2 Commits

Author	SHA1	Message	Date
Luke Tornquist	d34c9c2c9b	ci: shorten the PR critical path (desktop lanes, smoke shards, parallel relay artifacts) (#8156) ... ## Summary The PR critical path on main push [run 37635224099](https://github.com/block/buzz/actions/runs/37635224099) (`e097c4299`) was about 42 minutes: - **Desktop Core** took 41 min. It ran frontend checks, two clippy graphs, the Tauri suite, the mesh-llm suite and every compiled-flag state one after another. The compiled-flag step alone took 16m51. - **Relay artifact producer → Desktop E2E Integration** took about 31 min. The producer built the relay binaries (5m07), the backend integration archive (4m41) and the PostgreSQL archive (8m18) one after another. - **Desktop Smoke E2E** took 19–24 min per shard across 4 shards, with 82 min of tests in total. This PR attacks all three. ### 1. Run each compiled-flag state once (`Justfile`) | Previous step | Cost | Why it's redundant | |---|---|---| | Clean state: full `buzz_lib` suite (`--lib`) | 3m27 run | `desktop-tauri-test` runs the same suite in the same compile state one step earlier. When `BUZZ_TEST_EXPECTED_AGENT_ACCESS_OWNER_ONLY` is unset, the expected value already defaults to `false` in non-owner-only builds (`runtime/test_fixtures.rs`, `agents_tests.rs`). | | Internal state: auto-connect and owner-only built separately | 1m22 + 1m06 + 56s across three recompiles | Internal release packaging always sets both capabilities together, so building owner-only without auto-connect tests a configuration that never ships. The extra 56s happens because a bare `cargo test <filter>` builds every target, not just the `--lib` harness. | | Demo state: `cargo test --workspace` | ~30s more than `--lib` | `buzz-terminal` and the `tests/*.rs` integration tests don't read any `BUZZ_BUILD_*` variable or `buzz_lib`, so they ran the same binaries a second time. | A fourth cost was hidden: every recipe invocation spent ~1m20 recompiling `buzz-desktop` to the *same* unit hash `desktop-tauri-test` had just built. `_ensure-sidecar-stubs` ran `touch` on every sidecar stub on each run. `tauri-build` emits `rerun-if-changed` for each `externalBin`, so the new mtimes reran `build.rs` and recompiled the crate. Stubs are now created only when they're missing. This also speeds up any local `desktop-tauri-*` recipe run back-to-back. After this change: - **Clean state:** runs only the two compiled-flag assertions (`compiled_flag_matches_expected`, `compiled_policy_matches_expected`) with `expected=false`. No recompile. - **Internal state:** one compile with `BUZZ_BUILD_AUTO_CONNECT_DEFAULT_RELAY=1` and `BUZZ_BUILD_AGENT_ACCESS_OWNER_ONLY=1`. It runs both assertions with `expected=true`, then the full `buzz_lib` suite. The suite now runs with auto-connect enabled as well, which matches the shipped internal build more closely than before. - **Auto-connect-only state (assertions only):** auto-connect `true`, owner-only `false`. The two capabilities are documented as independent. Without this state, a `build.rs` block that read the other capability's variable would pass, because every other state sets both or neither. - **Demo state:** one compile with only `BUZZ_BUILD_DEMO_SLUG`. It asserts the slug and asserts both capabilities stay `false`, then runs the full `buzz_lib` suite. - The recipe clears the three toggled `BUZZ_BUILD_*` capability variables and any caller `BUZZ_TEST_EXPECTED_*` values up front. Each state exports its own values in a subshell. Every invocation uses `--lib`, so each state compiles one test harness. - Each ignored-assertion run requires exactly one passing test per name filter. Previously, a renamed or un-ignored assertion matched nothing and passed silently. - The 32-character demo-name rejection check is unchanged. The full-suite reruns in the internal and demo states are kept on purpose. They are the only coverage that the whole suite holds up in those two shipped build shapes. Estimate: the compiled-flag work goes from about 17 min to about 11. ### 2. Split Desktop Core into parallel, path-gated lanes (`_ci-desktop.yml`) - **Desktop JS:** `desktop-check`, `desktop-test` and `desktop-build`. These compile no Rust, but some node:test cases read relay and DB crate sources to pin shared constants (for example `relayReconnectReplay.test.mjs`). So the job runs on desktop *and* Rust changes, in parallel with the Tauri lanes. `lefthook.yml` notes this as a deliberate local deviation. - **Desktop Tauri (core | mesh | compiled-flags):** a matrix of parallel Rust lanes. - `core` and `compiled-flags` share the existing `desktop-core` Rust cache key, so they start warm. `compiled-flags` restores it read-only. - `mesh` keeps its own cache because `mesh-llm` pulls a different dependency graph. - **Smoke E2E and the Windows frontend build** compile no Rust and read no crate sources, so they no longer run for Rust-only changes. - **Smoke E2E goes from 4 to 10 shards.** Playwright shards by test count, not duration. A simulation from the per-test durations in run 37635224099 reproduces the observed 4-shard split exactly. With 10 shards, the slowest shard runs about 11 min of tests. - **The `Desktop` gate is stricter.** Each selected job must succeed and each unselected job must be skipped, so a mis-gated job can't pass silently. The required check context (`Desktop`) is unchanged. `scripts/ci-required-gates.test.mjs` pins each job's `if:` to the gate's selection expression, and checks which jobs a Rust-only, desktop-only, Tauri, push, and empty selection pick. - **Path-filter fixes:** - The `desktop` filter now includes `_ci-desktop.yml`, `ci.yml` and `Justfile` (which holds the desktop recipes), so edits to them run the frontend lanes. - The `rust` filter now matches `Justfile`. It still named the pre-rename `justfile`, so a `Justfile`-only change selected no Rust suites. - `scripts/ci-selection.test.mjs` gains `Justfile` and desktop-workflow scenarios, and the shared-workflow scenario now expects `desktop`. - **Dead step removed:** the failure-only Playwright artifact upload in the old Core job never had anything to upload, because that job never ran Playwright. ### 3. Build relay artifacts in parallel (`_ci-relay.yml`) - **The producer is now a 3-way matrix** (`relay-binaries`, `backend-integration-tests`, `postgres-tests`). - Each lane has its own exact-source cache entry and uploads `desktop-e2e-relay-<artifact>`. - Consumers download `desktop-e2e-relay-*` into `target/ci`, so their paths are unchanged. - The required `Desktop E2E Relay` context still reads the producer job's aggregate result. - **Rust cache saving:** all lanes restore the existing Rust cache key, but only the widest lane (`postgres-tests`) saves it. Otherwise the fastest lane could overwrite it with a narrower target dir. - **Desktop E2E Integration goes from 2 to 3 shards.** Shard labels now use `strategy.job-total`. ### 4. Bound degraded apt steps (`scripts/ci-apt-retry.sh`) - With more shards, more runs hit the occasional degraded Ubuntu mirror. On this PR, a Tauri lane ran `apt-get install` for 27 minutes: apt's `Acquire` timeouts kept firing, but apt kept re-queuing the requests and never failed. Separately, three Smoke shards stalled silently in `playwright install-deps` (no output after `InRelease`) until the job timeout. Healthy apt steps across four recent runs finished in at most 238s. - Every apt-backed step in Desktop and Integration now runs through `scripts/ci-apt-retry.sh`. The script writes the apt options to `apt.conf.d`, so Playwright's own apt-get calls get them too. It kills each attempt after 7 minutes (90s for `apt-get update`). Between attempts it stops any `apt-get` or `dpkg` the attempt left running, repairs dpkg under its own 60s deadline, and retries up to 3 times. The stop matters for `playwright install-deps`, which runs `sudo sh -c "apt-get update && apt-get install ..."`: sudo puts its command in a new process group, so the deadline's signal stops `sh` and apt-get survives holding the dpkg lock (seen on Smoke shard 8 of run 37696815557, and reproduced with real sudo). - Downloaded `.deb`s are kept between attempts, so a retry resumes where the last one stopped. Against a mirror that stays degraded, this bounds each wrapped command at about 25 minutes rather than recovering it. The Tauri lanes wrap two commands, `apt-get update` (about 8.5 minutes worst case) and `apt-get install` (about 25), so about 34 minutes together. Job timeouts are sized to fit that bound plus the job's own work: Tauri 55m (lanes take up to ~15m), Integration 40m, and Smoke 45m. A degraded mirror therefore ends with the wrapper's error instead of a silent job timeout. - **The retry has already fired on this PR.** In [run 37703795342](https://github.com/block/buzz/actions/runs/37703795342), `playwright install-deps` on Integration shard 1/3 hit the 420s attempt deadline. The second attempt started about 1.4s later, downloaded and configured the real packages, and succeeded. The step took 526s (8m46), and that shard took 17m24 while the other two took about 5m. Healthy installs have taken at most 238s, so 420s leaves headroom, and a slow mirror costs one attempt instead of the job. - Between attempts the script stops **every** `apt-get`/`dpkg` on the machine, not only the ones it started. That's intentional: on a hosted runner any other one would hold the same dpkg lock. It assumes an ephemeral runner, and the script says not to use it on a self-hosted runner or a developer machine. - `scripts/ci-apt-retry.test.mjs` covers the following with a fake `sudo`: success, retry, the attempt budget, a stalled attempt, a stalled dpkg repair, an apt-get that escapes the attempt's process group (on a retry and on the final attempt), and usage. It needs Linux (GNU `timeout`, `flock`, `setsid`, procps), so it skips on other platforms. CI runs it on Ubuntu. ### What got simpler - **Compiled-flag recipe:** two full-suite runs instead of three, six cargo invocations instead of eight, and sidecar stubs no longer invalidate the desktop build cache. - **Path selection:** each desktop job declares only the paths that actually affect it. The stale `justfile` filter and a dead artifact step are gone. - **Apt options:** set once in `apt.conf.d` by one script, instead of as repeated `-o Acquire...` flags on each apt call. - **Shard counts:** the shard count is written once per matrix, not repeated in names and commands. ### Cost - **More runners:** a full run uses more runner jobs. Smoke E2E goes from 4 to 10 jobs, Desktop Core from 1 to 4, the relay producer from 1 to 3, and Integration from 2 to 3. - **Fewer runners on single-area PRs:** desktop-only PRs skip the Rust lanes, and Rust-only PRs skip Smoke E2E and the Windows build. - **One extra Rust cache entry:** `desktop-mesh`. ### Related issue None found. ### Testing Blox, compiled-flag recipe at `77b9874b`: - [x] Recipe passes with output redirected to a file. The log keeps all 4 state headers, the ignored assertions ran 2/2/2/3, and both full suites passed (`3423 passed`). - [x] A second back-to-back run shows no clean-state `Compiling buzz-desktop`. - [x] Mutation tests all fail as required: - `build.rs` auto-connect reads the owner-only var: rc=101. - The reverse swap: rc=101. - Rename `compiled_policy_matches_expected`: rc=1. - Wrong owner-only expectation: rc=101. - [ ] `just ci` Workflows: - [x] `actionlint` on `_ci-desktop.yml`, `_ci-relay.yml` and `ci.yml`: clean. - [x] `scripts/test-ci-required-context-isolation.sh`: passes. - [x] GitHub Actions: runs [37701003713](https://github.com/block/buzz/actions/runs/37701003713) and [37703795342](https://github.com/block/buzz/actions/runs/37703795342) took 27m09 and 28m51. Before this PR, PR runs had a median of 40.0m (n=21) and main pushes took about 42m. These are observations across mixed runner images, not a measured causal speedup. - [x] Selection, gate and apt suites (74/74 on Linux), `actionlint` and `bash -n` on Blox at tree `db9725fe`. --------- Signed-off-by: Elrond <28d6302a099e5225b02c4155ac4236e4912603df2ab08dbfc2f4fef08ce598c8@buzz.block.builderlab.xyz> Co-authored-by: Elrond <28d6302a099e5225b02c4155ac4236e4912603df2ab08dbfc2f4fef08ce598c8@buzz.block.builderlab.xyz>	2026-10-08 11:33:37 -04:00
Tom Brow	02c6309f53	fix(ci): select runtime suites from PR changes only (#7843) ... A documentation-only update to #7809 selected desktop builds and E2E tests because path detection compared an old PR base SHA with GitHub's newer synthetic merge commit. The [failing run's path-detection log](https://github.com/block/buzz/actions/runs/35789957773/job/106955810630) includes four unrelated desktop files from `main`; `VISION_MOBILE.md` did not match a runtime filter. Use GitHub's PR file list for pull requests so unrelated changes in the synthetic merge commit cannot select runtime suites. The existing directory filters are unchanged: root and docs/ Markdown do not select runtime suites, while Markdown under runtime directories still selects its affected suites. Mixed code/documentation changes retain normal coverage. Always-on security, policy, and source-contract checks and full push-to-main coverage are unchanged. Added regression coverage runs the pinned paths-filter action against real fixture repositories, including a synthetic merge containing unrelated upstream desktop code. All 35 selection scenarios and 26 required-check scenarios pass; the new regressions failed before the fix. Existing required-context isolation, file-size policy, and security-review contract checks pass, as do script lint and workflow syntax validation. Full workflow lint reports the same two pre-existing shell-quoting findings in the untouched dead-token guard. A desktop E2E build was run to diagnose the unrelated required smoke failure. Related: #7809 (incident, unchanged) and #5756 (shared-input path coverage, separate scope). No matching issue found. PRs with at least 3,000 changed files now select every runtime suite, avoiding GitHub's PR-file-list ceiling. Boundary tests cover 2,999, 3,000, and 3,001 files, including a runtime file omitted beyond the API cap; disabling the safeguard makes the latter two regressions fail. Required checks now run and fail when path selection fails, is cancelled, or is skipped. Regression coverage exercises the workflow conditions and shell checks for all 13 required wrappers, plus an API-denial case against the pinned action. Mutating either the scheduling guard or the result check makes all 13 failure regressions fail. The required mention-settings smoke test expected a pin after explicitly disabling automatic mentions. Waiting for its old avatar to exit reproduced the CI failure 3/3; the test now checks that subsequent mentions remain manual and retains outgoing recipient-tag assertions. Corrected browser regression: 20/20 repeated runs; related picker unit tests: 18/18. No desktop production behavior changed. The always() requirement is now pinned in required-wrapper regression tests; deleting it fails all 26 gate cases. --------- Signed-off-by: Tom Brow <tomb@block.xyz>	2026-09-25 23:19:32 +00:00