SinghAbhinav04/dev-house

Hackeroom is a thoughtful attempt to turn Claude Code into a visible, configurable development team. Members can have separate roles, models, skills, budgets, and permission modes, while a deterministic orchestrator controls planning, review, coding, testing, auditing, and deployment. The implementation goes beyond prompt-based roleplay. A manifest-driven hook restricts tool use, protects locked artifacts, blocks unknown members, and limits most agents to the active project. Shared memory uses a compact index with detailed entries loaded on demand, and token costs are tracked by member and model. The Next.js interface offers office, squad, and team views without hiding the underlying terminal activity. Architecture, security, and setup documentation are unusually candid about how the system works and where its protection ends.
The security model still needs work before the application should be exposed beyond a trusted local machine. Its API routes have no authentication or origin checks, yet they can modify the roster, approve commands, start agents, and delete or reset project state. The documentation should require loopback-only access, and the application should add authorization and CSRF protection before supporting LAN access or tunnels. The shell hook is correctly described as a guardrail rather than a sandbox, but indirect command execution can bypass its pattern checks. The current behavior also treats unparseable agent output as approval. That fail-open path is especially risky during plan review, testing, and security auditing. Invalid or incomplete structured output should pause the run or require human confirmation.
Engineering quality is strong overall. The repository includes a lockfile, TypeScript checks, linting, production builds, CI, and 14 standalone test suites. The 54-case hook contract suite tests the real shell hook, which gives the most important boundary meaningful coverage. The next step should be route-level and browser tests covering roster changes, approvals, interrupted streams, concurrent requests, and destructive reset actions. The whole-file event store has no locking and prevents safe parallel workers, so an append-only log or transactional store would strengthen recovery and unlock the advertised team model. The README also says 13 test suites when 14 are present. Finally, security reports should have a private disclosure channel instead of directing potentially sensitive escape reports to a public issue.