Turbo / Nx test-suite survey (2026-09-10)

The owner’s ask: “go through all tests of nx and turbo, see if we miss any test or any core functionality, e.g. prevention of circular task deps.” Two survey passes over fresh sparse clones (Turborepo at crates/turborepo/tests, turborepo-lib, turborepo-cache, turborepo-scm; Nx at packages/nx/src, task-runner / hasher / project-graph / command-line / native), every upstream test classified against vx’s suites with grep evidence: COVERED, GAP-TEST (behaviour present, unpinned), GAP-CORE (behaviour absent), DIVERGENT (documented choice), REJECTED (owner’s list). This is the record of what came out and what was decided; docs/parity.md stays the user-facing map.

Counts. Turbo: 568 tests over 74 files — ~355 rejected (Cargo/Go/uv workspaces, prune, query, the daemon, TUI, futureFlags, globals, env modes, // tasks…), ~155 covered, ~28 divergent, 10 core gaps, 5 test gaps. Nx: ~380 cases in 146 rows — 78 covered, 41 rejected, 17 divergent, 7 core gaps, 3 test gaps. Circular task dependencies specifically: every cycle shape both suites pin is covered, and vx refuses at graph build rather than breaking edges (NX_IGNORE_CYCLES has no vx spelling, on purpose).

Fixed the same day (STATUS 102–105)

1. A literal directory in cache.inputs.files / outputs.files (src/, dist) matched nothing — the most common turbo.json shape migrated into a config that folded zero inputs and cached an empty artifact. A literal is now the file or its whole tree (Turbo globs_test.rs).
2. A negated workspace package glob (!packages/fixtures) inverted discovery: every manifest in the tree became a member. Negations subtract now, in discovery and the root-claim walk (Nx project-configuration-utils.spec.ts).
3. --affected diffed from the ref, not the merge base of ref and HEAD, so a branch off a moved trunk selected everyone else’s changes (Turbo affected_test.rs, Nx command-line-utils.spec.ts).
4. Path filters take a glob (./packages/*, {apps/**}), matched against the project’s own dir (Turbo infer_pkg_test.rs).
5. A --cache spec naming a remote axis with no remote layer says so (Turbo run_caching.rs).
6. Pins: the cycle message names the path; an unknown pkg#task target is named as such; --concurrency <n>%; pkg#task runs under --filter '!pkg'. Comparison rows: no bare-task cross-product; parallelism: false maps to a whole-budget reservation (then exec.resources; a schedule-plugin reservation since 2026-09-12).

Fixed (STATUS 106, same night)

• SIGINT/SIGTERM: SIGTERM then process.exit with no grace and no SIGKILL escalation, so a child that trapped TERM was orphaned by a CI cancellation (Turbo graceful_shutdown_test.rs). Now: SIGTERM, a VX_KILL_GRACE_MS wait (2 s, the persistent shutdown’s grace), SIGKILL, exit 130/143; a second signal skips the grace.
• A task whose command re-entered vx run forked without bound (Turbo recursive_turbo_test.rs). Now every child carries VX_RUN_WORKSPACE / VX_RUN_TASK and run() refuses its own root, naming the task — the terminating shape too, since a nested run is invisible to the outer graph; a different workspace is still fine.

Deferred, with the reasoning

• Project edges outside package.json (Nx implicitDependencies; an e2e package that tests app without depending on it was invisible to ...[ref]). DECIDED 2026-09-10 (STATUS 113): no new field — a cross-project dependsOn edge (e2e’s dependsOn: ['app#build']) now counts for ... / ^... selection, since the task graph already knows it; every surface agrees without a second declaration. A workspace-level default base stays rejected: --affected=<ref> is one flag and the origin/HEAD fallback covers the rest.
• Project-selector dependsOn (lib*#build, !lib1). vx materialises concrete edges and rejects a pattern in the pkg#task form; task-name patterns (^build.*) cover the common case. Revisit when a real workspace needs it.
• Default base configuration (defaultBase, NX_BASE). vx falls back origin/HEAD → HEAD~1; a workspace-level affectedBase is a small addition when someone asks.
• FORCE_COLOR synthesized for children (Nx defaults it on). vx forwards it only when set; a default would put ANSI into cached logs and every renderer. Explicit over magical.
• Structured log stream (--log-file, --json NDJSON). The telemetry seam carries every record; a sink plugin is the vx spelling. Core will not grow a second stream.
• Richer --dry=json / --summarize (per-file input hashes, command, env). vx show gives the command, vx why the input diff; a pre-execution file listing is a candidate, not a gap.
• watch + persistent divergence (documented, unpinned): a pin needs a watch harness that is not flaky; on the list with M7/M8.
• Scoped vs whole-repo enumeration equivalence: DONE (STATUS 108), tests/enumeration-equivalence.test.ts — twelve seeded trees, every worktree state, exact partitions and trusted OIDs in both modes.