decree-harness now reads your team's decisions.

npx decree-harness already scanned routes, OpenAPI specs, scripts, env vars and database models to generate an agent harness. It now also reads the decisions your team wrote down in ADRs, post-mortems, CLAUDE.md, AGENTS.md, Cursor rules and Copilot instructions, and stores them in decree.json.

Each decision has a slug id, a title, the constraint an agent must follow, a status (live, proposed or superseded), governs globs for the paths it applies to, and the source file it came from. ADRs keep the status they declare. Rules-file bullets and post-mortem lessons start as proposed.

You review them with decree-harness decisions, confirm the ones that hold with decisions confirm, and retire old ones with decisions supersede or decisions reject. The harness regenerates after each change.

Every generated target now includes a get_decisions tool. An agent passes the paths it is about to change and gets back at most eight live decisions for those paths, most specific first. decree-harness decisions for <path> shows exactly what it will get.

We scope decisions because loading one large rules file into every session has a cost. A February 2026 study from ETH Zurich (arXiv 2602.11988) found that repository context files did not generally improve task success on the coding tasks it tested, and raised inference cost by over 20% on average.

Run npx decree-harness in a repo to try it. The docs cover every command. The hosted workspace, which will hold decisions across all your repos, is still coming and open to design partners first.