muralim-ai2/harness-assist

Measure what loads into every request before the user types - discoverable skills, always-on instructions, agent prompt, MCP tools - and warn about skill and instruction bloat with fixes.

Review queued feature branches before pushing - correctness against acceptance criteria, hidden errors, size, security (OWASP), tests and dependencies - and return findings as JSON. Use before the pr-push skill.

Create or update PRDs, technical design specifications and test strategies using configured Frontier HVE templates; validate structure and provenance before handoff.

Run AI-built work as a participatory, verified loop - plan, build and evaluate each feature with tests and screenshots, offer manual checks, use discovery sprints for complex features, and coach toward the right stack and experts.

Explain what is being built and the concepts involved (frontend, API, middleware, data, PRs, CI) at the depth that suits the user, and recommend which experts to involve before moving from prototype to production.

Export the Frontier HVE agents, skills and guard hooks to Claude Code, Cursor or Codex, and connect Azure DevOps through Microsoft's Azure DevOps MCP server.

Turn a request into feature_list.json for the harness loop - independently verifiable features, each with a shell check that exits 0 only when the feature works. Use before loop.py init.

Build two or more approaches to one feature in separate git worktrees, let the user run each and tick the winner, then adopt only the ticked option. Use when a feature has genuinely different implementations.

Show verified feature branches as a checkbox list, push only the branches the user ticks, open stacked GitHub PRs, request Copilot code review, and report GitHub Actions checks. Unticked branches stay queued for a later session.

Stop a vibe-coded prototype from growing into an unmanageable or non-enterprise system - flags code over 5,000 lines, open-source components where licensed enterprise services are expected (for example FalkorDB, Tesseract, SQLite, local vector stores), and self-built infrastructure, then recommends a solution architect and a complexity report.

Rank the Frontier HVE skill library for the current task, show what is loaded within the loadout budget and what else is relevant, and let the user tick extra skills to load.

Run the project's checks the harness way - the current feature through loop.py verify, or every feature check as a regression run - and report exact commands and exit codes.