Why AI code integrity is the new category

The gap nobody named

Linters find style and familiar static mistakes. Type checkers enforce declared contracts. SAST finds known security patterns. Test suites exercise the behavior their authors chose to cover. Each is valuable, but none owns the moment when a coding agent says a change is done and a human has not reviewed it yet.

That moment has its own failure distribution:

Phantom dependencies: imports and APIs that look plausible but do not exist or were never declared.
Placeholder behavior: handlers, branches, and error paths that compile while doing no real work.
Silenced failures: catches, ignores, and assertions weakened to make a run look green.
Claim gaps: a coherent change that solves a neighboring problem instead of the task it was given.
Missing evidence: a passing status without proof that the intended tests ran or remained meaningful.
Hidden blast radius: a small edit whose callers and contracts extend beyond the visible diff.

AI code integrity is the category that asks whether agent work earned its claim before scarce reviewer attention is spent on it.

Why it is not just another scanner

The category needs more than a larger rule catalog.

It starts with the change. Diff scope distinguishes what the agent introduced from inherited repository debt.
It separates facts from judgment. Deterministic findings and evidence decide where they can; model output stays labeled and bounded.
It preserves intent. Claim Check evaluates explicit acceptance criteria instead of guessing success from a confident summary.
It lives in the agent loop. CLI commands, Skills, and Harness hooks can return findings before the agent declares completion.
It produces reviewable evidence. Human output, JSON, SARIF, reports, and attestations make uncertainty visible rather than converting it into a pass.

This complements linters, tests, SAST, vulnerability scanners, and human review. It does not replace them or certify that software is correct.

The Shipmoor shape

Shipmoor runs in your environment. Source code, diffs, prompts, findings, verdicts, evidence, paths, repository and branch names, test output, and acceptance criteria are excluded from normal Shipmoor authentication, entitlement, allowance, and analytics requests. Provider-backed checks use the coding-agent or model relationship you configure; Shipmoor does not host the model.

The product has four direct command families, not a fictional umbrella command:

shipmoor claim-check checks a change against approved intent;
shipmoor review provides advisory Code Review;
shipmoor test-evidence checks the integrity of test evidence; and
shipmoor blast maps Blast Radius.

Every supported product command requires an authorized account. A newly created personal account receives 30 days of full Pro access without a card. Afterward, Free includes unlimited deterministic Scan, Agent Harness, and Agent Skills, plus five monthly invocations for each command family above. The four counters are independent; reruns and agent invocations count like user invocations, while internal work composed inside one admitted Claim Check does not double-charge. Paid Pro is $19 per month or $190 per year and makes the four command families unlimited.

The trust boundary

The compact promise is: Free account. Local execution. No source upload.

That does not mean no service contact. Shipmoor contacts its service for account and device authorization, signed entitlements, Free allowance consumption, security, billing when you choose Pro, and a narrow analytics allowlist. Those requests carry operational metadata such as opaque account/device identifiers, CLI version and platform, command-family key, random invocation idempotency key, invocation source, and timestamp. They do not carry project content or command results.

This boundary keeps local execution useful without pretending that account-based access and reliable allowances can operate with no network control plane.

Start at the review boundary

Run deterministic Scan immediately after an agent edit. Add Skills or Harness hooks when the agent should invoke that checkpoint itself. Use Review for a second reading, Test Evidence for the test signal, Blast Radius for impact, and Claim Check when the task’s obligations need a verdict.

Start free or read how the workflow fits together.