Implementation Plan: Harness event normalization

Branch: t3/harness-integration-fixes · Date: 2026-10-08 · Spec: spec.md

Summary

Repair OpenCode provider event normalization in the shared installer shim. Retain the existing pure adapter and IPC contracts. Execute installed source through both public plugin entrypoints with an isolated fake reporter.

Technical Context

Rust 2021/minimum 1.92; generated JavaScript; Linux. Existing serde/nix integration library, existing server testkit. Node required for behavioral plugin tests, added to reference setup and doctor. No paid provider requests.

Constitution Check

Spec and clarification review complete, then plan/tasks before implementation. Tests use existing public installer and hook-event seams, red before green. No OS calls in pure adapters; no model or architecture changes, so no new ADR. Accurate attention states follow cmux/herdr; existing native presentation remains.

Structure and sequence

1. Regression test on installed V2 setup → error translation fix.
2. Regression test on installed V1 server → identity normalization fix.
3. Test unknown events, success, unmanaged-pane no-op and unload cancellation.
4. Document contracts and Node test dependency, independent review, full gate.

Risks

Event fixtures are grounded in tagged provider schemas, but do not certify a paid live session. Node tests verify actual generated JavaScript, not source substrings. Reporter timeout remains bounded and user config stays untouched.

Audit refinement: answered approvals

Grok found that Store::answer_decision resumes a task but leaves a blocked pane blocked. Add a Blocked → Working transition there, shared by native and TypeText answer paths; retain stale-answer and non-answer clear behavior. Test first at Store and native reporter IPC seams. Stabilize the existing GTK focus test by waiting for mapping before its focus assertion with the shared bounded wait helper.

Display validation refinement: the host Xvfb default is 640×480. Explicitly set 1920×1080×24 for GTK scenes and refresh benchmark so desktop-width assertions remain meaningful across distributions; no production GUI behavior changes.