[Mikola Lysenko (mikolalysenko)]

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

Refs #1606
Refs #1607

Hosted rescans no longer send the same reference request twice. This is one slice of #1606/#1607, so both use Refs and stay open for the rest (see Not fixed here).

Root cause

A hosted scan asked the patch API's reference endpoint about the same patch uuids twice:

1. The withdrawn-pin sweep (Restore withdrawn hosted pins upstream during scan #1592, scan/withdrawn.rs) asks about every uuid the lockfiles pin.
2. The redirect (run_redirect_selected_untracked in scan/hosted.rs) then asks about every selected uuid.

On a rescan, the pins and the selection are the same patches, so the second request repeated the first. That cost one extra network round trip on every hosted rescan, in every ecosystem, including the bun and yarn-berry rescan scenarios in these issues. #1693 found this while profiling them.

Changes

• The sweep keeps the answers it fetched as KnownReferences. It keeps only settled answers (granted, reused, withdrawn, not_found, forbidden), and only for uuids it actually asked about.
• Scan hands those answers to the redirect through rollout::Gate::with_references, in both the --json arm and the human arm. The redirect asks only about the selected uuids that have no known answer. When the sweep covered all of them, it sends no request.
• Some uuids are asked about again:
  • pending_build or an unknown status, because the answer may change while the run goes on (a build that finishes mid-run still gets pinned)
  • a uuid the response left out (a 404 chunk answers nothing)
• get --mode hosted runs no sweep and passes no gate, so it is unchanged.

Test evidence

Red → green: cargo test -p socket-patch-cli --all-features --test in_process_scan_withdrawn

• a_json_rescan_asks_about_a_pinned_uuid_once and a_human_rescan_asks_about_a_pinned_uuid_once failed on main with two identical requests ([[UUID], [UUID]] vs [[UUID]]). Both pass with the fix.
• New behavior-pinning tests, which pass:
  • the_redirect_asks_only_about_uuids_the_sweep_did_not: a replacement uuid is still looked up, and only it
  • a_pending_build_answer_is_asked_again
  • unit test withdrawn::tests::only_settled_answers_are_known
• 13/13 in the suite pass.

Local checks on 4fd3d54:

• cargo fmt --all -- --check: clean
• cargo clippy --workspace --all-features -- -D warnings: clean
• cargo test --workspace --all-features --no-fail-fast: 14,867 passed, 11 failed. All 11 are environmental, and none touch the changed code:
  • 9 assert that a write to a 0o555 directory fails, which can't happen in the sandbox because it runs as root. They are the same set Speed up bun and yarn-berry rescans (#1606, #1607) #1693 lists: copy_tree, vlt_heal, pypi_poetry, pypi_requirements, in_process_redirect, repair.
  • 2 are in e2e_bun_lockb: binary_shared_bundled_record_hosted_pin_is_managed (the sandbox proxy refuses the real patch host) and workspace_text_migration_heals_on_rerun, which fails the same way with main's sources.
• I did not run a bench: the change removes one request per hosted rescan, and the tests count requests exactly.

Per-issue checklist

• Perf regression: bun-isolated/rescan wall +21% (cc938aa6..17793173; #1279 candidate) #1606 (bun rescan): duplicate reference request removed, covered by the *_rescan_asks_about_a_pinned_uuid_once tests (the path every ecosystem shares)
• Perf regression: yarn-berry/rescan wall +21% (a43cf327..17793173; #1321/#1323/#1397 candidates) #1607 (yarn-berry rescan): same

Not fixed here (why the issues stay open)

• bun: the .bun store liveness walk (With Bun's isolated linker, vex refuses every hosted patch as not_applied after the usual in-place bun install, because it checks orphaned node_modules/.bun registry entries that Bun never removes (regression from #496) #599), as listed in Speed up bun and yarn-berry rescans (#1606, #1607) #1693
• yarn-berry: peak RSS +10%

/code-review high

The review reported 8 findings. I fixed 6:

• stale pending_build answers
• treating a 404 chunk as an answer
• a doc that contradicted itself
• clones in answers_for, now a consuming into_answers
• references taken out of the sweep two different ways, now done once
• a duplicated test helper

I skipped 2:

• Capped-rollout predicate. On a failed lookup it still checks every selected row. This is the same behavior as before. Narrowing it to the fetched rows would drop the candidates for rows that did get answers.
• Move the cache into ApiClient. The only duplicate is scan's sweep→redirect handoff, and get runs no sweep. A client-wide memo would also have to handle when an answer goes stale, for every caller.

---

Generated by Claude Code

[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 4fd3d54. Configure here.}