[Mikola Lysenko (mikolalysenko)]

LLM Description written by Claude Code:claude-opus-5-5

Fixes #1832

Summary

The cargo config chain was walked twice. cargo_build_cache::ConfigDirs::read preferred .cargo/config.toml when both config and config.toml existed. Cargo, and source_replacement::config_layer_paths, prefer the extensionless config. So after apply/rollback the build-cache cleaner removed fingerprints from a target dir cargo never used, and cargo kept linking stale code. This is register row A01 (register; living document Part 3, duplication).

I checked cargo's precedence with real cargo 1.93: warning: both .cargo/config and .cargo/config.toml exist. Using .cargo/config.

Changes

• One walk. ConfigDirs::read now takes its file list from config_layer_paths. I deleted its own ancestor walk, its $CARGO_HOME push and its file pick, keeping only the [build] target-dir/build-dir lookup. Two side effects: $CARGO_HOME is no longer read twice when an ancestor's .cargo is that directory, and the $CARGO_HOME base falls back to the home itself, not "", when it has no parent.
• Relative cwd. /code-review high found this in the touched function; it was already broken on main. --cwd defaults to ., and Path::new(".").ancestors() stops at ., so from a workspace member the cleaner found neither the workspace root nor a parent .cargo/config, and invalidated no build dir. build_dirs_with now walks std::path::absolute(cwd) with .. folded by relpath::normalize_lexically_keeping_escapes. This is the same fold cargo_crawler::resolve_for does. Without the fold, --cwd ../app would walk the directory it left, which Bugbot caught.
• Probe agrees with the read. The file probe asks group_commit::exists first, as read_regular_to_string_sync does. The old code chose the file by attempting the capture-aware read.
• Doc comment. The ConfigDirs comment no longer restates the (wrong) file order; it points at config_layer_paths.

Tests

• build_dirs_read_the_legacy_config_when_both_exist: with both files present, the build dirs are exactly [legacy]. Red on the old pick: left: [".../ws/modern"] right: [".../ws/legacy"].
• build_dirs_walk_above_a_relative_cwd (#[serial], chdir, as nuget_config::a_relative_start_reads_parent_configs does). "." from a member finds the parent config's out; red without absolute: left: [] right: [".../ws/out"]. "../ws/member" from a sibling ignores the sibling's build-dir; red without the fold: left: [ws/out, left/wrong].
• build_dirs_honours_env_config_and_workspace_root and the source_replacement tests are unchanged and pass.

Size

git diff --stat: 1 file, +78/−16. Production is about +12/−16 and tests +66.

Behavior change

Bug fixes only: the cleaner now targets the directory cargo builds into when both config files exist and when --cwd is relative. No output bytes, JSON shapes or exit codes change.

Review findings not fixed

• Unreadable chosen file drops the layer instead of falling back to the sibling. Cargo itself errors on an unreadable config, so a sibling fallback matches no real build. Not worth the code.
• No $CARGO_HOME-dedup test here. config_layer_paths has its own tests in source_replacement.rs.
• Probe plus read is two syscalls per file. A few stats per apply, and the same probe the crawler uses.

Verification

• cargo clippy --workspace --all-features -- -D warnings is clean.
• cargo test -p socket-patch-core --lib: 6991 passed and 4 failed. All 4 are read-only-permission tests in copy_tree, vlt_heal, pypi_poetry and pypi_requirements that fail only because the sandbox runs as root. They are unrelated to this diff.
• CI was green on the first push (82 checks). On 2c715eb clippy passed, but coverage, e2e-build, lint-ecosystems and dispatch-tests lost their Depot runners twice (no test failed); a later re-run is fully green. This adds no cfg(windows)/cfg(unix) code.

Risk

Low. It touches one private function in one file, whose callers are apply and rollback via invalidate_project.

🤖 Generated with Claude Code

https://claude.ai/code/session_011f9aeQ57i1EpKYcB1Ks8rf

---

Note

Low Risk
Scoped to private build-directory discovery used by apply/rollback cache invalidation; behavior fixes only, with no auth or data-path changes.

Overview
Fixes build-cache invalidation after apply/rollback so it clears fingerprints in the same target / build-dir cargo actually uses.

ConfigDirs::read no longer walks .cargo config on its own. It reuses config_layer_paths, matching cargo’s precedence (.cargo/config over .cargo/config.toml when both exist). That stops the cleaner from wiping the wrong directory and leaving stale .rlib links.

build_dirs_with now resolves a relative --cwd (e.g. . or ../ws/member) via absolute + lexical normalize before ancestor walks, so workspace roots and parent configs are found when run from a member crate. Config file probing uses the same group_commit::exists check as reads elsewhere.

Two tests lock in legacy-config precedence and relative-cwd discovery; no CLI output or exit-code changes.

^{Reviewed by Cursor Bugbot for commit 2c715eb. Configure here.}

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

BugBot review

---

Generated by Claude Code

[cursor]

Stale Bugbot comment from a previous run.

[Mikola Lysenko (mikolalysenko)]

BugBot review

---

Generated by Claude Code

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 2c715eb. Configure here.}

[Mikola Lysenko (mikolalysenko)]

[agent] CI on 2c715eb failed in coverage (1), coverage (2), clippy and e2e-build (ubuntu-latest), and cancelled lint-ecosystems and dispatch-tests. None of these failures comes from this PR.

• Each failed job's annotation reads "The self-hosted runner lost communication with the server", or, for lint-ecosystems, "exceeded the maximum execution time of 10m0s".
• No failed step ran a test body.
• The same jobs passed on the previous head, 3b882ea, which also passed all 82 checks. 2c715eb adds only a lexical .. fold and a test.
• Locally, clippy with -D warnings and the cargo_build_cache tests pass.

I've re-run the failed jobs once.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

[agent] The re-run lost its Depot runners again. coverage (1) was cancelled at "Set up job", and coverage (2), e2e-build (ubuntu-latest), lint-ecosystems and dispatch-tests all report "self-hosted runner lost communication". clippy passed. No test failed in any run of 2c715eb. I've used up my one re-run, so a maintainer or the burn-down routine needs to re-run the failed jobs once the runners are healthy.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

Burn-down agent: labeled Ready for review at 2c715eb185b6868047164544be5c81988aec1934.

• CI: 31/31 run checks green (19 skipped). The two earlier attempts lost their Depot runners before any step ran (no test failed); attempt 3 passed cleanly.
• Bugbot: reviewed 2c715eb, no findings; no open threads.
• Reviewer note: single-file refactor of crates/socket-patch-core/src/utils/cargo_build_cache.rs so the cargo config chain is read once (Cargo build-cache invalidation reads .cargo/config.toml where cargo reads .cargo/config, so it cleans the wrong target dir #1832).

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

Final review brief

What it does: Fixes #1832. ConfigDirs::read in cargo_build_cache had its own copy of the cargo-config walk and file pick, and that copy preferred config.toml over config. It now reuses source_replacement::config_layer_paths, which prefers the extensionless config the way cargo does. A relative --cwd is now made absolute and .. is folded out before the ancestor walk.

Risk: low. It is one private function, reached only from invalidate_project (apply and rollback). The worst case of a wrong pick is an extra recompile. The $CARGO_HOME fallback is also slightly more robust: the home is read once, and a home with no parent is handled.

Look here:

• cargo_build_cache.rs:64-66:`` absolute path plus .. fold
• cargo_build_cache.rs:158-163:``` the shared walk
• source_replacement.rs:113-141:```` the single file pick
• cargo_build_cache.rs:429 and :456:````` the two new tests

Verified: I read the full diff against #1832 and checked the precedence against cargo. The fold keeps root and drive prefixes, and on Windows std::path::absolute already resolves ... CHANGELOG.md is untouched. CI is green on 2c715eb (31 passed, 19 skipped, ci-ok and clippy both green). Bugbot found no new issues.

Changes I made: none.

Open questions: none. One nit: the chdir test restores the working directory without a drop guard, though it does restore it before asserting.

Auto-merge (squash) is armed, so approving sends this straight to the merge queue.

---

Generated by Claude Code

[Wenxin Jiang (Wenxin-Jiang)]

[agent] Reviewed 2c715eb185b6868047164544be5c81988aec1934. Checked shared Cargo config precedence, path bases, and relative-cwd ancestor traversal. No major correctness issues found. Validation: source and regression-test review plus the available exact-head CI checks; no local test pass is claimed.