Skip to content
Verification gate · CodeTrussRev A

Every agent change gets a first pass.

CodeTruss is the deterministic first-pass verification gate for AI-written code. Routine changes clear the checks and pass; material changes escalate for human sign-off — and every verdict leaves a signed receipt you can re-verify locally.

Boundary
Local first-pass gate
History
Explicit receipt sync
Health
Full-codebase audit

Specimen receipt · Boundary

Signed

PASS

Task
Add input validation to /api/orders
Scope
3 allowed · 0 unexpected
Analysis
15 registry analyzers · local
Verification
npm test · exit 0
Signature
integrity-signed · re-verify: codetruss verify latest
CodeTruss certification stamp

Example signed Boundary receipt with verdict PASS: task, scope, analysis, verification, and signature rows. Values are illustrative.

local by default — explicit sync only

Exhibit 0 — Watch a boundary hold


One agent turn, replayed exactly: what the agent wrote, what the boundary classified while it wrote it, and the single reason the receipt was signed REVIEW_REQUIRED.

A replayed agent turn. The agent is asked to add a password-reset endpoint. It writes src/auth/reset.ts and edits src/lib/email/send.ts, both inside the allowed scope, then edits src/billing/webhooks.ts, which is outside it. CodeTruss returns a fast scope check during the turn, then runs all 15 local registry analyzers when the turn stops: one LOW finding from the todos analyzer at src/auth/reset.ts:41, and the configured verification command pnpm test exits 0. The signed receipt records the verdict REVIEW_REQUIRED for one reason — 1 file(s) changed outside approved scope: src/billing/webhooks.ts — and exits 1. Running codetruss verify latest re-checks the signature. Local receipts run a local SAST subset but omit the hosted symbol graph and the hosted-only passes, and report hosted Health scores as not available.

Specimen session · one agent turn

Boundary · CodeTruss

Local

Agent session

Claude · CodeTruss hooks active

Specimen session · deterministic replay · analyzer ids, scope words, and verdict logic are the shipped CLI’s

0 false positives across 8 repositories, 7 of them not ours — true of that corpus, and no longer the whole story. A larger ten-repository corpus we did not choose found false positives, including a CRITICAL SQL-injection report on a parameterized drizzle query. Fixed in CLI 0.2.53. 6 of 9 bugs caught on a corpus we wrote ourselves. No other tool was run, so none of it says we score higher than anyone. Read the benchmark and the correction

Boundary · History · Health

One verification gate, from prompt to codebase.

BoundaryLocal · before the PR
Record the task and starting Git state, enforce scope, flag sensitive surfaces, run 15 registry analyzers plus your verification commands, and write an integrity-signed receipt.
HistoryExplicit sync · organization scoped
Receipts stay local until you choose one to sync. Sync sends a privacy-minimized copy for organization history and shared review; it never uploads the patch or verification output.
HealthHosted · whole codebase
Run a read-only repository audit with the registry suite plus graph, SAST, and route-authorization passes, health scores, architecture maps, findings, and an issue roadmap.

Install + activate · macOS + Linux

curl -fsSL https://codetruss.com/install.sh | sh
codetruss setup
  • 15 registry analyzers stay local
  • One guided setup, then automatic checks
  • Integrity-signed Markdown + JSON

§ 01What you receive


When a change escalates, the next question is the codebase around it. One hosted audit answers that in four exhibits. Deliverables — not a lecture.

Exhibit A

Architecture map

Every module, edge, and cycle — drawn from the real symbol graph, not a stale wiki page.

Example architecture map: modules app, lib, api and prisma with import edges, and a three-file import cycle flagged in red.

Exhibit B

Health certificate

Five scores that tell you where the risk actually lives, weakest marked in red.

Certificate of Health

Snapshot a1b2c3 · Read-only


STRUCT
HEALTH
TESTS
SEC
DOCS

Example health certificate with scores: structure 87, health 74, tests 71, security 64 (weakest, marked in red), docs 93.

Exhibit C

Issue roadmap

A punch list of findings, ranked and filed as labeled GitHub issues under a milestone.

FindingSevEff
  • Vendored dir committed: .agent/ (1,862 files) HIGHM
  • No CI pipeline detected HIGHS
  • Deep nesting in geminiService.ts MEDS

Ranked by severity × effort · filed as GitHub issues

Exhibit D

Fix PRs

Approve a finding and CodeTruss opens the branch and the pull request. Nothing ships without you.

codetruss/fix-a1b2c3 → mainOpt-in
- const apiKey = 'sk-live-9f3…'- writeFileSync(out, data)+ const apiKey = process.env.API_KEY+ if (!apiKey) throw new Error('API_KEY missing')+ await writeFile(out, data)

§ 02Method


  1. Connect

    Install the GitHub App on the repositories you choose — and nothing else.

    LEAST-PRIVILEGE GITHUB APP · REVOCABLE

  2. Snapshot

    CodeTruss takes a read-only tarball of your default branch. Your code is never executed.

    READ-ONLY TARBALL · NEVER EXECUTED

  3. Analyze

    A policy-selected analyzer suite walks the symbol graph and source: structure, import cycles, tests, secrets, SAST, broken access control on Prisma and Drizzle routes, and known vulnerabilities.

    STATIC ANALYSIS · SAST · SYMBOL GRAPH · AUTHZ GRAPH · OSV

  4. Roadmap

    Findings become labeled GitHub issues under a milestone, with a tracking issue on top.

    LABELED ISSUES · MILESTONE · TRACKING ISSUE

§ 03The deliverable


The audit ends as a typeset report you can hand to a client, a CTO, or your future self: scores, architecture, findings, and a prioritized plan, enriched with AI review and shared as a read-only link. Freelancers and agencies deliver it as the codebase audit their engagement is billed on. Don’t take our word for it: the specimen shown here is a real audit of a repository you already trust.

The specimen is a small, well-maintained library — the calibration point for what a healthy codebase scores. For what the same method finds in a real AI-assisted SaaS, read an independent teardown of a pinned public repository: auditing an AI-built SaaS.

Specimen · Healthy-repo calibration

expressjs/cors


OVERALL
DEBT
ARCH
SEC
DOCS

Shared read-only

send it to the client as-is — or white-label it

§ 04Fee schedule


Free

Try CodeTruss on a single repository.

1 repository · 5 scans a month · health scores & basic reports

$0

Pro

For developers and small teams shipping agent-written code.

☛ recommended

10 repositories · 50 scans · full reports, exports, fix PRs, your own AI keys

$19 / mo

Agency

For agencies delivering audits to a book of clients.

Client workspaces · white-label reports · 200 repositories · 25 seats

$249 / mo


Full fee schedule

§ 05Engineer-led option


Shipping an AI-assisted release and need more than a report? The Release Assurance pilot is a fixed-scope engagement: an engineer contracts your exact release state, exercises up to three workflows that cannot fail, repairs a bounded set of defects, and hands you an evidence-backed go, conditional-go, or no-go decision in ten business days.

See the fixed pilot scope

Prepared by CodeTruss — the deterministic first-pass gate for AI-written code.

Signed

Footnotes

† Read-only tarball snapshot. Your code is never executed. GitHub App access is least-privilege and revocable.