GeezerVison

Assisted-by: Claude Code:claude-opus-5-5

Agent-mode apply, scan and vex looked for a `cargo vendor` tree only at
`<cwd>/vendor`. A custom vendor directory, a workspace root's vendor
directory seen from a member, or a `vendor/` that no config wires all
made apply patch a copy cargo ignores while reporting success, and made
vex judge the wrong copy.

The crawler now follows the `[source.crates-io] replace-with` chain of
every cargo config cargo reads (each ancestor's `.cargo/config*`, then
`$CARGO_HOME`'s) and crawls the directory source it ends on, resolved
the way cargo resolves config paths. Without a directory replacement it
crawls the registry cache, and a configured directory that is missing
yields nothing rather than a copy cargo doesn't read.

Refs #338, #455

Assisted-by: Claude Code:claude-opus-5-5

A project whose .cargo/config replaces crates.io with a directory
source (`cargo vendor`) builds every crate from that directory and
never reads a registry. `scan --mode hosted` still pinned the crate
to a per-patch Socket registry and reported success, so every fresh
`cargo build --frozen --offline` failed with "no matching package".

The hosted cargo rewriter now follows the project config's
`[source.crates-io] replace-with` chain and, when it ends on a
directory source, skips every cargo dependency with nothing written
and a `redirect_cargo_directory_source` warning that points at agent
mode. A registry mirror replacement is still redirected.

Refs #455

Assisted-by: Claude Code:claude-opus-5-5

Adds real-cargo e2e coverage for a custom-named vendor directory and
for apply run from a workspace member, and documents how agent mode
finds the directory source and the new hosted refusal code.

Refs #338, #455

Assisted-by: Claude Code:claude-opus-5-5

Three fixtures staged a bare `vendor/` with no source replacement and
relied on the crawler treating it as cargo's source. They now write the
`.cargo/config.toml` that `cargo vendor` asks for, and the resolver
strips a BOM through the shared text helper.

Refs #338

Assisted-by: Claude Code:claude-opus-5-5

Rollback and prune look for an agent-mode patch left in the shared
registry cache when the project crawl searches only a `cargo vendor`
directory. That check still looked for `<cwd>/vendor`, so a custom or
workspace-root vendor directory skipped it, and an unwired `vendor/`
triggered it needlessly. It now asks the crawler for the directory
source the project's cargo configs name.

The hosted refusal's message now says why a Socket registry pin breaks
an offline vendored build, instead of claiming cargo never reads it.

Refs #338, #455

Assisted-by: Claude Code:claude-opus-5-5

cursor Bot left a comment •

edited

Loading

Uh oh!

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

The dispatch, all-ecosystem rollback and scan-scope tests staged a
bare `vendor/` crate tree and relied on the crawler treating it as
cargo's source. They now write the source replacement `cargo vendor`
asks for, so they keep exercising the vendored-crate path.

Refs #338

Assisted-by: Claude Code:claude-opus-5-5

cursor Bot left a comment •

edited

Loading

Uh oh!

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

A --cwd such as `b/../app` walked `b`'s ancestors and could pick up a
sibling directory's `.cargo/config.toml`, so the crawl followed a
source replacement cargo never sees from `app`. The path is folded
lexically before the walk.

Refs #338

Assisted-by: Claude Code:claude-opus-5-5

cursor Bot left a comment •

edited

Loading

Uh oh!

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

Folding `..` before the config walk also moved the reported vendor
directory onto the folded path, so path-scoped scan and rollback, which
strip the given cwd as a prefix, saw the vendored crates as outside the
project. The folded path now only picks the config files; a directory
under it is reported back under the cwd as given.

Refs #338

Assisted-by: Claude Code:claude-opus-5-5

cursor Bot left a comment •

edited

Loading

Uh oh!

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

The new prefix test compared crawler output with a path that still
held `..`, but on Windows `std::path::absolute` folds it, so the
crawler (and path-scoped commands) report the folded path there. The
test now builds its expectation with `absolute` too.

Refs #338

Assisted-by: Claude Code:claude-opus-5-5

cursor Bot left a comment •

edited

Loading

Uh oh!

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

The hosted refusal for a `cargo vendor` project only read the project's
own `.cargo/config*`. A directory source set in a monorepo root's config
or in `$CARGO_HOME` was missed, and a `cargo local-registry` was never
recognised, so `scan --mode hosted` still pinned the Socket registry
and every `cargo build --frozen --offline` broke.

The hosted scan now resolves the source replacement from every config
cargo reads for a build there, read-only, and passes it into the cargo
rewriter. A directory source or local registry from any of them skips
every cargo dependency with `redirect_cargo_directory_source`. Agent mode
keeps crawling the registry cache for a local registry, which cargo
unpacks there. The process-level resolve now lives in
`source_replacement` so the crawler and the hosted scan share it.

Refs #455

Assisted-by: Claude Code:claude-opus-5-5

cursor Bot left a comment •

edited

Loading

Uh oh!

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

The format models are pure by rule (formats::architecture_tests), but
the last commit put the file-reading resolve into
formats/cargo/source_replacement.rs. It now lives in the cargo crawler
as `CargoCrawler::crates_io_source_for`, which the hosted scan calls;
the format module keeps only the pure chain resolution.

Refs #455

Assisted-by: Claude Code:claude-opus-5-5

cursor Bot left a comment