Skip to content

y-b-1/orchestra

v3.6.2

Deliver multi-request prompts fast: every independent ask in parallel, one independent review, /goal, and a guard that blocks merging into the default branch, a second Fable critic, untyped agents during a run, and a push before the test gate passes.

build

Gives the deeper builder rules and the mechanical, lever, migrate, cleanup, sensitive and repair modes. Load when you are a builder and your brief names a mode, or a unit fails its checks twice. Not for a routine first build that passes.

critique

Gives the rules for challenging a spec. Load when you are the critic. Not for builders, reviewers or investigators.

decide

Settles a choice the ask leaves open, interviews the owner, keeps the glossary and decision records, and compares design shapes. Load when two asks contradict, an ask has two valid product shapes, an owner suggestion would break existing behavior, the owner asks to be grilled or interviewed, a new term or interface needs a name, or the owner wants a prototype or a questionnaire for someone else. Not for choices the ask already settles.

diagnose

Gives a loop for finding the cause of a bug, fixing the root cause and proving it, plus correct mode for an owner correction. Load when a bug's cause is unknown, a test fails for an unclear reason, a fix did not hold, or the owner corrects a mistake. Not for an ask that already names the exact fix.

frontend-proof

Gives the proof and accessibility rules for interactive UI. Load when building a new interactive component (dialog, menu, drawer, form, keyboard flow) or when the ask is about accessibility. Not for copy, color, spacing or label changes.

investigate

Gives investigation rules by question type: files per ask, a how walkthrough with a placement answer, a why answer from git history and PRs, or research against primary sources. Load when you are an investigator or the main session answers a how, where, why or docs question. Not for builders (the cleanup builder may load why mode), and not for finding a bug's cause (use diagnose).

large-work

Gives a way to size, split, hand off and prove work too big for one builder. Load when an ask needs more than about 20 minutes of one builder's work, has open unknowns, runs long or overnight, the context nears its limit, or the owner asks to triage issues or PRs, compare candidate designs or write a handoff. Not for prompts that are long but made of small asks.

operate

Gives the operator rules for bounded long checks, reading failures and publishing work. Load when you are the operator. Not for builders running their own unit checks.

orchestra

Deliver a multi-request prompt fast by running every independent ask in parallel, then one independent review. Use whenever a prompt holds more than one ask (several changes, fixes, features, docs or tests), or when the user selects Orchestra or /goal. For the main session only; never loads in a subagent.

plan

Rules for the planner: a short spec, a product options note, a type and signature sketch, and a ticket plan with dependencies. Load when you are the planner on a planned ask or on a large feature whose shape is a real choice. Not for builders, reviewers or investigators.

review

Gives review lenses (correctness, data and contracts, architecture, standards, tests, security, accessibility, performance, and detailed lenses in references) and the rules every review follows. Load when you are a reviewer and your brief names one or more lenses. Not for a re-review with no lenses named.

ship

Load when opening, updating or fixing a pull request: red CI, review comments, merge conflicts, or a PR that must be easy to review. Also load when the owner asks to set up pre-commit hooks or git guardrails. Not for the final message of a run with no PR.

test-quality

Gives the rules for tests that can fail and the test-hygiene pass. Load when doing the hygiene pass before the PR or when writing tests that must be able to fail. Not for a routine unit that adds one behavior test.

verify

Load when a flow has many states and events (booking, payments, sync, multi-step forms), the repo has no way to run and drive the real app, a change touches a flow, route, command or UI the repo's driver covers, or the owner asks to audit or update the driver. Not for a change that needs no driver update.