mirror of
https://github.com/mims-harvard/ToolUniverse.git
synced 2026-09-19 07:31:47 +08:00
Shanghua Gao
406add8bf0
Pinned third-party MCP tools were stored only as a contract_sha256. A hash can show that an upstream server changed a tool but never what it changed, so a maintainer asked to re-pin had nothing to review and could only accept blind -- which is what re-pinning the nine drifted Folklore and Genomic Intelligence contracts actually amounted to. Adds src/tooluniverse/data/mcp_contracts.lock.json recording the reviewed inputSchema/outputSchema of all 27 pinned tools across the four loaders, so upstream drift shows up as a readable diff instead of two hex strings. scripts/sync_mcp_contracts.py has three modes: --check-config (offline; the lockfile and the configs must agree), --check-live (diff the servers against the lockfile) and --update (re-record, producing the diff to review). The new workflow moves detection off the user's critical path. Contract verification currently fires inside the user's process, where nobody can review a third-party schema or re-pin it; Folklore and Genomic Intelligence were each serving zero tools for an unknown period, found only by accident, because the weekly health check cannot see them -- MCP loader URLs are opt-in env vars CI never sets. check-config gates every PR; check-live runs daily and fails on real drift while skipping servers it cannot reach, so a vendor outage or rate limit does not read as drift. The drift report calls out new input fields specifically: the agent fills in any parameter a server advertises, so an added field is a new path for data to leave the session.