GeezerVison

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

Fixes #1521

Summary

Bun 1.3.0 and 1.3.1 honour ! workspace exclusions, but socket-patch read their locks as pre-1.3 Bun. A project excluded from a Bun workspace with ! therefore couldn't be patched from either directory. The hosted scan was refused, the vendored scan failed, and the suggested root run reported success with nothing pinned. This PR makes the untagged lock's own workspace list decide which Bun wrote it.

Root cause

lock_dialect in crates/socket-patch-core/src/hosted/governing_root.rs works out whether a Bun root lock honours ! from a version tag alone: a top-level configVersion in bun.lock, or the cNfGvRsN trailer in bun.lockb (#1496). Bun 1.3.0 and 1.3.1 write neither tag (1.3.2 is the first release that does), so their locks read as pre-1.3, and package_json_workspace_root counted the excluded member as governed by the root lock. Both the hosted refusal (redirect_workspace_lockfile_elsewhere) and the vendored stray-lock check (member_stray_lock → vendor_lockfile_missing) go through that function.

Fix

New lock_lists_member helper. Only a directory that a ! pattern excludes reads differently between the two Bun dialects, and only an untagged lock is ambiguous. For that case:

• If the lock lists the member, or any other !-excluded directory, it was written by Bun before 1.3, which lists every member it matched. The member is governed.
• If it lists none of them, it is a 1.3.0/1.3.1 lock and the member is a standalone project.
• If the list can't be read (unparseable or missing section), today's wider pre-1.3 reading is kept.

Workspace keys are normalized (./ prefix, \\ separators) before comparing. Real bun.lockb files written by Bun 1.2.23, 1.3.0 and 1.3.1 for the same ["packages/*","!packages/b"] root are added under tests/fixtures/bun-lockb/negated-workspace. They were generated with the official @oven/bun-linux-x64 builds; the text-lock outputs are inlined in the test.

Accepted trade-off: suppose a stale Bun 1.2 lock was last written before every !-excluded directory existed. Then it lists none of them and is read as 1.3.0/1.3.1, so the member is not refused. Nothing in such a lock tells it apart from a 1.3.0 lock. This matches 5.0.0, before #1480. The fix keeps the refusal whenever the lock lists any excluded directory.

Test evidence

• untagged_bun_lock_reads_its_workspace_list (new). Covers 1.2.23 / 1.3.0 text locks and 1.2.23 / 1.3.0 / 1.3.1 binary locks: the hosted refusal, member_stray_lock (the vendored path), the unexcluded member packages/a (still refused), and that each verdict comes from a parsed list. Red without the fix: bun.lock of 219 bytes: left: Some("redirect_workspace_lockfile_elsewhere") right: None. Green with it.
• untagged_bun_lock_listing_any_excluded_dir_is_pre_1_3 (new): the per-lock decision and ./ key normalization.
• bun_lockb_with_a_config_version_tag_honours_negated_workspaces: the untagged 1.2.23 row now uses the real 1.2.23 lock for this workspace (it lists packages/b; still refused). The old fixture was a non-workspace project lock.
• bun_config_version_is_read_at_the_top_level_only: its pre-1.3 lock now lists packages/b, as a real one does, so the test still checks only that a nested configVersion isn't read as the tag.
• VEX discovery golden regenerated for the new fixture directory (empty entries).

Commands run locally:

• cargo fmt --all -- --check and cargo clippy --workspace --all-features -- -D warnings: clean.
• cargo test -p socket-patch-core --all-features --lib: 6473 passed. The 5 failures were the 4 chmod-based tests that don't hold under root in this sandbox (copy_tree, vlt_heal, pypi_poetry, pypi_requirements) plus the golden corpus, now regenerated.
• in_process_redirect_pnpm (38), in_process_vendor_bun_takeover (40), mode_migration_bun (35): all pass.
• e2e_bun_lockb: 23/24. binary_shared_bundled_record_hosted_pin_is_managed needs patches-api.socket.dev, which the sandbox can't reach; it doesn't touch this code path.
• A full cargo test --workspace ran out of the sandbox's disk, so CI covers the rest.

No wrapper (npm/, pypi/, gem/) changes: they only dispatch to the binary.

Per-issue checklist

• Bun 1.3.0 and 1.3.1 locks have no configVersion, so a "!"-excluded member is still refused in hosted and vendored mode after the #1492 fix (regression since #1480) #1521 hosted member refused → untagged_bun_lock_reads_its_workspace_list (code row)
• Bun 1.3.0 and 1.3.1 locks have no configVersion, so a "!"-excluded member is still refused in hosted and vendored mode after the #1492 fix (regression since #1480) #1521 vendored vendor_lockfile_missing → same test (member_stray_lock row)
• Bun 1.3.0 and 1.3.1 locks have no configVersion, so a "!"-excluded member is still refused in hosted and vendored mode after the #1492 fix (regression since #1480) #1521 text and bun.lockb variants, 1.3.0 and 1.3.1 → fixture rows

/code-review high

Fixed: per-lock version decision, key normalization, the shared read_regular_to_bytes helper, and the stale WorkspacesDialect docs. Not fixed: (1) the stale-1.2-lock case above, which can't be told apart from the lock bytes; (2) the repeated lock reads. They happen only for a !-excluded member under an untagged Bun lock, so caching isn't worth the extra code.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FEZwdsnHoJB8vCKkMFipaY

Generated by Claude Code