Skip to content

Latest commit

 

History

History

fixtures

README.md

Target-drift matched-pair fixtures

Sanitized, fully synthetic server.json A/B pairs for the target-drift signal discussed in modelcontextprotocol/registry#1706. No real publishers, no token-bearing URLs, no real tool arguments — registry-schema-shaped only.

Each case directory holds:

  • a.server.json — the earlier version (by publishedAt). Absent in the UNKNOWN case.
  • b.server.json — the current version.
  • expected.json — the targetDrift signal the Registry should emit (server.json-derivable only), and the action a client must take.

A registry reference diff and a client gate (e.g. the matched-pair suite in peach420fuzz/objekts-production-desk) should consume the same pairs; mismatches become reproducible cases.

The through-line

The registry signal is necessary, not sufficient. It is a cheap shared hint that lets a client skip re-diffing history. It never replaces the client re-pinning against its own exact target and its live tools/list. A false from the registry must never suppress client re-consent, and a missing signal is UNKNOWN, never "unchanged."

Cases

# Case Registry sees Client action Point
01 query-key-set-change remotesChanged: true re-consent Registry can report query keys (not values) safely.
02 query-value-rotation remotesChanged: false re-consent Same key, new value. Registry sees nothing; pinned target still changed. The invariant case.
03 same-scope-package-rename packageOwnersChanged: false (+ identifier change) re-consent Owner scope unchanged, installed identity changed.
04 non-semver-ordering remotesChanged: true re-consent "Previous" resolved by publishedAt, not the text version sort.
05 tools-only-drift all false re-consent Target byte-identical; tool contract changed — only the client (via tools/list) can see it.
06 unknown-no-baseline UNKNOWN (absent) ask No parseable prior version. Absent ≠ unchanged; never auto-pass.
07 version-ordering-record remotesChanged: true (by publishedAt) re-consent A 3-version history where a naive version-string sort picks the wrong current+previous and silently passes, while publishedAt ordering catches the real origin change.

Case 05 — executable tools/list arrays

Case 05's drift lives only in tools/list, so the pair ships client-side arrays alongside the server.json:

  • 05-tools-only-drift/a.tools.json / b.tools.json — the result.tools arrays a client observes (A = baseline, B = current).
  • 05-tools-only-drift/tools.expected.json — the tool delta a client must act on: send_email added (write/destructive) and read_doc's readOnlyHint flipped true → false under the same name and inputSchema. The registry cannot see any of it.

Case 07 — version-ordering record

  • 07-version-ordering-record/versions.json — a synthetic GET /servers/{name}/versions record for one server, listed in descending version-string order on purpose to bait a text/semver-sort consumer. Current is the latest publishedAt, not the highest version string.
  • 07-version-ordering-record/expected.json — correct resolution by publishedAt (remotesChanged: true) vs the naive version-desc trap (remotesChanged: false, a dangerous silent pass). The divergence is the test.

Comparison policy encoded here

  • comparedTo.previousPublished and comparedTo.first are resolved by publishedAt (GET /servers/{id}/versions is ordered by publication date; publishedAt is a first-class field), not by the text version sort the list endpoint uses.
  • Remote URLs compare origin, port, path, and query keys — never query values.
  • Packages compare owner scope (packageOwnersChanged) and identifier; a new version under the same identifier is a routine version change, not a target change.