What’s New.

Your agent now knows how to drive VibeHard, and every step says what comes next.

• — Your agent is told what to do. A connected agent now receives VibeHard’s working rules when it connects, and three ready-made prompts (vibehard-build, vibehard-status, vibehard-next) in agents that list MCP prompts.
• — Every step says what comes next. vibehard_progress returns next_action: the current step (for example “PRD, step 2 of 8”), one plain sentence, the exact tool to call, and whether VibeHard is waiting on the agent or on you.
• — Research is labelled honestly. vibehard_build_start accepts research_evidence, an agent’s own research, shown to the interview and the PRD as not REA. REA findings stay for apps that can actually be run or inspected, and a build that mentions replacing or rebuilding an app asks for the exact location and your rights to analyze it.
• — Rules travel with the project. Folders VibeHard creates now contain AGENTS.md and a CLAUDE.md that points at it. VibeHard never touches your own global files.
• — Capabilities are reported. vibehard_session_open and vibehard_install_info list what this install supports, and the first connection after an install carries a one-time welcome.
• — Installer finish screen. A short “You’re ready” block says what to tell your agent first, links the first-build walkthrough and names vibehard-agent doctor. Installer prerequisite checks (part 2) are included.
• — Honest install info. vibehard_install_info reports the absolute path of the installed binary, matching what the installer registers.

Upgrading: From 1.7.1 or later: arrives automatically within 30 minutes, or run vibehard-agent update.

Intel Macs: stay on 2.0.0 for now. 2.0.3 has no Intel Mac build.

Safer planning, sturdier repairs, Rust builds that clear the supply-chain gate, and apps that are proprietary by default.

Why it matters: Fewer builds stop on problems the factory can handle, and what it builds is marked as yours alone unless you say otherwise.

• — Planning asks before it builds. Disputed high-severity planning findings now wait for a person’s ruling instead of being decided silently.
• — Repairs follow the whole chain. The repair loop handles blocked inputs and command-line contract findings the way the core loop does, so fewer builds stop on a finding the factory can fix itself.
• — More work at once. The daemon runs four build runners by default, and a build waiting on a person no longer holds a runner.
• — Rust builds get through the supply-chain gate. Rust projects the factory creates now declare a licence, so the gate no longer stops them on their own crate. The vendored crate registry was also fixed.
• — Proprietary by default. Apps the factory builds are marked all rights reserved to the client unless the client asks for an open-source licence: LicenseRef-Proprietary with publish = false for Cargo, UNLICENSED with private: true for npm, and a LICENSE file. Nothing is open-sourced on your behalf.
• — Redesign epochs. A redesign request now starts a clean epoch of the plan.
• — Safer tests. The factory’s own tests no longer touch your real ~/.vibehard or service manager.
• — Passed our official acceptance run: one run, no retries, frozen spec; 0 of 44 gates blocked (12 passed; 32 did not apply to that small command-line project); the build audit accepted it; all three output fixtures matched byte for byte; a separate reviewer confirmed the evidence.

Note: In that acceptance run the plan approvals came from the test harness, not a person, and a test agent drove the build. It verifies the factory code; it does not certify the downloadable build, which is built and signed separately.

Upgrading: From 1.7.1 or later (including 2.0.0 and 2.0.1): arrives automatically within 30 minutes, or run vibehard-agent update.

Intel Macs: stay on 2.0.0 for now. 2.0.2 has no Intel Mac build. If automatic updates are on, the 2.0.0 updater on an Intel Mac picks the Apple Silicon package, so about every 6 hours it downloads the 2.0.2 Apple Silicon package, fails its pre-install check, and stays on 2.0.0. To stop these attempts, set "auto_update": "off" in ~/.vibehard-agent/config.json.

The first 2.0 release whose factory code passed our official acceptance run: clearer failures, cleaner snapshots, a stricter command-line contract.

Why it matters: Your agent sees the full reason a check failed, so fewer repair rounds are spent guessing, and the release passed our official acceptance run before it shipped.

• — Gate failures now include both standard output and standard error.
• — A production environment setting (NODE_ENV) inherited from your shell no longer stops test runners and type checkers from being installed.
• — Installed packages and build output are excluded from snapshots, which fixes the snapshot error some projects hit during repair.
• — Generated command-line tools must answer both --help and --version.
• — The TypeScript scaffold writes a .gitignore and creates its output folder first.
• — Passed our official acceptance run: one run, no retries, frozen spec; 0 of 44 gates blocked (12 passed; 32 did not apply to that small command-line project); the build audit accepted it; all three output fixtures matched byte for byte; a separate reviewer confirmed the evidence.

Note: In that acceptance run the plan approvals came from the test harness, not a person, and a test agent drove the build. It verifies the factory code; it does not certify the downloadable build, which is built and signed separately.

Upgrading: From 1.7.1 or later (including 2.0.0): arrives automatically within 30 minutes, or run vibehard-agent update.

Intel Macs: stay on 2.0.0 for now. 2.0.1 has no Intel Mac build. If automatic updates are on, the 2.0.0 updater on an Intel Mac picks the Apple Silicon package, so about every 6 hours it downloads the 2.0.1 Apple Silicon package, fails its pre-install check, and stays on 2.0.0. To stop these attempts, set "auto_update": "off" in ~/.vibehard-agent/config.json.

Groundwork for agentic workflows inside the factory.

Why it matters: 2.0.0 introduced the groundwork for factory-built workflows. Those workflows are not yet offered as a hosted service. 2.0.1 adds the verified factory fixes; we recommend upgrading.

• — A workflow request tool for your agent, alongside the software request.
• — A workflow compiler with eight workflow-specific gates and signed workflow bundles.
• — Buy-before-build now covers 22 component categories, with a new review that flags plans which rebuild something they could adopt.
• — Installer fixes.

Upgrading: Installed copies move to 2.0.1 automatically; we recommend upgrading.

A stricter planning audit.

Why it matters: The plan review got harder to fool, so problems in a plan surface before code is written instead of after.

• — An unchanged section of a revised plan can no longer count as re-verified.
• — Credit from an earlier review no longer carries over to a changed plan.
• — Quiet drift in the spec, in buy-or-build decisions, or in language choices is detected.
• — When the plan review finds a contradiction in the spec, it now asks you instead of getting stuck.
• — The factory rejects tests that can never fail.

Note: Our acceptance runs on 1.7.4 stopped safely at the approval stage instead of completing. The first fully completed official run came with the fixes now shipped in 2.0.1.

Upgrading: Superseded by 2.0.1.

Go and Elixir join the build languages; jobs survive a restart.

Why it matters: More of your stack can go through the factory, and an interrupted build picks up where it stopped.

• — Go and Elixir are supported build languages, alongside TypeScript, Rust and Python.
• — A job interrupted by a restart of the factory’s background service resumes instead of failing.
• — Safer rollback: update rollback targets are checked so they cannot land outside the install folder.
• — On Linux, the build sandbox starts from an empty root with an explicit allow-list.
• — The secret-scanning and static-analysis gates read large compiled files in full.

Upgrading: Superseded by 2.0.1.

Automatic updates work end to end.

Why it matters: You no longer have to remember to update: installs check every 30 minutes, verify the signature and apply the update.

• — The background update check now reaches its update engine in the shipped program.
• — Fixed a race in supervised rollback and a containment check in rollback path validation.
• — The daemon start-identity check now works on Mac.
• — The packaged program is smoke-tested as a compiled file before release.

Upgrading: Superseded by 2.0.1.

Update machinery and a clean first boot.

Why it matters: Fresh installs start cleanly, and the groundwork for automatic updates arrived.

• — Automatic-update timer introduced. In this build its checks did not complete; that was fixed in 1.7.1, and manual updates worked.
• — The embedded database files are included in the download, so a fresh install boots cleanly.
• — Better socket diagnostics for the background service.

Upgrading: Run vibehard-agent update once; updates are automatic after that.