gitlawb-zero is winding up a 0.4.0 release, and the commit stream shows a project maturing its distribution and its security boundaries at the same time.
Shipping the native binary the right way
A merged PR ships the native binary as platform optionalDependencies on npm. That is the correct packaging move for a tool with per-platform builds: instead of one bloated package that downloads everything, npm pulls only the binary for your OS/arch. Faster installs, smaller footprint, and no “why did a Linux binary land on my Windows machine” confusion. For a security-focused coding agent, clean distribution also means fewer opportunities for a bad binary to ride along.
Sandbox: tell the truth about Windows failures
Two sandbox commits landed in the window. One classifies silent wrapped Windows command failures as sandbox denials — fixing the misdiagnosis where a failed command was reported as a blocked one (the same class of bug we flagged on the Hermes side). The other uses a WRITE_RESTRICTED token when no DenyRead paths are configured, tightening the default write posture. Together they show the project taking least-privilege seriously instead of defaulting open.
This pairs with the trust-boundary work the project shipped earlier — the gitlawb-zero MCP trust-boundary beware (v0.4.0) is the security writeup that frames why “deny and prove” is the right default, and the stale-lock denial-of-service prevention is the same hardening instinct applied to the session layer.
A GitHub Copilot provider enters
An open issue and a corresponding PR add GitHub Copilot provider and authentication support. That widens gitlawb-zero from its own model backend to a router you can point at Copilot — useful if your team already pays for a Copilot seat and wants the agent’s DevSecOps workflow on top.
Where it fits
gitlawb-zero’s angle has always been the DevSecOps framing — scan, guard, and constrain the agent rather than just accelerate it. The deep-dive on the gitlawb-zero agent lays out that model, and this 0.4.0 batch is the practical proof it is still moving. The gitlawb-zero vs Codex comparison is a useful read on how its local-first, security-first stance contrasts with Codex’s cloud worktree parallelism. If you want an agent whose defaults lean toward “deny and prove” rather than “run and hope,” this is the branch to track. The Hermes security writeup is a good companion read on why provider isolation and sandbox truthfulness are the same battle.
What to watch in the 0.4.0 cycle
The Copilot provider is the most interesting addition because it changes who gitlawb-zero is for. Until now the pitch was “bring your own model, own your agent.” Adding a Copilot backend means teams that already pay GitHub can layer gitlawb-zero’s DevSecOps guardrails on top of a seat they have anyway — no second model bill. That lowers the adoption barrier without giving up the local-first, audit-friendly posture. If the 0.4.0 line holds, expect the next moves to be about default-deny presets that teams can drop into CI without hand-tuning the sandbox.