[calvin-archastro]

Review on ArchCode

Problem and author intent

Two scheduling problems seen in firstlanding CI.

1. A pnpm-workspace.yaml edit ran every Node project in full.
   • The lockfile rule in src/related/engine.rs marked every nodejs project under the file's directory with "pnpm-workspace.yaml changed; running in full".
   • That includes projects the workspace excludes, which the file can't affect.
   • pnpm-lock.yaml edits were already narrowed to the importers they change; the workspace file was not.
   • Example: firstlanding run 42439191201 only added one entry to packages:, yet it selected 112 dependents, including the 10-minute Rust e2e lane.
2. Execution admission was a race.
   • Every ready target polled the slot locks every 20 ms, and the first poll after a slot freed won.
   • So a target gating a long chain could wait behind targets that gate nothing.

What changed

1. pnpm-workspace rule (src/related/engine.rs, src/git/affected.rs, src/plugins/nodejs.rs):
   • The root project is always reached.
   • Only the packages: list changed → only packages whose membership changed are reached. The lockfile rule judges their importers.
   • Any other setting changed (overrides, patches, catalogs, linker, hoisting) → reached:
     • every package listed on either side;
     • unlisted projects that resolve from the root node_modules (no pnpm-workspace.yaml, pnpm-lock.yaml or node_modules of their own).
   • An unreadable packages: list reaches everything, as before.
   • Comments are stripped only where YAML does (# after whitespace, outside quotes), so a # inside a value can't hide an edit.
2. Admission order (src/executor/admission_order.rs, admission.rs):
   • Before: whichever waiting target's poll came first after a slot freed won.
   • After: a per-invocation gate orders waiting targets by the longest chain of dependents they gate, then by request order. ReadyQueue also releases longest chain first.
   • A target blocked on an exclusive resource steps out of the line, so it can't hold up the others or deadlock with the resource's holder.
   • Other invocations on the host still compete through the OS locks.
   • Unix-only pieces are behind #[cfg(unix)].

Scope and risk

Planner and executor. Medium risk.

• The workspace rule now selects fewer projects. The settings branch keeps every package that could resolve differently, and an unparseable file still means everything.
• Admission changes start order, not concurrency.

User impact

Fewer targets for packages:-only workspace edits, and chain-gating targets start earlier. No CLI or config change.

Testing

• Canonical end-to-end proofs: both drive the real aster binary.
  • a_pnpm_workspace_edit_runs_only_the_packages_it_can_reach in tests/related.rs covers packages:-only edits, settings edits, unlisted nested projects and # in values.
  • the_target_gating_the_longest_chain_takes_a_slot_before_independent_targets in tests/execution_capacity.rs runs 4 slots and 9 ready targets, with z-gate gating a chain. Against the old binary it failed 6 of 6 times, with a different random admitted set each time.
• firstlanding, run 42439191201's diff, --related --dry-run:
  • "pnpm-workspace.yaml changed" lines: 100 → 0;
  • dependents: 112 → 20;
  • the e2e-rust lane is no longer selected.
• Simulated clock over the real ReadyQueue (12 independent 5 s targets, a 1 s → 10 s → 10 s chain, 4 slots): 36 s → 25 s.
• The CONTRIBUTING checks pass: fmt, clippy -D warnings, cargo doc -D warnings, and cargo test --all-targets --all-features.
• Not proven here:
  • The Windows build was not compiled locally (the C headers it needs are missing); it rests on the cfg audit and this PR's CI.
  • CI effect in firstlanding needs a release and a pin bump.
  • Priority counts targets, not seconds, so a chain's last long target still queues in request order. Duration weighting needs per-target run times that Aster doesn't keep.

Follow-ups

Merge order: independent of the plan replay PR. Both touch src/git/affected.rs and tests/related.rs, so whichever merges second may need a rebase.

🤖 Generated with Claude Code

[calvin-archastro]

ArchDev is watching this pull request. State: settled (cycle 1 of 1 for head b15c81c).

when	trigger	decision	result
2026-11-11T00:52:03.018Z	verify head	enqueue job_feba52fc52d91bc33b7a0320	no change

Paused by the no-autofix label or a draft; never resolves reviewer threads. Jobs: job_feba52fc52d91bc33b7a0320.