Inspiration
Codex already provides an execution boundary through sandboxing and permissions. Dollar focuses on the human layer: helping developers understand what a risky action means for their actual project, choose a safer path, and recover when needed.
What it does
Dollar is a local-first Windows desktop companion and safety supervision layer for supported Codex local tool actions. It turns technical agent operations into understandable decisions with project context, approvals, audit, snapshots, and recovery.
For a request such as git clean -fdx, Dollar keeps the action fail-closed, explains likely project impact, and can offer a locally validated safe preview instead of the destructive original command.
What we built during OpenAI Build Week
Dollar existed before Build Week with deterministic risk classification, local approvals, snapshots, rollback, and project-health tools. Those capabilities are runtime foundations and are not claimed as Build Week work.
During Build Week, Dollar added:
- GPT-5.6 Safety Briefs using
gpt-5.6-sol - minimized and redacted action context with strict Structured Outputs and
store: false - a locally validated safer-action workflow for exact supported
git cleancommands - Codex
PreToolUseupdated-input rewriting with a second Hook-side validation - new audit fields for model, response, latency, strategy, replacement, and user decision
- a failure-safe Judge Mode with three fixed scenarios
If a user chooses Run safe preview, local code constructs the allowlisted replacement git clean -ndx; the server recomputes the mapping and the Codex Hook independently revalidates it before returning updatedInput. GPT-5.6 recommends a strategy enum only—it never supplies a command for execution.
How Codex and GPT-5.6 were used
Codex was the primary implementation environment: it audited the Electron architecture, established the dated pre-event baseline, implemented the strict safety schema and redaction boundary, built the strategy validator and Hook-side revalidation, integrated the approval UI, and generated/refined the core, integration, fail-closed, and Renderer evaluation suite.
GPT-5.6 is advisory, not the safety authority. It receives only minimized, redacted action metadata and returns a concise project-specific explanation plus an enum strategy. Local deterministic code remains responsible for risk classification, command construction, validation, approval, fail-closed behavior, and audit. If GPT-5.6 is offline, times out, refuses, or returns malformed output, Dollar's local protection continues unchanged.
How to test
- Download the signed Windows x64 Portable from the release link below.
- Launch it—administrator access is not required.
- Choose Open Build Week Demo from the Dollar menu.
- Run Risky delete + safe preview.
- Review the Safety Brief, choose Run safe preview, and confirm the audit records
git clean -fdx → git clean -ndx.
The supported submission platform is Windows 10/11 x64. The Build Week edition is isolated from the Steam/production edition with its own app ID, executable, user-data directory, Hook, token, audit state, port, and release directory.
Privacy and limitations
Risk rules, approvals, snapshots, project history, audit, and rollback remain on the device. Optional Safety Briefs send only a minimized, redacted action summary to OpenAI. Persisted API credentials use Windows-backed Electron safeStorage; the Renderer cannot read a stored key back.
Dollar supervises supported Codex local Hook paths and does not replace Codex's sandbox or claim to be a universal security boundary.
Log in or sign up for Devpost to join the conversation.