ADR-0006 — Scripted LanguageModelV2 for credential-free live-path testing

Status: accepted · 2026-07-30

Unit tests reach the browser half (surface registration, availability, bindings) and the server half (exposure, policy, approvals) separately, but neither exercises the seam between them: the chat route → Mastra loop → domain toolset → client-tool suspension → browser dispatch → reconciliation. That path is where the interesting failures live, and CI must never depend on a model provider to reach it.

Decision: MODEL_PROVIDER=mock (test-only) wires a hand-written scripted LanguageModelV2 into the same Mastra agent live mode uses. The script follows one scenario — read the ageing ladder, narrow the table, read it back, summarize — and deliberately batches a server tool and a client tool in the same message, because that is the case that suspends the run mid-step and makes the host answer the server call itself (ADR-0009). Playwright runs the suite in this mode, so the entire production pipeline — route, per-request catalog composition, collision check, NDJSON frames, browser dispatch, oRPC execution, reconciliation — is exercised with zero credentials.

ai/test's MockLanguageModelV2 is not used: it drags msw into the runtime graph. The mock is ~80 lines of plain LanguageModelV2 implementation.