What CodeVetter embeds instead of building, what it refuses, and why — the hub page for per-category evaluations.

Third-party tooling decisions

CodeVetter should embed proven tools rather than reimplement them, but the product’s positioning narrows the field hard. Two constraints disqualify most of the market before features are even considered:

1. Privacy is the product. CodeVetter verifies code on the user’s machine. Any tool that transmits source, manifests, or lockfiles to a third party is rejected regardless of quality.
2. Local-first and offline. The desktop app works without a network. Tools requiring a hosted engine, an account, or a live API are rejected.

A third constraint governs licensing: this is a commercial desktop product that redistributes what it bundles. AGPL, SSPL, and non-commercial licenses are blockers. MPL and GPL are subprocess-only at best.

Research verified 2026-08-30. Licensing, pricing, and maintenance facts decay — re-check before acting on anything here. Items the research could not confirm are marked UNVERIFIED on the detail pages and should be treated as open questions, not findings.

Status vocabulary

• Approved — qualified for a bounded lane, but not executable there yet.
• Trialled — exercised against this repository with measured output, but not part of a maintained path.
• Wired — invoked by a tracked local command, hook, workflow, or product adapter.
• Configured — registered with a maintainer client, but not yet callable because authentication, product authorization, or a client restart is pending.
• Rejected — disqualified for the named lane. Rejection in the private customer-code product lane does not automatically reject a public-repository maintainer aid.

Product adoption

These are product-lane decisions. The three collector resources now also have an unreleased implementation and qualification receipt. A signed/notarized production archive remains a separate release gate.

Repository and maintainer tooling

These tools subsidize discovery and evidence collection. CodeVetter still owns receipt qualification, taxonomy, budgets, and the final measurable verdict. The maintainer client also registers CodeVetter’s repository-scoped runtime MCP as codevetter-runtime-exec and its non-executing receipt MCP as codevetter-verification; the names preserve the execution boundary. The product-generated packaged sidecar is registered separately as codevetter-history. Newly added registrations appear as callable client tools in the next Codex session, so the packaged sidecar was also protocol-smoked directly in this session. GitHub’s browser OAuth is complete, and its token remains in the server process rather than Codex configuration. Future repository scopes must still be enabled in CodeVetter and copied from its opaque generated configuration rather than synthesizing database arguments.

The first CodeQL run also exposed a cleartext localStorage API-key field. Source tracing showed that the old browser gateway execution path had already been removed and only its non-functional Settings panel remained. The remediation removes that panel and allowlist-migrates the shared record to rubric fields instead of adding a credential dependency to preserve dead behavior; issue #194 remains open until pushed CodeQL evidence confirms the alert is closed.

The OSV baseline changes the order of operations for repository dependency work: fix or classify the measured lockfile baseline before adding another scanner. The tracked baseline evidence records 52 advisory/package matches and the exact database hashes; generated SARIF and receipts remain ignored scratch. cargo-audit remains the highest-leverage product embedding candidate because the rustsec crate runs inside the existing Rust backend with no process boundary or sidecar to codesign. Its SARIF 2.1.0 output is real but undocumented in the README and absent from the changelog — pin >= 0.22.0 and trust the source, not the docs.

Reject

Adopt only with explicit configuration

• Trivy ships telemetry on by default, contacting check.trivy.dev. Requires --skip-version-check --disable-telemetry. Separately, --offline-scan does not mean offline — it only suppresses dependency-identification API calls, not DB downloads or telemetry.
• Terminal-Bench needs network access at verification time. Its task template sets network_mode = "public", all 89 TB-2.0 tasks set allow_internet = true, and the verifier’s own test.sh runs apt-get update and curls astral.sh. These are package-registry dependencies rather than live websites, so pre-baking images is tractable — but it is real work, not a flag.
• cdxgen v13 requires pnpm >= 11 and node >= 24, which conflicts with this repo’s pnpm@10.33.2. Pin @cyclonedx/cdxgen@12.x, or prefer Syft.

What not to outsource

Changed-line coverage — the metric that answers “did the agent’s change actually get exercised?” — should be computed in the Rust backend, not delegated. It is a small deterministic join: git diff -U0 yields changed lines, LCOV yields hit counts, intersect them. It is the core verdict input, it must be reproducible and explainable inside the evidence bundle, and it must not depend on a Python install existing on the user’s machine. Use diff-cover (Apache-2.0) as a cross-check oracle in the test suite instead of a runtime dependency.

Two traps that produce false verdicts if ignored are documented in coverage: V8’s loaded-files-only blind spot, and LCOV’s line-granularity loss on dense lines.

Related

• codebase-context-tools-landscape.md — April 2026 survey of codebase indexing and context tooling. Its DeepWiki assessment was independently reconfirmed in August 2026.
• failed-approaches.md — constraints left behind by things that broke. Check before adopting anything that touches the package manager or the data layer.