AdversaryGate

What the AdversaryGate MCP server does

The AdversaryGate MCP server acts as a verification gate for AI coding workflows. It examines a baseline and a patch, runs relevant tests, and emits one of three decisions: MERGE, BLOCK, or INCONCLUSIVE. The design treats uncertainty as a distinct result rather than converting missing evidence into a pass.

A merge requires all evaluated claims to pass, the full repository suite to pass on the patch, measured diff coverage of at least 80%, and sufficient suite strength when that metric can be calculated. A failing claim or a collateral regression produces BLOCK. Harness errors, missing measurements, weak tests, and a baseline that was already failing produce INCONCLUSIVE.

How it works

The gate uses a baseline-versus-patch model. Tests that existed on the baseline are restored when needed so the patch cannot change its own oracle. This protection also applies to test helpers and declared test-support paths. When a patch changes an existing test file, the collateral suite can run the baseline version against the new code as well.

Evidence classifications distinguish a test that failed before and passes after the patch (fixed) from one that passes on both sides (no_regression) or a newly added test (discarded). The resulting evidence records whether a claimed fix was actually proven. Decision precedence is BLOCK, then INCONCLUSIVE, then MERGE, so direct failure outranks missing evidence.

Setup and configuration

Install the package with pip:

pip install adversary-gate

The gate is intended for Python/pytest repositories. For another stack, provide a --test-command; the command can be run, but the resulting decision remains INCONCLUSIVE rather than MERGE.

Coverage evidence must be measured to clear the coverage requirement. The documented demonstration uses diff and coverage JSON inputs, with both artefacts hashed using SHA-256. The repository also includes a demonstration script that runs three scenarios as subprocesses and checks their exit codes.

Tools and capabilities

The AdversaryGate MCP server supports these documented capabilities:

• Run the repository’s test harness against baseline and patch states.
• Detect failed claims, collection errors, syntax errors, missing files, timeouts, and other unverified harness results.
• Require measured diff coverage and evaluate suite strength before allowing a merge.
• Protect baseline tests, helpers, and configured test-support files from being replaced by the patch under review.
• Report structured evidence, including test outcome classifications, claims_fixed, fix_proven, and oracle information.
• Return process outcomes that distinguish mergeable, blocked, and not-yet-verifiable work.

Limitations and notes

AdversaryGate MCP server coverage is explicitly strongest for Python/pytest repositories. Other language stacks may execute their test command but cannot receive MERGE according to the provided documentation. INCONCLUSIVE is an intentional safety result and uses exit code 2, allowing CI to treat the work as not ready rather than as a successful run.

The provided material does not document a transport, named MCP client, environment variables, or a separate configuration service. It also does not establish that every test framework or repository layout is supported. A legitimate behavior change that rewrites existing tests may be blocked because the baseline tests are treated as the contract requiring human review.