realdiff

realdiff [repo] [--base ref] [--pr ref] [options]realdiff detect reporealdiff warm repo --target ref [--cache-dir dir]realdiff baseline write --findings file [options]realdiff post [--provider=name] --findings file

realdiff finds runtime behavior changes that a source diff does not show. It checks out both sides of a pull request, instruments each build, runs the same existing tests, records method arguments and return values, then diffs those traces. It does not mutate production code, generate tests, or infer behavior from source.The public executable is a thin Rust launcher. It owns argument routing, repository config loading, and detect, then starts a sibling realdiff-managed binary for ref resolution, builds, caches, instrumentation, and posting. Language tracers emit one NDJSON contract; a streaming Rust engine matches events, subtracts self-noise (the base build is run more than once), collapses a changed call tree onto the frontier (the first changed member whose descendants behaved identically), and writes findings.json.Supported languages in the v0.4.0 unified CLI are .NET 8 (Mono.Cecil IL weaving, xUnit), Java (ASM javaagent, Maven/Gradle), Node/TypeScript (CommonJS/ESM Babel hooks), Go (stable AST rewrite into a cache), Rust (stable syn rewrite into a cache), and Python 3.12+ (PEP 669 sys.monitoring, no bytecode rewrite). Mixed-language repositories and multiple entry points are refused rather than guessed; set language and workdir in .realdiff/config.yml.Findings in unedited files are unexpected; changes in edited files are expected. A behavior gap is a changed value that at least one executing test did not assert on. Comments on pull requests use a high-confidence policy by default.

--base ref

Git ref of the stable side (the pull request's base).

Git ref of the proposed change (typically HEAD).

Branch to instrument for warm (cache the noise-baseline traces only).

Resolve the repository path and refs from GitHub Actions or Azure Pipelines environment variables.

Override the temporary work directory used for isolated checkouts.

Write canonical machine-readable findings JSON.

Opt in to a local base-trace cache. Persistence is off unless this is set.

Expire cached traces after a stated window (for example 12h or 7d). Default is one day when caching is enabled.

Keep working traces for a stated hours/days window. Traces are deleted after analysis by default.

Keep temporary Git worktrees. Traces are still deleted unless --keep-traces is set.

Use this suppression file instead of .realdiff/baseline.yml.

Inspect the raw result without applying the suppression baseline.

Do not restore or store base traces even if --cache-dir is set.

Include all unsuppressed findings in provider comments (default comments are high-confidence only).

Print the effective language, work directory, entry point, commands, test projects, and scope. detect-language is an alias.

Run the three baseline instrumented tests for --target and store them in --cache-dir, without a pull-request comparison.

Merge 30-day acknowledgements for every currently actionable member. --expires window changes the lifetime; --no-expiry makes them permanent.

Post a previously written findings.json through a CI provider (for example --provider=azuredevops).

Print help from the managed CLI and exit.

.realdiff/config.yml (optional) overrides detection field by field:language / workdir / build / test

Language id (dotnet, java, node, go, rust), repository-relative work directory, and the exact build/test commands run on both sides.

.NET test project globs, or Java source directories when inference is not enough.

Extra tracing scope. Go exclusions may name exact .go files.

Extra display-redaction rules (digests still use the real value).

Nested realdiff.baseline/2 document, same schema as .realdiff/baseline.yml.

0

Analysis completed; no unexpected behavior changes.

Analysis completed; behavior findings exist.

Evidence could not support a verdict (also used by the launcher for invalid usage).

Could not instrument the repository (also used if realdiff-managed cannot start).

The unmodified repository did not build in this environment.

Only methods the existing tests execute have evidence. Unexecuted edited files report zero traced members, never "unchanged". Incomplete path attribution, source information, call-tree integrity, or coverage is exit 3 (refused), not a clean result.The extracted release directory must stay together: the launcher starts realdiff-managed beside it and needs the tracers and native engine. --engine was removed; the Rust engine is the only path.Traces can contain application values. Redaction is on by default for credential-shaped names and strings, but method names, paths, and the fact that a secret changed remain visible. Treat unredacted traces as sensitive build output. Target tests run with the CI agent's permissions; RealDiff is not a sandbox.Status is early preview (v0.4.0). Python 3.11 and older are refused. Node worker threads, .NET type initializers, and several other callable shapes are recorded as skipped coverage boundaries.

Developed by Nitin Issac as RealDiff (MIT). Earlier releases were named BehaviorDiff; v0.3.0 introduced the Rust launcher and engine, and v0.4.0 added the Python tracer and the all-language container ghcr.io/issacnitin/realdiff.

git(1), git-diff(1), diff(1), diffoscope(1), gh(1), dotnet(1)

Source code · Documentation