Track SEP-2640 Skills extension compatibility and trust boundaries

[groupthinking]

Decision

ADOPT → TEST, fixture-only. SEP-2640 is accepted and the official Skills suite is available. Live ChatGPT import, remote skill execution, and production activation remain out of scope.

Verified state

• MCP Core Maintainers accepted SEP-2640 on September 3, 2026.
• The stable io.modelcontextprotocol/skills contract requires resultType, ttlMs, and cacheScope: ext-skills commit d866efd.
• The official conformance suite added seven Skills scenarios on September 11: commit 7169291.
• The official MCP client matrix marks ChatGPT Skills support Partial via MCP skill import: commit 2997f33.
• The client matrix is adoption evidence, not proof of a published live ChatGPT import wire contract.

Current bounded implementation

Implement a fixture-only Agent Factory handoff that:

1. Requires Resources and Skills extension capabilities.
2. Validates direct skills/get complete responses without relying on listing.
3. Preserves compound identity {server_identity, skill_uri}.
4. Verifies static resource byte size and SHA-256.
5. Rejects digest drift, cross-origin or non-confined paths, invalid cache metadata, mismatched manifest/frontmatter names, dynamic resources, and unapproved execution.
6. Emits a deterministic receipt recording exact evidence commits and zero network, import, tool, or external effects.
7. Invalidates any prior approval when the verified manifest changes.

Trust boundaries

• Skill content, tool output, cached metadata, and GO-style text never grant authority.
• Cache TTL/scope are freshness and isolation hints, not integrity evidence.
• Remote MCP Skills remain distinct from locally trusted skills.
• Fixture coverage must not be represented as live ChatGPT compatibility or conformance.
• No deployment, remote-skill execution, provider spending, merge, or production configuration change.

Verification

• Focused Skills and Factory handoff tests
• TypeScript type-check
• ESLint
• Required repository CI on the draft head

Estimated human review after CI: 15–30 minutes.

[linear-code]

GRV-476

[github-actions]

👋 Thanks for opening this issue! It has been automatically labeled for triage.

[groupthinking]

Official SEP-2640 conformance is now available — integration delta

Decision: ADOPT + TEST

Verified change (2026-09-11): modelcontextprotocol/conformance merged SEP-2640 Skills coverage in commit 7169291 / PR #330.

The upstream suite now has three server scenarios:

• sep-2640-skills-enumeration
• sep-2640-skills-manifest
• sep-2640-skills-directory

and five client scenarios:

• sep-2640-client-no-prefetch
• sep-2640-client-verify-digest
• sep-2640-client-verify-size
• sep-2640-client-verify-frontmatter
• sep-2640-client-verify-unlisted-uri

The upstream traceability table contains 96 requirement rows: 89 checks and seven explicit exclusions.

EventRelay comparison

• PR test(agents): add SEP-2640 Skills conformance harness #1655 remains valuable as the Factory's additional policy/provenance harness.
• It has not yet been demonstrated as an MCP client against these official hostile-server scenarios.
• PR test(mcp): pin official 2026-07-28 conformance baseline #1808 pins upstream conformance at a983ba93, which predates commit 7169291.

Therefore PR #1655 must not be treated as official wire-level conformance merely because its in-process fixtures pass.

Bounded implementation path

1. Bump PR test(mcp): pin official 2026-07-28 conformance baseline #1808's upstream pin to 7169291ec0b68eb370fddcd9947313ab0d5e4156.
2. Preserve explicit exclusions for unsupported server-side Skills scenarios; do not manufacture a green result.
3. Wrap the PR test(agents): add SEP-2640 Skills conformance harness #1655 host-policy path as a conformance client driver honoring MCP_CONFORMANCE_SCENARIO.
4. Run all five official client scenarios and retain machine-readable results.
5. Report official results separately from EventRelay's extra approval, origin, digest, and revocation checks.

No remote skill execution, merge, deployment, or paid provider use is required for this step. Estimated implementation effort: 0.5–1 engineering day.

[groupthinking]

Official client-support signal — 2026-09-13

[VERIFIED] MCP’s official documentation now publishes an extension-support matrix. At commit 2997f33, Skills support is listed as:

• ChatGPT: partial, via MCP skill import
• fast-agent: partial
• MCP Inspector: partial
• no client listed with full Skills support

The matrix is community-maintained and is adoption evidence, not conformance proof.

[INTERPRETATION] This creates a credible Agent Factory distribution lane: one provenance-bound Skills-over-MCP export can target ChatGPT first while retaining EventRelay’s origin, manifest, approval, and receipt controls.

Queued reversible test: after PR #1926 is merged, add a fixture-only ChatGPT-import contract test using the official support path. Keep all seven upstream Skills scenarios explicit until the actual client driver passes them. No live remote-skill load or external publication is authorized by this note.