Review Procedure

Thirdpass review coverage is built from individual file reviews. The current procedure reference is file-focused-review/v1.

The unit of review is one target file from one package archive. The agent focuses on that file, while the rest of the package archive remains available for context. The recorded result is evidence about the target file in that package version.

Review Scope

Each selected file gets its own focused review run. When a command reviews several files, the CLI runs a separate file review for each target and bundles the results into one saved review record.

Any finding from a run must be about the target file. The agent can inspect other files to understand the target file, but those context files need their own review runs before they count as covered.

This keeps coverage inspectable. A reader can see which files were reviewed, which file hashes were covered, and which files still need review.

What The Review Looks For

The procedure looks for concrete supply-chain relevant behavior, including:

• install-time or build-time behavior that runs commands or changes the environment
• network access
• credential, token, or secret access
• dynamic code loading or execution
• hidden intent, obfuscation, or tampering behavior
• behavior that changes files outside the package's expected work area

A reported finding needs a clear connection to package behavior. It should explain what triggers the behavior, what happens, and why that matters. Unsupported concerns are not recorded as findings.

Context

The target file is reviewed in context. Useful context may include package metadata, build scripts, dependency declarations, imports, callers, nearby modules, tests, examples, generated files, and documentation.

Context matters because a file can be harmless on its own but important when it is imported, executed, packaged, or reached from install-time code.

Recorded Evidence

Each file review records structured evidence:

• the registry, package name, package version, and package archive hash
• the target file path and file hash
• the review procedure version
• the reviewer identity used for the public review record
• the agent name, model, and reasoning effort
• a summary and confidence level
• review comments and source selections when there are findings
• runtime and token metrics when the agent reports them

The CLI checks that comments point to the target file. It then adds package and file hashes, review metadata, and runtime details when available.

That structure lets Thirdpass show what was reviewed. Successful raw agent transcripts are not stored as part of the shared review evidence.

Clean Coverage

A clean result means the target file was inspected and no in-scope issue was reported for that file in that package version.

Clean coverage is still useful evidence. It distinguishes files that were reviewed from files that have not been reviewed yet. It does not prove that the package is safe, and it does not apply to other package versions.

Package Coverage

Package-level coverage is calculated from accepted file reviews. A package version reaches full coverage only when accepted file reviews match the package archive, match the reviewed file hashes, and cover every file in the archive manifest.

The manifest tracks regular files: ordinary file entries, rather than directories or archive metadata. Thirdpass checks file count, line count, and byte coverage against that manifest.

For cargo-vet exports, Thirdpass emits a package-level audit only after that coverage condition is met. See cargo-vet Methodology for how file reviews become cargo-vet audit evidence.

Local First, Then Shared

Before submitting, the CLI saves the review locally. The server validates the submission and stores it as pending or approved based on moderation rules. Approved reviews become shared coverage.

If submission fails, the local pending review remains available for retry.

Limits

Thirdpass reviews are evidence, not proof. A review can miss behavior. A clean review result is not a formal verification result, a full dynamic analysis, or a general security certification.

The evidence is tied to the package archive and file hashes that were reviewed. Other package versions, changed files, and unreviewed files need their own coverage.

Confidence can improve as more important files are reviewed and as independent reviews cover the same package version.