Repository navigation
rust.sh shares one CARGO_TARGET_DIR across worktrees, so a scratch-copy build poisons the next run #906
Description
scripts/rust.sh runs every build in one container with CARGO_TARGET_DIR=/target and the source mounted at /work/nostream. A build of a scratch copy of the tree — the shape a reviewer uses to prove a fix by reverting it — writes into that same target directory, and the next worktree run then reads artifacts keyed to the scratch path.
Measured during the 4.0 rewrite (006-nostream-core, round 3): a revert build followed by the worktree suite gave 51 passed / 2 failed; touch-ing the affected sources and re-running gave 53 passed / 0 failed, with no source change. The failures were stale artifacts, not defects.
Two candidate fixes:
- Isolate the target for a scratch run: pass
CARGO_TARGET_DIR=/target-$(basename "$PWD")(or a caller-suppliedCARGO_TARGET_DIR) instead of a single shared/target. - Keep the shared target and document the requirement in the script header: after building anything outside the job worktree,
touchthe sources before the next worktree run.
(2) is a one-line comment; (1) costs a cold build per distinct path but removes the trap. The reviewer that hit it lost one round-trip and nearly reported a regression that did not exist.
Activity
chappie-daemon commentedon Oct 20, 2026
Fixed by #914, merged as
6ef609a.scripts/rust.shnow derives the target directory from the checkout path —/target-<basename>-<cksum>, mounted from the same key underRUST_CACHE_HOST— so two trees no longer share fingerprints.CARGO_TARGET_DIRstill overrides (absolute only; relative exits 64), andRUST_CACHE_VISIBLEnames the client-side spelling when the cache root is not visible from the calling shell, in which case a daemon-only root is warned about and left to the daemon rather than refused.Measured on a scratch copy with
run_transactionreverted: the old script's worktree ran the scratch binary with noCompilingline and failedop 117 (transaction) differs from the TypeScript capture;touchthen cleared it. With the fix the two runs wrotetarget-scratch-906-562595450andtarget-906-rust-target-53259279and neither read the other. Cost: one cold build per checkout, 36 s forcargo test --workspace --all-features --no-run, 21 s forcargo check -p nostream-storageon a second path.The same pattern still exists outside this repository in
containers/rust-builder/rb-run; filed separately as cameri/phoenix-server#48, since it cannot be fixed here.