Port model/agent capture to v2 Model.Ref shape

[frap129]

Part of #70

Question

The v2 context hook's SessionContext.model is a Model.Ref ({ providerID, id, variant? } — Model.ID is under id, verified in @opencode-ai/schema@beta model.js), but captureSessionContext (src/runtime/chat-capture.ts) reads input.model?.modelID and orchestrator.ts declares model?: { modelID?: string }. Under the real hook, lastModelID is never captured, so every model:-conditioned rule silently fails to match.

Fix the capture (read model.id, keep modelID as the internal RuleMatchContext field name), update the V2SessionContextInput type, and adjust test fixtures that pass { modelID: 'x' } as the hook's model ref. Also check whether the variant suffix (model#high parsing) matters for model: matching and decide accept/defer.

[frap129]

Resolution

Fix applied on the v2 branch working tree (part of the uncommitted v2 port):

• captureSessionContext (src/runtime/chat-capture.ts) now reads input.model?.id (Model.Ref) and keeps modelID as the internal RuleMatchContext/CapturedTurnContext field name. lastModelID is now actually captured under the real hook — model:-conditioned rules match again.
• V2SessionContextInput.model (src/runtime/orchestrator.ts) retyped to { id?: string; providerID?: string; variant?: string }.
• Test fixtures updated: every context-hook model ref { modelID: 'x' } → { id: 'x' } (index.test.ts harness types + call sites, index.runtime.test.ts, index.integration.test.ts, chat-message-persistence.test.ts, chat-capture.test.ts), and history-assistant mocks of session.context() returns model: { providerID: 'p', modelID: 'm' } → { providerID: 'p', id: 'm' } (real SessionMessageAssistant.model is also a ModelRef).

Verified API facts (beta-18999 typings)

• @opencode-ai/schema/dist/model.d.ts: Model.Ref = { id: Model.ID, providerID: Provider.ID, variant?: Model.VariantID } — no modelID key.
• Model.Ref.parse("provider/model#high") → { providerID, id, variant: "high" } — the variant lives in its own key; id is never variant-contaminated.
• @opencode-ai/client SessionMessageAssistant.model: ModelRef = { id, providerID, variant? }.

Variant decision: accept as-is

Because Model.Ref.id is variant-free by construction, model: frontmatter matching needs no variant handling: rules match on the bare model id regardless of the selected reasoning variant, exactly as intended. No parse/strip logic added; a regression test pins the bare-id capture from a variant-bearing ref.

Left alone (verified not matching sources)

• Flat modelID fields in message info fixtures: v2 history messages persist flat modelID (v1 session schema), and withModelObject in the codec is presentation-only (decorates injected transient messages), never read for matching.

Verification: aube run lint, aubx tsc --noEmit, aube run test:run — 548 tests / 35 files pass.