readyset-flow0.17.0

Readyset is a standalone omp extension, invoked as /readyset. It turns a rough idea into a grounded, reviewable, executable change — with a code-enforced review gate between "written" and "executing", not just a prompt asking nicely.

Rough idea → GRILL → EXPLORE → PROPOSE → REVIEW → EXECUTE → ARCHIVE

Install

npm install --save-dev readyset-flow
npx readyset-flow install

Installs globally, tied to ~/.omp/ — available in every repo you work in. Run npx readyset-flow configure afterward to set a default language or model (optional), or npx readyset-flow uninstall to remove it later.

Use

/readyset let users export their data as CSV

• Grill interrogates the idea until it's unambiguous — a real Q&A, not a rubber stamp. It also proposes a lane (--lane fast|full to force one): fast folds Explore into Propose and writes a smaller artifact set; full is the workflow below. When it finishes, Readyset offers to carry straight on into Explore & Propose — no second /readyset run, no re-picking the brainstorm.
• Explore greps the actual repo before anything gets proposed, instead of assuming.
• Propose writes a proposal, design, specs, and tasks — grounded in what Explore found, and with a ## Files This Change Will Touch scope contract the gate checks against.
• Review stops for your Approve / Refine / Discard — Discard is the default, so nothing executes without a deliberate look.
• Execute is handed off to core omp: approving dispatches tasks.md to omp's native runtime and Readyset exits, so subagents, parallel tool calls, and live task-checklist updates work as usual. Readyset runs the project's own tests itself — a baseline right after approve, then again as the execution progresses — instead of trusting a self-reported note. The handoff closes on its own: it settles when the model signals readyset_done, when every task gets checked, or when your next /readyset supersedes it, no stall inference. When it's done, Readyset's risk-based review policy recommends /readyset --review <change-id> or skips it with an honest reason.

Pin a cheaper model per phase with --phase-model grill=... --phase-model explore=... (see the full guide).

Already have a brainstorm sitting in .ai/brainstorms/? Run /readyset with no idea and pick it — Grill is skipped, since the ambiguity's already resolved.

Why

Plans drift from the repo, reviews get skipped under pressure, and "done" ends up meaning "the model said so." Readyset turns each of those into a structural gate instead of a habit: ambiguity gets interrogated before anything is written, every claim about the repo is grounded in a dedicated explore phase, nothing executes without an explicit human approve, and "done" needs verification, not just a claim. The code-review turn is risk-based and on demand: it evaluates its triggers against the working tree and runs when /readyset --review <change-id> asks for it, skipping with an honest stub when nothing raises a flag.

Benchmark

Readyset has a headless benchmark harness (readyset-bench) comparing it against native /plan across bugfixes, storage migrations, cross-cutting refactors, and ambiguous feature requests. The numbers from the last run are not currently trusted as a measurement of the current version — see docs/BENCHMARK.md for why, and for the methodology a clean rerun would use.

Learn more

• Full guide — design philosophy, configuration, model/language settings, the review gate in detail, CLI validation, uninstalling, what Readyset deliberately doesn't do, and the full file/package layout.
• Benchmark deep-dive — methodology, per-task breakdown (T01–T12), judge dimensions, request clarity impact, and resource trade-offs.
• Brainstorming outside omp — write a brainstorm in Claude Cowork before you ever open omp.

License

MIT — see the LICENSE file. src/skill/mattpocock-grilling.md is vendored separately under its own MIT license (Matt Pocock, full text kept in the file itself).

Resolve ambiguity before planning. Ground the plan in the real repo. Make the proposal reviewable. Require human approval before execution. Keep runtime evidence separate from claims of correctness. Then:

/readyset