docs(dev): the plan records the nextest answers and that the sponsor texts are done

N1 is answered: macos-nextest runs in nix on a Mac (383 tests on the Mac Mini), the Windows
unit tests run in the Windows guest through nix run .#windows-nextest (964 tests, 964
passed, one run) with the check windows-nextest-build cross-compiling the 48 test
executables, and the neutral crates' tests run in linux-nextest (1532 tests). The remaining
open point is only whether the first CI also gets a hosted windows-latest cargo test job.
The R6 follow-up (the sponsor-only [FEAT] texts) and 2.2a are marked done.

Verified: documentation only; the numbers are from the agents' reports (win-nextest, and the
Mac Mini run of macos-nextest).
This commit is contained in:
LGUG2Z committed 2026-11-05 15:03:27 -08:00
1 parent 5d85e3f719
commit 598b57ff9b
1 file changed
+6 -6
+6 -6
View File
@@ -35,7 +35,7 @@ the first version where they differ):**
Sections 4 and 5 and the checklists are split accordingly: **DEV-LINE LANDING** (steps L0 to L6) and **RELEASE** (steps
5. **R1 refined (later message): "yes, but I will push myself".** Publishing the KDE and macOS history is approved. **The owner performs every push and tag command**;
agents prepare them (5.3) and never run them. Secret scan: see the fact row in section 0 (gitleaks, default rules, 2832 commits, no leaks; its limits are listed there).
6. **R6 follow-up: "you can update these".** An agent, `sponsor-texts`, is rewriting the sponsor-only `[FEAT]` texts of 2.2a in its own workspace: **in progress**; this plan does not touch those files.
6. **R6 follow-up: "you can update these".** An agent, `sponsor-texts`, rewrote the sponsor-only `[FEAT]` texts of 2.2a: **done** (revisions on the dev line, the owner answered its three OWNER notes: sponsors get what the README lists, no private macOS repository, the macOS nightly survives from the joined repository).
7. **R8: "master is already frozen"** (done by the owner): the "freeze" step is removed; merging the ready PRs, the path-map comments and closing the 8 dependabot PRs stay as steps (L1).
8. **R9: "I will self host but this will come later".** The first CI is hosted Linux plus hosted macOS only (the cheap check set `checkSets.cheap` is the PR gate). The Windows e2e, the KDE VM tiers and the canary are a **later phase** on a self-hosted KVM runner, nightly or manual, with the evaluation-ISO caveat (3.3, A.2).
9. **R10: "PR only is good"**: the `hotfix/*` and `feature/*` push triggers are dropped in the drafts. **R12: "GitHub cache sounds good"**: `actions/cache` with per-platform keys, no Cachix for now.
@@ -193,7 +193,7 @@ The `[FEAT]:` sponsor-only auto-close policy (`.github/workflows/feature-check.y
#### 2.2a R6 task list: the sponsor-only `[FEAT]` policy for one public project with no private tier
**Status: in progress.** The owner said "you can update these"; the agent `sponsor-texts` is doing the rewrite in its own workspace. This plan does not edit those files; when its revisions land, check the list below against them and tick it off.
**Status: done.** The owner said "you can update these"; the agent `sponsor-texts` rewrote the texts (its revisions are on the dev line); check the list below against them when reviewing.
Found by `git grep -i sponsor` on the tip (`59b01fdb`, plus the checks of `2185556efba6`):
1. `.github/workflows/feature-check.yaml:22` (message text): names "current GitHub sponsors on the $5/month tier and above", "valid individual commercial use licence" and "approved contributors". Rewrite so it does not imply a private repository or early access, says it is for all platforms (Windows, Linux, macOS), and links the one licence page. Keep the `[FEAT]: ` title prefix check and the close-and-lock behaviour.
@@ -276,7 +276,7 @@ Principles: (1) **required check = one job, `required`, that depends on the othe
| `deny` | `ubuntu-24.04` | | PR, push `master`, schedule | `EmbarkStudios/cargo-deny-action@v2` with the one root `deny.toml` (the cutover merged them) |
| `cheap` (matrix over the sets `fmt deny windows-lint windows-test linux-lint linux-test linux-options`) | `ubuntu-24.04` | | PR, push `master` | `nix build .#checkSets.<set>`: the cheap check set is the PR gate (`checkSets.cheap` = `defaultCheckSets` in `nix/checks.nix`; docs/dev/flake-checks.md). Windows clippy is the 3.5-minute pre-landing gate with warm dependencies (docs/dev/build-and-test.md section 1) |
| `windows-cross` | `ubuntu-24.04` | | PR, push | `nix build .#windows-komorebi .#windows-komorebic .#windows-komorebi-bar` (cross-built with xwin). The hosted `windows-latest` cargo job of the old workflow (fmt, clippy, `cargo test`, MSI) is NOT in the first CI (R9); the MSI moves to the release workflow (A.3). What that leaves untested is decision N1 |
| `macos` | `macos-14` (hosted arm64) | | PR, push | `nix build .#checkSets.macos-{lint,test,options}` and `.#komorebi .#komorebic .#komorebi-bar`: build, clippy and the darwin nix checks. `macos-nextest` runs no test today (N1). **No daemon is started**: nothing in CI can verify tiling on macOS. A hosted macOS runner can compile and build: **(inferred, never tried here)** |
| `macos` | `macos-14` (hosted arm64) | | PR, push | `nix build .#checkSets.macos-{lint,test,options}` and `.#komorebi .#komorebic .#komorebi-bar`: build, clippy and the darwin nix checks. `macos-nextest` runs 383 tests on the hosted macOS job (N1). **No daemon is started**: nothing in CI can verify tiling on macOS. A hosted macOS runner can compile and build: **(inferred, never tried here)** |
| `linux-build` | `ubuntu-24.04`, `ubuntu-24.04-arm` (optional) | | PR, push | the old KDE `build` job with `.#komorebi` and `.#komorebic`, patchelf to the system loader, `--version` |
| `vm-fast` (16 suites), `vm-slow` (18 suites) | **self-hosted with KVM (later phase)** | `discover` | nightly, manual | `legacyPackages.<system>.vm.<name>`; tiers from `legacyPackages.<system>.vmTiers`; the old design (one hosted runner per suite, 45/60 min timeouts, `rerun-no-kvm`) is the fallback if hosted KVM is used instead. KVM on hosted runners is documented for public repositories only (docs/dev/old-ci/linux/actions/nix/action.yaml:20-21); some hosts have none |
| `e2e-windows` | **self-hosted (later phase)**, label `kvm-windows-e2e` | | manual, nightly subset, weekly full, before a release | see below: needs KVM, the evaluation ISO and a locally built golden image (README:9-26); cannot run on a hosted runner **(inferred)** |
@@ -295,7 +295,7 @@ Known failures (docs/dev/build-and-test.md section 6; they do not block the dev-
**Caching (decided, R12).** `actions/cache` (through `nix-community/cache-nix-action`, as the old KDE action did) with per-platform key prefixes (`nix-linux-<arch>-`, `nix-macos-<arch>-`, keyed on `flake.lock` and `**/Cargo.lock`, `gc-max-store-size-linux: 8G`), and `Swatinem/rust-cache@v2` for any cargo job. No Cachix for now. No Cachix was ever used by the old repositories (action.yaml comment). One repository now carries a Linux and a Darwin closure and the GitHub per-repository cache quota is small **(inferred: 10 GB for GitHub Actions caches; check current limits)**: the per-platform prefixes keep one platform from evicting the other; measure the hit rate on the first runs and revisit only if PR time exceeds 30 minutes.
**N1. The nextest decision: IN PROGRESS, answers to be recorded (not decided here).** `windows-nextest` was removed from the flake because it ran no test (the Windows crates' tests cannot run on a Linux builder; docs/dev/flake.md "Old name -> new name", flake-checks.md `windows-test`), and `macos-nextest` runs none either (as told by the coordinator; not re-read in `nix/macos/`). With the first CI hosted-Linux/macOS only, the Rust unit tests of the Windows daemon crates and the macOS crates are then run nowhere in CI (the Linux ones run in `linux-nextest`; the shared `crates/` run natively on Linux). The owner asked two feasibility questions that engineers are answering now: (1) can macOS nextest run through nix on a Mac with nix-darwin; (2) can the Windows unit tests run in a Windows VM from NixOS. Record the answers here when they arrive; until then the first CI is as drafted (no `cargo test` jobs for Windows or macOS), and the old hosted options (the `windows-latest` job of the old workflow running `cargo test`, a hosted macOS `cargo test --locked -p komorebi-macos -p komorebic -p komorebi-client -p komorebi-bar-core -p komorebi-bar`, 665 passed on the Mac Mini at candidate 8) stay available as fallbacks.
**N1. The nextest decision: ANSWERED (2026-11-01).** The owner asked: can nextest run on a Mac with nix-darwin, and can the Windows unit tests run in a Windows VM from NixOS; if so, keep them. (1) **macOS: yes.** `macos-nextest` now has `doCheck = true` (`nix/macos/flake-module.nix`); built on the Mac Mini (aarch64-darwin, nix, no daemon): exit 0 in about a minute, 383 tests run, 383 passed, 1 skipped (the komorebi-macos library; the other packages' tests ran earlier through cargo, 753 in all). It needs a darwin builder, so the first CI runs it on the hosted macOS job (`checkSets.macos-test`). (2) **Windows: yes, in the Windows guest of the e2e harness, not as a pure nix check** (a Windows image cannot be in the store). The check `windows-nextest-build` (set `windows-test`) cross-compiles the 48 Windows test executables with xwin on a Linux builder (about 90 s warm); the app `windows-nextest` (`nix run .#windows-nextest -- OUT`, with `E2E_GOLDEN`, `KOMOREBI_E2E_SCALE`, `KOMOREBI_E2E_STATE` as for e2e-windows) uploads them with their data files to the guest and runs every test in its own process: 964 tests in 37 executables, 964 passed, 0 failed, 202 s with a restored guest state (one run, not soaked). It needs the same self-hosted KVM machine as the Windows e2e (later phase, R9). Also: the neutral crates (`komorebi-layouts`, `-scrolling`, `-themes`, `komorebic-core` and the shared `crates/*`) now run in `linux-nextest` (1532 tests, all pass), so they no longer depend on the Windows CI. What remains untested in the FIRST CI: the Windows-only crates' tests (they run in the self-hosted phase or in a hosted `windows-latest` `cargo test` job, the old fallback) and the macOS-only code at runtime.
**Cost and time.** Public repositories do not bill hosted Linux and Windows minutes **(inferred, GitHub's public-repo policy; macOS hosted runners are also free for public repos at the time of writing, check)**. The private KDE repository did bill. I found no billing data and do not give a cost figure. Times that were read: Windows workflow 22-30 minutes; KDE `vm-fast` timeout 45 min per suite in parallel; the PR critical path is therefore about the slowest of: Windows build 22-30 min, a VM suite up to 45 min, nix checks (Windows clippy 3.5 min warm, cold is **(unknown)**; the KDE shard run out of resources at about six minutes before it was sharded, ci.yaml comment lines 41-43).
@@ -392,7 +392,7 @@ Steps (owner): 1. Starlight has the guides of all three platforms (Windows from
| # | Step | Reversible? | Gate before it (go/no-go) |
|---|---|---|---|
| L0 | The macOS token is not rotated (owner decision); N1 (nextest) is in progress and need not be answered to land; the R6 text rewrites (2.2a, agent `sponsor-texts`, in progress) land before CI goes on | yes | the user's answers |
| L0 | The macOS token is not rotated (owner decision); N1 (nextest) is answered (3.3); the R6 text rewrites (2.2a) are done | yes | the user's answers |
| L1 | `master` is already frozen (done by the owner). Merge the ready PRs, post the path-map comment on the rest, close the 8 dependabot PRs; record the `master` sha | yes | `master` sha unchanged since the freeze |
| L2 | Scratch-repo dry run (5.4) of push and CI (the release part is rehearsed later, E0) | yes (delete the scratch repository) | CI files drafted |
| L3 | Push `komomono` (fast-forward) | **partly**: the private KDE and macOS histories become public | the owner confirms the push itself (decision 1) |
@@ -483,7 +483,7 @@ Decided by the owner (kept for the record): **R1** no private sponsors repositor
1. **The push itself (owner).** The owner runs L3 (`komomono`) and L4 (`master`), then L5 (ruleset, `legacy/windows-0.1`), using the commands of 5.3; confirm when.
2. **The macOS hand-check (owner; gates only the 0.2.0 RELEASE).** About 5 hours on a real Mac (`docs/dev/macos-hand-check-plan.md`); macOS is on the dev line meanwhile; the owner releases 0.2.0 when it is ready.
3. **The self-hosted Windows runner (later).** KVM, the evaluation ISO (expires every 90 days), then the Windows e2e, KDE VM tiers and canary as nightly/manual jobs (A.2).
4. **N1, the nextest decision: in progress, answers to be recorded.** The owner asked two feasibility questions that engineers are answering now: macOS nextest via nix on a Mac with nix-darwin, and Windows unit tests in a Windows VM from NixOS. Not decided here (3.3). Record the answers in 3.3 when they arrive.
4. **N1, the nextest decision: answered** (3.3): macOS nextest runs in nix on a Mac (383 tests); the Windows unit tests run in the Windows VM through `nix run .#windows-nextest` (964 tests, self-hosted phase); the neutral crates' tests run in `linux-nextest` (1532). Nothing is left to decide except whether to add a hosted `windows-latest` `cargo test` job to the first CI (not needed for the dev-line landing).
5. **Release-time items** (decided at the release, not now): re-open the known failures (R4: the four stale e2e markers, `fullscreen`, the lock-hold stall) and whether they block; the release notes (4.5); the full gates (5.2); the macOS nightly workflow and formula (4.8); retiring `gh-pages` (4.9); the old KDE repository archive (2.5).
What could not be determined: active user counts, the macOS repo's 13 "open issues", the cost of hosted minutes, the hosted-runner wall time of the KDE suites, why the KDE `ci` push runs fail in 7 to 10 seconds, whether Chocolatey or other third-party packages exist, the role of `PAT` and `SCOOP_TOKEN`, DNS for `komorebi.lgug2z.com` and `komorebi-for-mac.lgug2z.com`, the real names of the flake outputs (placeholders until `flake-one` lands), whether the 19 community PRs merge cleanly after the renames, the installed jj version.