The self-maintaining supply chain for AI agents.

One MCP/API tool change can silently break multiple agents across repositories. Code search alone cannot see shared skills, ownership, business meaning, or governance policy.

Validated workflow at a glance

The complete proof chain, from the changed contract to independent evaluation, in one read.

01 Changed API contract
customer-context.lookup_customer · 1.0.0 → 2.0.0
02 3 affected consumers
Known live scenario mapped through DataHub
03 3 actual repair repositories
2 direct · 1 shared-skill target
04 5 accepted Codex candidates
3 known · 2 no-gold holdout
05 All 7 evaluation layers passed
98 of 98 checks passed

Two proofs, one boundary

The known layout proves the live provider path. The materially different holdout proves the same evidence-driven workflow without a gold patch.

Scenario	Changed tool	Affected	Repair targets	Memorable proof	Status
customer-context.lookup_customer Known live scenario	customer-context.lookup_customer 1.0.0 → 2.0.0	3 consumers	3 repositories	124500 cents became $1,245.00	Human review
operations-context.get_job_status No-gold holdout	operations-context.get_job_status 1.0.0 → 2.0.0	2 consumers	2 repositories	6500 milliseconds became 6.5 seconds	Human review

Why DataHub is indispensable

DataHub is the organizational context graph connecting agents, skills, tools, repositories, owners, schemas, business definitions, and governance policies.

Code search sees

Calls inside one repository

Local imports and call sites
Fields referenced in source
Repository-native tests

MCPatch plus DataHub sees

The organizational dependency chain

Direct and shared-skill consumers
Repository ownership and reviewers
Business units, semantics, and PII policy

No merge or deployment

Every accepted result stops at a verified draft pull request for human review. This replay performs no DataHub, GitHub, Codex, Docker, internet, or local-network activity.