Oconee Runtime — AI Governance (IDE extension)

AI action governance for coding agents — brings Oconee Runtime's real-time policy enforcement to your editor. It watches AI-assisted activity (prompts, pastes, dependency installs, git operations, file mutations) and evaluates it against your organization's policies, surfacing warn and block decisions inline before risky actions complete.

Supported editors

A single VS Code-compatible build runs unchanged on the entire VS Code family. The host editor is auto-detected at activation (see src/core/hostEditor.ts) and reported for diagnostics.

JetBrains is not a VS Code fork and cannot run this extension. It is served by a native plugin that reuses this same detection/enforcement engine via a Node sidecar (src/sidecar) — no logic is duplicated.

Why Open VSX for Cursor & Windsurf?

Cursor and Windsurf cannot install from Microsoft's Visual Studio Marketplace (licensing), so they pull extensions from the Open VSX registry (open-vsx.org). Publishing there is what makes Oconee installable in those editors. Users can also side-load the .vsix directly:

Cursor / Windsurf → Extensions panel → "…" menu → Install from VSIX…

Features

• Prompt & secret scanning — flags secrets, API keys, prompt-injection, and untrusted-source content before they reach an AI tool.
• @oconee chat participant — type @oconee <text> in Copilot Chat (VS Code 1.90+) to scan a prompt for secrets and policy violations before you send it.
• Guarded model provider — the Oconee Runtime (Guarded) language-model provider evaluates prompts pre-flight and can block them before they ever reach the model.
• Claude Code guard (Agent Runtime Guard) — an enforcement hook for Claude Code that blocks prompts and tool calls that violate policy before they reach the model, using the same detection engine as the IDE and browser extensions. Detection + audit run on every plan; blocking is Growth+ and above. See the bundled CLAUDE-CODE-GUARD.md.
• Dependency risk enforcement — warns/blocks on typosquatting and packages with sensitive capabilities.
• Git guardrails — surfaces force pushes, destructive operations, commits touching sensitive files, and pushes to untrusted remotes.
• Editor DLP — passively scan files as you open/edit/save them and show inline squiggles for detected secrets, including in source code. On by default (oconeeRuntime.editorDlp.enabled). Only real violations are flagged — provider API keys, private keys, JWTs, SSNs, hard-coded credential literals (quoted or high-entropy — not apiKey = getKey()), and strong passwords — so ordinary code stays quiet. .env files are inverted: secrets are expected there so they're not flagged — instead any misplaced code / non-KEY=VALUE line is flagged (code in a secrets file is the anomaly). Other config files (.ini/.conf/.npmrc/…) stay exempt. Turn it off to scan only on demand via Scan Workspace / Scan Selection. Content sent to AI tools is always scanned regardless.
• Workspace sensitivity awareness — escalates enforcement in high/critical repositories.
• Side panel & sign-in — the Oconee Runtime shield in the Activity Bar shows live protection status (Protection Active / Offline); sign in with your Oconee dashboard credentials.
• Managed deployment (MDM) — ship org-wide defaults and lock the Claude Code guard via managed settings; see the bundled managed-settings.example.json.
• Telemetry — privacy-safe signals (enums and boolean flags only — never raw code, paths, URLs, or credentials) sent to your Oconee Runtime backend.

Governed AI tools

Tool identity is classified in src/core/toolClassifier.ts and is governable per-org from Settings → Tool access (allow / warn / block):

copilot, cursor, windsurf, codex (OpenAI Codex — CLI / IDE / cloud), claude_code, chatgpt, claude, gemini.

Codex is classified as an execution-enabled, high-risk agent (it can read the workspace, write files, and run commands), the same tier as Claude Code.

Setup

1. Install the extension.
2. Open Settings (Cmd+,) → search Oconee Runtime, or run Oconee Runtime: Open Settings from the Command Palette.
3. Sign in from the Oconee Runtime side panel (Activity Bar shield) or run Oconee Runtime: Sign In — this is the same login as the dashboard.
4. Configure (optional — the defaults point at the hosted backend):
   • oconeeRuntime.apiEndpoint — your Oconee Runtime API URL (default https://api.oconeeruntime.com).
   • oconeeRuntime.apiKey — an API key for your organization.
   • oconeeRuntime.telemetryEnabled — toggle telemetry transmission (default true).
   • oconeeRuntime.editorDlp.enabled — passive in-editor secret squiggles (default false — see Editor DLP above).
   • oconeeRuntime.logLevel — debug / info / warn / error (default info).

Commands

The side panel is opened from the Oconee Runtime shield in the Activity Bar (there is no "Open Panel" command); it auto-reveals on first run after an install or upgrade.

Build & package

pnpm --filter oconee-runtime build      # compile TypeScript → dist/
pnpm --filter oconee-runtime package    # produce oconee-runtime.vsix (needs vsce)

Publish

One artifact, two registries — both gated on a token being present:

# VS Code Marketplace
VSCE_PAT=<token> pnpm --filter oconee-runtime publish:vsce

# Open VSX (enables Cursor & Windsurf)
OVSX_PAT=<token> pnpm --filter oconee-runtime publish:ovsx

Or do build + (optional) publish in one step via the release packager:

PUBLISH=1 VSCE_PAT=… OVSX_PAT=… bash scripts/build-release.sh 0.1.0

Privacy

Oconee Runtime is built to never transmit user-authored content. Detected signals are reduced to opaque enum values and boolean flags; raw prompts, code, file paths, branch names, commit messages, remote URLs, and credentials are never stored or sent.

Support

Issues: https://github.com/oconeesoftware/ai-action-guard/issues Learn more: https://www.oconeeruntime.com