[Mikola Lysenko (mikolalysenko)]

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

Stacked on #2015 (which is on #2013).

Summary

Hosted scans no longer call /patches/view. A view carries every patched file's blob contents, and v5 keeps no hosted ledger, so the scan was fetching full patch bodies only to feed checks that run within the same scan. Each of those checks now uses data the scan already has, or fetches its own:

Consumer	Before	Now
In-run --vex	records from the scan's view fetches	fetches the records it attests itself via vex_sources::fetch_records, the same path standalone vex uses; output unchanged
Gem stale-install probe	patched file hashes from the view	sha256 of the archive RubyGems installed from (<home>/cache/<leaf>.gem, and committed vendor/cache) vs. the grant's artifact sha256
Python stale-install probe	patched file hashes from the view	the patched wheel's own RECORD, downloaded through its grant URL (sha256-checked) only when an installed copy exists and its WHEEL tags match; never under --offline
vlt warm-tree heal	record byte checks	vlt's hidden lock alone: the pinned integrity means healthy, anything else means stale, and no readable hidden lock means undeterminable
In-memory engine / Node binding	fetchPatch provider, presence-checked only	not called; the binding still accepts a fetchPatch method and ignores it, so existing hosts (depscan's autopatch provider) keep working

When an agent-mode manifest holds a record for the same patch, the gem and Python probes use that local record instead, so an agent→hosted migration stays quiet.

Vendored mode, agent mode, get <uuid>, vendor eject and standalone vex are unchanged and still use the view.

Behavior changes

• record_fetch_failed no longer exists. A view failure during in-run VEX is reported as vex_record_fetch_failed.
• A hosted scan with --vex where every record fetch fails ends with a record_unavailable skip instead of manifest_not_found.
• Review this one: the Python probe now downloads a patched wheel, which a hosted scan never did before, but only for a confirmed patch with an installed copy whose tags match, at most once per patch. An artifact-free alternative would be the installed dist's direct_url.json archive hash, but not every installer writes one, so I kept the wheel RECORD.
• vlt: a FIFO or symlinked hidden lock used to get its store entry invalidated when a record proved the bytes stale. Now the entry is kept and the advisory asks for vlt ci, as it does when there is no hidden lock (new test). The advisory names the missing or unreadable hidden lock.

CLI_CONTRACT.md, index.d.ts and docs/testing/poetry-compatibility.md are updated.

Validation

Check	Result
cargo fmt --check	pass
cargo clippy --workspace --all-features -- -D warnings	pass
core --lib (isolated HOME, see #2015 for the NuGet cache test)	7137 passed
CLI lib	860 passed
e2e_redirect_gem_stale_install	42 passed
e2e_vex_lockfile	354 passed
hosted_memory_engine	49 passed
in_process_redirect	134 passed
pdm / pipenv / poetry redirect suites	6 / 11 / 9 passed
scan	248 passed
Node smoke test	14/14 passed

An earlier run also passed cli, get, cli_parse_scan, cli_parse_get, hosted_memory_parity, hosted_memory_rollout, in_process_rollback_hosted, in_process_scan, in_process_vendor_bun_takeover, mode_migration_vlt and mode_migration_bun.

Tests that used a view request as proof that a redirect was confirmed now assert zero view calls, or confirm the redirect through a re-scan's envelope. Fixtures that lay down a gem install by hand now write the cache archive RubyGems would. The pdm, pipenv and poetry stale-install tests now serve a real patched wheel from the mock server.

Not run: the hatch and uv real-toolchain e2e suites, which already serve real wheels.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes hosted post-rewrite behavior (new wheel downloads, vlt heal rules, and in-run VEX fetch timing) without altering vendored/agent modes or standalone vex’s overall model.

Overview
Hosted scan / get --mode hosted no longer call /patches/view after a rewrite. v5 does not persist hosted records, and views were only feeding same-run checks that now use data the scan already has or fetch narrowly on their own.

In-run --vex takes confirmed pin UUIDs and loads records through vex_sources::fetch_records (like standalone vex), instead of a pre-built map from the scan. record_fetch_failed is gone; view failures show up as vex_record_fetch_failed / record_unavailable.

Stale-install probes no longer depend on fetched patch bodies: gems compare the installed cache/<leaf>.gem (and committed vendor cache) to grant artifact shas, with agent manifest.json afterHash still used when present; Python downloads the pinned wheel once ( RECORD + matching WHEEL tags) via the grant URL when a copy is installed—not under --offline; vlt heal/stale checks use node_modules/.vlt-lock.json integrity only, so missing/unreadable hidden lock keeps store entries and advises vlt ci instead of record-driven invalidation.

Contract docs and tests are updated (including zero-view assertions and gem cache fixtures).

^{Reviewed by Cursor Bugbot for commit fb3d418. Configure here.}

---

Generated by Claude Code

[cursor]

Stale Bugbot comment from a previous run.

[Mikola Lysenko (mikolalysenko)]

bugbot run

---

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 fb3d418. Configure here.}

[Mikola Lysenko (mikolalysenko)]

Labeled Ready for review by the burn-down agent.

• Head: fb3d4183852150c00139b671240b93019ba88655
• CI: 49/49 check runs green (30 success, 19 skipped by path filters)
• Fixed in fb3d418:
  • e2e_redirect_vlt_build rule (c) without a hidden lock still expected the old record-based heal. This PR makes that case undeterminable (copies kept, advisory says run vlt ci), so the e2e now matches the in-process test and checks that vlt ci installs the patched package.
  • Bugbot (medium): the wheel probe read site-packages files with bare std::fs reads, so a FIFO could block the scan. It now uses the regular-file guard, with a FIFO regression test.
• Bugbot: re-reviewed fb3d418, no new issues.
• Reviewer note: stacked on Refuse vendored patches the same way everywhere #2015, which is stacked on Stop downloading vlt artifacts during hosted scans #2013. heal_rule_c_no_record now covers the same scenario as heal_rule_c_no_hidden_lock because patch records are never fetched; consider dropping one of them.
• Slack announcement not sent this run (Slack send tool unavailable); next run retries.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

Final review brief (for the human approver)

What it does: hosted scans stop calling /patches/view, which returned the full patched file contents just to feed checks inside the same scan. In-run --vex now fetches its own records, the same way standalone vex does. The gem stale-install probe compares the sha256 of the cached .gem with the grant's. The Python probe downloads the patched wheel, verifies its sha256 and reads its RECORD. The vlt heal trusts only vlt's hidden lock. Stacked on #2015, so it merges into that branch, not main.

Risk: medium. I found no security problem and nothing that makes the probes warn wrongly. The risk is in the Python probe: it now judges an installed copy only when its WHEEL tags and dist-info name exactly match the first patched wheel for that patch. Where they don't, a stale venv goes unwarned with no signal (see the open questions).

Look here

• wheel_record.rs:40-50: checks the sha256 before it parses anything. Downloads are capped at 256MB and RECORD/WHEEL reads at 16MB. Nothing is extracted. RECORD paths go through is_safe_relative_subpath.
• wheel_record.rs:134-171:`` the tag gate.
• python.rs:91-110:``` how the override is chosen and the --offline skip.
• hosted.rs:387-407:```` any cached .gem whose sha isn't a patched one now warns as stale.
• vlt_heal.rs:258-286:````` with no readable hidden lock the result is now Undeterminable (it used to be `Healthy`), and the in-run VEX drops it. That errs on the safe side.

Verified: I read the PR's own diff (42843ef..fb3d4183852150c00139b671240b93019ba88655, 31 files). No view calls remain on the scan, memory-engine or Node paths. --offline and --dry-run are honoured. CRLF in RECORD is handled. CHANGELOG.md is untouched and there is no debug leftover. Ran locally: wheel_record unit tests (5), CLI stale (23), in_process_redirect_poetry and e2e_redirect_gem_stale_install, all passing. The poetry test depends on the served wheel, so it exercises the new probe. CI: 30 passed and 19 skipped by path filter (49 checks). Bugbot: passing, and its one thread is resolved.

Changes I made: none.

Open questions / nits (none block merging)

1. When a patch ships several platform wheels, or the patched wheel changes its dist-info name or tags (PyYAML vs pyyaml, py2.py3 → py3), the probe never judges the installed copy. Try every matching .whl override? This could be a follow-up.
2. Every re-scan downloads the whole patched wheel again for each pypi candidate that has an installed copy and no local record. That costs about what the old view fetch did.
3. No test covers the --offline skip of the wheel probe, or a gem with no .gem under <home>/cache (git or path sources).
4. Nit: hosted.rs:974-976 still says "patch records (fetched below)", but that fetch is gone.

Auto-merge is not armed. This PR targets #2015's branch, which has no review rule, so arming it would merge it immediately. Approve and merge it after #2015, or retarget it to main once #2013 and #2015 land.

---

Generated by Claude Code