates on every commit — it has no stable publication date.	Editor's draft + commit log

Proposed edits

• README.md — MCP entry marked released, plus server/discover, MRTR, cacheable list results. WebMCP entry relabelled as an undated editor's draft (retrieved 3 Aug 2026).
• docs/whitepaper.md — same two citation fixes; changelog URL pinned to the versioned path; two new paragraphs in the MCP section covering server/discover and MRTR.
• docs/implementation-guide.md — deprecation caveat on Pattern 11. The pattern itself is not rewritten; that is an authoring decision.
• DRIFT.md — server/discover narrows (but does not resolve) the open application/mcp+json question; new [DRIFT] entry for the Pattern 11 / whitepaper inconsistency.

Verification

Every claim above was confirmed against a primary source (official spec changelog, official protocol blog, or the spec repo's own commit log). The skeptic pass cut three candidate findings that secondary sources reported but primary sources did not support:

1. navigator.modelContext — several secondary write-ups state WebMCP exposes navigator.modelContext. The spec's IDL defines partial interface Document { readonly attribute ModelContext modelContext; }. The framework's existing document.modelContext is correct and was left unchanged.
2. OSWorld-Verified SOTA figures (e.g. "Claude Mythos Preview 79.6%", "GPT-5.4 75.0% vs. a 72.4% human baseline") — available only from third-party leaderboard aggregators, partly self-reported, no primary confirmation. Cut entirely; no OSWorld SOTA claim was added.
3. WebMCP Chrome origin trial (Chrome 149–156) — asserted by secondary blogs; Chrome Platform Status did not render a verifiable feature entry. Cut.

Also checked and found not to need changes: schema.org v30.0 (March 2026, outside window; framework pins no version), A2A v1.0 (already cited correctly), WAI-ARIA — 1.3 is a Working Draft only, so the framework's citation of 1.2 (the current Recommendation) is correct and was deliberately not "upgraded".

Conventions respected

Benchmark numbers remain directional projections; data-agent-* remains supplementary; npm not pnpm; drift recorded in DRIFT.md only.

Checks

• npm test — 122 passed, 7 suites, run before opening.
• No changes to package.json, package-lock.json, or .github/workflows.
• Unrelated Prettier churn (README badge links, a TS union wrap in the implementation guide) was reverted to keep the diff scoped.

🤖 Generated with Claude Code