A skill library for orchestrator → subagent work

You just talk.
It orchestrates.

orchflows turns Claude Code, Codex, or Grok Build into an orchestrator — subagents that work in parallel, real verification of everything they produce, and workflows that improve themselves.

Install Read the docs

No skill names to memorize · two callables do every job

orchflows

$ ship dark mode # one root ticket that cuts itself into tickets ├─ parallel subagents ├─ one review pass └─ final verification against the ask external check passed — done

The problem

Skill libraries break four ways.

orchflows answers each in structure, not with a bigger prompt file. Full rationale in Design rationale.

01 · Over-specific

Handcrafted prompts

As models get smarter, overprescription degrades output quality.

↳ Tiny composable skills where each word fights for its life.

02 · Manual

Chained by hand

Run /brainstorm, then /to-prd, then… you shepherd every hop.

↳ Autorouting instead of memorized chains.

03 · Domain-locked

One workflow per domain

Writing a blog and shipping a feature are the same shape.

↳ Domain-blind workflows retargeted by data standards.

04 · Static

Run 100 hits the same snags

Nothing learns. The hundredth run stubs the same toe as the first.

↳ Self-improvement wired into every run.

The mechanism

Two callables. One return path.

Control flow — loops, branches, retries — is the calling workflow's own prose, not a new callable. Detail in Workflows.

orch-do produces

Builds each unit of work; in planning mode, freezes the root ticket before any child starts.

orch-judge challenges

The read-only twin. It attacks Goal and evidence — adversarial review stands between a claim and "done".

Nothing marks "done" except an external check passing. Goal, executor evidence, and the applicable independent path all sit between a claim and completion.

Every ticket attempt closes through one dispatch-outcome envelope — the single path every return crosses. tickets.py do is the whole outbound half; tickets.py land imports, adjudicates, and retires. Full lifecycle in Tickets.

Four layers · data, never control flow

Retarget the pipeline by stamping a standard.

A narrow waist of seven data contracts sits between everything; the domain switch is pure prose, never a code fork. See Architecture and Standards.

L3 example-workflows/ Named workflows, callable like any skill.

L2 standards/ Per-domain data — code, content, research, design, data — never control flow.

L1 skills/ Everything callable: kernel primitives and domain-blind workflows.

L0 contracts/ The narrow waist — pure data shapes, the only interface between everything above.

StandardDeliversEvidence

orch-codecode · tests and checksgit, one worktree per work item

orch-contentdocumentsartifact evidence

orch-dataanalysesreproduction evidence · datasets pinned by digest

orch-designrendered UIcapture evidence per view × breakpoint × state

orch-researchanswerssource evidence · lane packets

Documentation

Browse the library.

The docs mirror the repository's own vocabulary and reading order — not a parallel rewrite.

OverviewReading order by role, and what lives where in the tree. InstallationInstall, update, uninstall, and the per-host surfaces. ConceptsThe vocabulary — every term, defined once. ArchitectureFour tiers, the narrow waist, and dependency direction. StandardsFive domains, three projections, and narrowing. TicketsAnatomy, lifecycle, dispatch protocol, and review. WorkflowsCallables, routing lanes, self-improvement, tournaments. Design rationaleWhy it is shaped this way, and the reader it serves.

Install

Clone, install, talk.

A user install that configures whichever CLI it finds. Full guide in Installation.

$ git clone https://github.com/DanMcInerney/orchflows
$ cd orchflows
$ ./install.sh # install.cmd on Windows

WindowsmacOSLinuxClaude CodeCodexGrok Build

To update: git pull, rerun. · Uninstall: python install.py --user --uninstall

You just talk.It orchestrates.