GeezerVison

Detect project patch ownership across pnpm, Yarn, patch-package, npm, Bun, Composer, Cargo and Go. Skip protected scan candidates and refuse explicit mutations unless --allow-user-patches is passed. Preserve native backend refusals and require explicit consent before overwriting unknown installed file contents.

BREAKING CHANGE: declared project patches are protected by default and agent apply now rejects unknown file-content drift. Document the next-major-release requirement, detection limits and invocation-only consent in the CLI contract.

cursor Bot left a comment •

edited

Loading

Uh oh!

There was an error while loading. Please reload this page.

The url-spelled crates-io [patch] table and ancestor .cargo/config
[patch] fixtures are now caught earlier by the new UserPatchIndex,
which refuses with user_patch_conflict before the cargo-specific
guards run. Pass --allow-user-patches so these tests keep exercising
the guards they were written for (cargo_manifest_patch_source_alias
and user_authored_patch_entry); the default refusal is covered by
tests/user_patches.rs.

Co-Authored-By: Claude <[email protected]>

- apply: release apply.lock before reporting a user_patch_conflict, like
  the manifest-load failures above it, so telemetry can't block siblings.
- core: match Cargo [patch]/[replace] names ignoring case and -/_, since
  crates.io treats those as the same crate.
- core: follow go.work `use` members so a member go.mod replace protects
  its module from the workspace root too.
- scan: hide user_patch_conflict/override warnings under --silent, matching
  user_patches::check; the JSON envelope still carries them.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

The #480 [patch] override shapes are now caught first by the
user-patch gate (user_patch_conflict), so they no longer reach
redirect_cargo_dep_overridden. Scan with --allow-user-patches so the
shapes keep exercising the redirect guards, and ignore the resulting
user_patch_override notice; tests/user_patches.rs covers the gate.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

From /code-review high on #1646:

- A `cd` after patch-package (`patch-package && cd ios && pod install`)
  no longer marks the patch directory unresolvable and blocks all npm.
- Yarn alias locators (`alias@npm:[email protected]`) protect `real`, the name
  the crawler reports, instead of a name that never matches.
- Cargo [patch] paths use vendor's path_is_socket_owned, so a user path
  like `.socket/vendor/../../forks/x` is no longer treated as Socket's.
- yarn.lock is scanned line by line for patch resolutions instead of
  parsed (every Berry lock has compat patches); pnpm locks parse with a
  budget large enough for big monorepos; package-lock/bun.lock are only
  parsed when they mention `patched`, so a stub or broken lock with no
  patch evidence no longer protects every npm package.
- Workspace globs walk directories only, take only dirs with a
  package.json as members, flag a symlink only when it would be a
  member, and resolve `..` lexically (re-rooting `../shared/*`).
- Approved user_patch_override warnings print once: check() no longer
  prints, callers do, and a nested apply leaves them to its caller.
- The hosted re-check refuses through the hosted module's refuse(), so
  the scan envelope is kept and telemetry precedes the JSON error.
- A scan whose every package is protected says so instead of "No
  packages found. Run your package manager's install first."
- Go hosted/build tests for replace guards now pass --allow-user-patches
  so they keep testing the guards behind the consent gate.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

Two review findings on #1646, resolved as decided with the maintainer:

- apply no longer refuses a whole run over one protected manifest entry
  (a postinstall apply would fail every install). Protected records sit
  the run out with a user_patch_conflict warning, everything else
  applies, and skips alone exit 0. The manifest on disk is unchanged,
  and skipped Go entries are kept out of the stale-redirect reconcile.
  `apply --check` stays read-only and verifies every entry. Explicit
  `get` and `vendor` still refuse; vendor now reverts dropped entries
  before its check so a protected record can't block that cleanup.
- The default mismatch policy is the new MismatchPolicy::SocketContent:
  Strict, except a file whose hash names a stored .socket/blobs blob
  (another Socket patch's output, as after an update to a newer patch
  for the same release) is overwritten with the usual
  content_mismatch_overwritten warning. --strict stays fully strict.

Contract and usage docs updated to match.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

- A nested apply under `get` now reports the older manifest entries it
  skips for project patches (stderr and the JSON envelope via the run
  report); `get` checks only its own selection. `scan` already reports
  every protected installed package, so its nested apply stays quiet.
- vendor: a conflict found after reconcile_dropped already wrote is
  recorded as per-package `failed` events instead of a top-level error,
  which the contract reserves for runs where nothing happened.
- patch-package `cd` detection matches `cd` as a command, not the tail
  of a word such as `abcd`.
- The ancestor workspace walk never reaches the home directory or above,
  so a checkout without `.git` doesn't read `~/.cargo/config.toml` or a
  stray `~/patches/` as project evidence.
- Discovery is skipped when no in-scope purl is in an ecosystem the
  detector reads (an empty or pypi-only manifest on postinstall).
- npm/Bun locks are read once instead of twice.
- The scan message for protected packages no longer claims they were
  the only reason the scan came up empty.
- Contract: a pruned or unshared blob means earlier Socket output reads
  as drift; `apply --force` is the explicit way through.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

cursor Bot left a comment •

edited

Loading

Uh oh!

There was an error while loading. Please reload this page.

Clarify that the flag accepts conflicts with existing project patches or local modifications. Rename the parser field and nested forwarding, diagnostics, contract, usage examples and existing tests consistently.

Preserve reason-bearing skip events across standalone and nested apply. Recheck conflicts under the apply lock and report late project patches in human and JSON output; count only actual applications, including mixed failures and multiple installed copies.

Document the event contract and cover conflicts introduced during API fetches with deterministic regression tests.

cursor Bot left a comment •

edited

Loading

Uh oh!

There was an error while loading. Please reload this page.

Resolve conflicts with main's manifest/ledger gitignore checks (#1638)
and the gem variant test split (#1642):

- vendor eject runs main's ledger gitignore refusal first, then the
  user-patch consent check, so both guards stay in place.
- CLI_CONTRACT keeps the PR's skipped-apply event wording and main's
  manifest_gitignored warning paragraph.
- get_failure_paths keeps both sides' new tests.
- The gem singleton mismatch tests (now split across
  cli_gem_variant_mismatch_policy.rs and apply_render.rs on main) pass
  --allow-patch-conflicts, since this PR gates the mismatch overwrite
  behind explicit consent.

Co-Authored-By: Claude <[email protected]>

The nested apply in get / agent scan covers the whole manifest, so
record_apply_outcome was copying every leftover skip (already_patched,
vendored, package_not_installed) for packages the run never selected
into the envelope, inflating summary.skipped. Keep only the
user_patch_conflict skips for unselected packages, which are the
conflicts found after selection that the caller needs to see
(Bugbot finding on #1646).

Co-Authored-By: Claude <[email protected]>

cursor Bot left a comment •

edited

Loading

Uh oh!

There was an error while loading. Please reload this page.

`apply --check` verifies every manifest entry, including packages the
project patched itself, but a plain `apply` skips those packages. A
drift on one of them pointed at a remedy that could never clear it.

The drift detail now names the project patch evidence, and the remedy
line says such a package applies only with --allow-patch-conflicts
after review, or by merging the fixes manually. Check still verifies
every entry, as decided for #1646.

Co-Authored-By: Claude <[email protected]>

cursor Bot left a comment •

edited

Loading

Uh oh!

There was an error while loading. Please reload this page.

The previous remedy still quoted a bare `socket-patch apply`, which
skips every project-patched package, and offered a manual merge that
cannot make installed bytes match afterHash. When a drifting package
has a project patch, the quoted command now carries
--allow-patch-conflicts (after a review prompt), so running it
verbatim clears the drift. The integration test runs that remedy and
re-checks.

Co-Authored-By: Claude <[email protected]>

cursor Bot left a comment