What task are you trying to do?

Cut the wall time of `ci / unit` on PRs. It currently spends ~4 minutes per PR run, which adds friction on every iteration.

What do you do today?

Recent run on PR #72 (representative):

• Total job time: 232s
• Checkout + setup + cache + install: 26s
• `bun turbo test:ci`: 186s
• Artifact upload + cache post: 20s

Per-package breakdown of the 186s test block:

`opencode` owns roughly 93% of the test time. `desktop-electron` runs ~1.1s/test on 29 tests, which suggests real I/O or process spawning. Also observed on the same run: the turbo cache step reported `Cache not found for input keys: turbo-Linux-...` with all three restore-key fallbacks missing. A YAML-only PR (where no package source actually changed) should in principle cache-hit every `test:ci` task and finish in seconds, but today it runs the full suite from scratch.

What would a good result look like?

1. Docs / YAML-only PRs finish `ci / unit` in under 45 seconds via turbo cache hits.
2. PRs that only touch a single package run `test:ci` for that package only (plus its dependents), not the whole workspace.
3. Full-cold runs (first touch after cache eviction) drop meaningfully by parallelising `bun test` across cores.
4. Each lever below is verified with a before/after benchmark in the PR description, not "looks right to me".

Which audience does this matter to most?

Maintainers (CI velocity).

Extra context

Optimization levers, ranked by expected win

• P1 — Fix turbo cache hit rate. The cache missed on a YAML-only PR that should have hit. Likely causes to investigate: `passThroughEnv: ["*"]` on every `test:ci` task in `turbo.json` (overly broad env invalidation), the `${{ github.sha }}` suffix in the cache key combined with restore-key prefixes that may not match across workspaces, and missing per-task `inputs` declarations. Verify: run a no-op docs PR against a warmed cache and confirm `test:ci` shows as `FULL TURBO`.
• P2 — Use `turbo run --filter` with affected-packages detection. `bun turbo test:ci --filter=...[origin/dev]` would skip packages unaffected by the current diff. Needs careful review of how turbo computes "affected" for workspaces that import each other. Verify: PR touching only `packages/app/` should not trigger `opencode:test:ci`.
• P3 — Increase `bun test` file-level concurrency. 172.8s / 178 files ≈ 0.97s/file suggests files are running close to serial on the 4-vCPU runner. Try `--concurrency=4` or the equivalent runtime flag. Verify: benchmark opencode tests with and without the flag locally on a 4-core container, confirm no flakiness regression.
• P3 — Sandbox or mock the slow desktop-electron tests. 1.1s/test on 29 tests is high and likely reflects process spawning or real FS access. Identify the worst offenders with `bun test --timeout --verbose` and convert to unit-level mocks where possible.

Out of scope

• Cross-OS caching (Windows is deferred per the Phase 1 macOS-only scope in AGENTS.md).
• Parallel jobs via a test matrix: turbo already parallelises packages within one job; splitting into multiple jobs adds runner cold-start overhead that likely eats the savings.

Sequencing recommendation

Benchmark before any change. Land the turbo cache fix first in its own PR (biggest win, lowest correctness risk). The `--filter` and concurrency changes are each their own PR with before/after numbers.