RUNCHECK_ · AGENT AUDIT

Show the code. Know the limits.

For external agents and custom RUNCHECK projects: submit a GitHub repository for a commit-scoped code audit with findings, coverage and evidence. This is separate from automatic RUNCHECK platform certification. Provide the repository URL, project name and review goal; RUNCHECK detects public or private repository access automatically.

NEW REQUEST

New code audit request

Follow the four steps below to request an audit of one GitHub repository. All settings are shown: only the repository, project name, review goal and final confirmation are required. Existing requests and GitHub connections stay in USER. Never paste API keys, bot tokens, private keys or source code.

01 · REPOSITORY & PROJECTAudit an existing agent repository.

GitHub repository root Required. Open the repository homepage and copy its URL, for example https://github.com/owner/repository. Do not use a branch, file, token or invitation URL. RUNCHECK detects public or private access automatically.Project name Required. Use a name you will recognize in My code audits and on the completed public result.Requested function and review scope Required. In at least 20 characters, explain what the agent should do and what you want reviewed. Example: “A Telegram assistant that answers documentation questions. Check credential handling, access restrictions and dependencies.” Do not paste source code or secrets.

Branch, commit and review focus (optional)

Not sure what to choose? Leave the branch and commit empty, and keep the overall audit priority. RUNCHECK will review the repository's default branch.

Source branch (optional)Only fill this in to review a specific branch. Leave empty for the default branch. The review stays on the selected branch; it never falls back to main. Changing the branch requires a new request.Approved commit SHA (optional)Only for a specific approved release. Also enter its branch above. Use the full 40-character lowercase commit SHA. If the branch head differs, the audit stops instead of reviewing another release.Review priorityKeep “Overall bounded code audit” for a general review. Another choice adds a focus to your private brief; it does not change the bounded scan or turn on the optional external checks in step 03.

02 · REPOSITORY ACCESSNo need to reconnect an already authorized repository.

Public or private? RUNCHECK detects it.

Public repository · no connection needed

Paste its URL in step 01 and continue. You do not need to connect a GitHub account.

Private repository · read-only permission needed

RUNCHECK checks your signed wallet's existing authorization for this exact URL. Already authorized? No action is needed here. If access is missing or needs renewal, the request is saved with the next step; it is not scanned until access is verified.

Authorize a private repository

Skip this if the repository is public or already authorized. This button is only for a private repository that needs new or renewed read-only access.

Save your request before leaving this form. You can submit it below even while private access is pending. Opening GitHub leaves this page, so unsaved form entries may be lost.

Connect and sign with the wallet that owns your audit request.
Open GitHub using the button below. Choose Only select repositories, select the repository for this request and finish the authorization.
GitHub returns you to My GitHub connection. Open your saved request in My code audits and use VERIFY GITHUB APP + QUEUE to check access and queue the audit.

Your connected accounts, repository list and revoke controls stay in My GitHub connection; they are not shown in this new-request form.

Access is read-only and limited to your selected repositories. If you have not saved a request yet, return here after authorizing and submit a new one. Never paste a GitHub token, password or source code.

Privacy and publication

Completed audit results are published automatically. Public repositories have public reports; private repositories have redacted public summaries and wallet-private details. Pending requests are not published.

What becomes public?: Public-repository reports show bounded commit evidence and fixed-category findings from already public code. Private reports show only project name, commit, check date, bounded status and coverage counts. Queued, access-pending and unfinished requests are never public reports.
What stays private?: For a private repository, your brief, repository URL, detailed findings, source paths and access or provider-consent details are visible only to you and authorized RUNCHECK administrators. Full evidence remains behind your signed USER access.
How is the code accessed?: A repository URL alone is not authorization. RUNCHECK uses narrowly scoped, server-side read access and never executes repository code. Private source is read transiently and is not stored by this triage flow. Never paste a GitHub token, credentials or source code.
Can I stop access later?: You can withdraw the request and revoke repository access in USER. An already completed result remains as monitoring-stopped historical evidence.

This is not a signed NDA, code escrow agreement or warranty; agree those separately before sharing sensitive source.

03 · OPTIONAL ANALYSISBoth choices start off. Neither is required.

Choose any extra checks

You can submit the standard audit with both options unchecked. Each checkbox grants a separate permission for this request; choosing one does not enable the other. Neither option guarantees a passed result.

I allow the optional dependency advisory check through OSV.dev for this request.

AI source review · optional

RUNCHECK RepoScan Intelligence looks for code-risk signals in selected files. Unlike the dependency check, this option shares selected source code outside the local scan. Leave it unchecked if you do not want that transfer.

If you opt in, RUNCHECK sends a commit-bound selection from this repository, after RUNCHECK automatically detects whether it is public or private to the access-controlled RUNCHECK RepoScan Intelligence gateway. The review uses up to 8 batches, each with at most 64 files, 64 KiB per file and 128 KiB of source. The total source-transfer ceiling is 1 MiB, with a shared 15-second deadline and 8,192 output tokens across all batches. Eligible files outside these limits remain explicitly unreviewed. Obvious credential literals are redacted before transfer; files and secrets outside those patterns still require your own review.

Consent version: repository-semantic-consent-v2 · analysis profile: sha256:204bfd305a7ec0d8ec1fd790e2f642980ae9aee24071e7d5deba82f92755ca09.

The model receives source as untrusted data and has no RUNCHECK tools or actions. Its fixed-category, file-and-line signals are AI inferences for human review, not a safety, “NON LARP”, authenticity or runtime verdict. Provider processing and retention terms may apply. The separate OSV dependency consent and local static checks do not authorize this transfer.

I explicitly authorize this bounded source transfer for this repository review and current analysis profile. I have authority to share the selected source and understand that an analysis-profile change requires fresh consent.

04 · CONFIRMConfirm authority before the request is saved.

I have permission to request this work. I understand that automated triage is limited, automatic public detection does not verify my ownership, and a human audit needs separate scope confirmation.

What happens after you submit?

The request is saved in USER → My code audits. If repository access is ready, the bounded scan is queued. If private access is missing, the saved request tells you how to authorize it. You can leave this page and follow the request in USER; a saved request is not yet a completed audit.

You can prepare this form now. Connect and sign with your wallet in the header to submit it.

AUDIT LIMITS

Read evidence, not a promise

Every result is tied to one commit and states exactly what was checked and what was not.

COMMIT PINNEDA new commit on the selected branch invalidates old evidence.

BOUNDEDCoverage and exclusions remain visible in every report.

READ ONLYRepository code and install hooks are never executed.

What this audit checks and what it does not

“Non-LARP” means asking for evidence, not promising that any agent is safe. This automated pass checks a bounded commit-pinned archive for specific risk patterns. It never grants a capability certificate or unlocks a token launch.

Each result names the exact selected-branch commit and the ruleset. A detected new commit invalidates the old result and triggers a new bounded scan.
Unchanged repositories become eligible for a periodic check every 24 hours, processing at most one request every 15 minutes globally. Queues may delay a check; last-check and next-check times are shown. Optional signed GitHub push webhooks invalidate old evidence immediately and queue a fresh scan within the same global budget. An administrator must configure each repository webhook; daily checks remain the fallback.
No commits for 30 days adds an inactivity warning, not a fraud accusation. Stale monitoring is also highlighted.
Red means a critical rule matched and needs investigation; amber means warnings, partial coverage or pending work. “No rule matches” is not a complete security clearance.
Up to 1,000 supported text files, 256 KiB each, within a 20 MiB compressed / 64 MiB expanded archive and 16 MiB scanned-text budget. Exclusions and partial coverage stay visible. Optional consented OSV lookups check at most 100 exact dependency versions for known advisories. No sandbox execution, financial safety check or proof that deployed code matches the repository.
RUNCHECK RepoScan Static adds a bounded local static/offline pass. It does not send repository source to a model provider or execute repository code. Its status and profile appear separately in each report.
RUNCHECK RepoScan Intelligence is available through the managed analysis gateway. It runs only after separate, versioned consent and reports AI inferences independently.
Trusted deployment-host evidence, when present, binds a signed read-only image measurement to the reviewed commit. It trusts the configured host operator and is neither hardware attestation nor a guarantee of runtime safety.

GitHub read-only archive API · Provider rate limits · OSV advisory lookup