What a check finds

Every check runs the hosted policy catalog over the files the change writes, plus any repository rules the repository declares. The catalog is a pinned set of Bifrost policies: structural and data-flow rules that run on the source as written, with no build and no language model. The service pins the catalog version, so a check run twice on the same code gives the same result.

Kinds of findings

Policy ids name their pack, so a finding’s id tells you what kind it is.

Pack	What it reports	Examples
bifrost.security.*	A value from a request or other untrusted source reaches a sensitive operation without passing through a sanitizer. The finding names the path, across files.	A servlet parameter reaching the SQL text of Statement.executeQuery (bifrost.security.java.servlet-parameter-to-jdbc); code, LDAP, XML, XPath, and log-forging injection; a request value written to the application’s own store and read back into a query.
bifrost.security.secret-like-code	A literal that looks like a credential, matched by pattern.	An API key or private key pasted into source.
bifrost.correctness.*	Code that is wrong on at least one path.	Dynamic evaluation of a string; unsafe deserialization; a resource acquired and still open at an exit of the procedure (resource-lifecycle); a Go data race; an overwritten value that is never read; a discarded pure result; a contradictory condition; an unreachable statement; a member access on a Python receiver whose class provably lacks it.
bifrost.maintainability.*	Code that will cost more than it should.	Dead code; oversized code.
Code smells	Patterns worth a second look.	Loop-invariant sorting; regular-expression compilation, file reads, parsing, database or network calls, subprocesses, or sleeps beneath nested loops.

Rules are added and tightened across Bifrost releases. The Bifrost policy documentation describes each built-in pack and the rule language they are written in.

Severity and certainty are different

Severity (error, warning, note) is the rule’s judgment of how bad the pattern is. --fail-on picks which severities fail a check; the default, finding, fails on any new finding at all.

Certainty is Bifrost’s judgment of whether this location is really an instance of the pattern:

Certainty	Meaning	What to do
definite	Bifrost proved the pattern or relation holds at this location.	Fix it.
possible	The match depends on something Bifrost could not settle: an ambiguous receiver or dispatch, name-based resolution, several candidate declarations, or a “may” relation.	Read the code before changing it. It can be a false positive.
heuristic	A text-pattern match, such as a secret-like string.	Confirm it the same way.

A possible finding is still reported and still fails the check, because an agent that reads the code can usually settle it in seconds. When it is wrong, ignore it and later checks leave it out.

What a finding looks like

src/main/java/orders/ExportServlet.java:48  request parameter reaches SQL text  definite · error
  ID: 0f8fad5b-d9cb-469f-a165-70867728950e
  bifrost.security.java.servlet-parameter-to-jdbc: req.getParameter("q") reaches
  Statement.executeQuery(String) through OrderQuery.build (OrderQuery.java:21).

A linter sees two files that each look fine. The data-flow rule sees one path, and reports it once, on the line the change wrote, with the step that proves it. A path through an open dispatch that Bifrost could not resolve is reported as possible; a binding it could not resolve at all makes the rule inconclusive rather than clean (see Read a report).

Languages

C, C++, C#, Go, Java, JavaScript, Kotlin, PHP, Python, Ruby, Rust, Scala, and TypeScript. Coverage differs by language and by rule: each rule claims the languages it has been validated on, and runs only there. The Bifrost documentation records what each front end resolves.

What a check does not report

• Findings that were already in the code. They are counted in the summary (72 already in the code), never listed, and never fail the check.
• Findings in files the change does not write. The check analyzes the files the patch touches, in the context of the whole repository.
• Anything a model might say. There is no AI review in a check, and no code is sent to one.

The web app’s own scans of a repository are a different surface: broader, triggered by pushes and pull requests or on a schedule, and paid in credits. See The web app.