Try first. Know what loaded. Adopt deliberately.
Spindle makes agent skills understandable, temporary when you want them temporary, exact at runtime, and reversible after adoption. Composition is one part of a complete evidence-bearing lifecycle.
Agent "skills" — the reusable instruction files a coding agent loads to do a job well — multiply fast. Left unmanaged, every repository inherits the same growing pile: skills it will never use, skills that quietly contradict each other, and a context budget spent on instructions instead of work. Spindle treats skills as behavioral dependencies you inspect, try, verify, and only then maintain—not directories you copy into an ambient pile and forget.
The ordinary workflow begins before installation. spindle inspect
explains a candidate without executing it. try creates one
disposable read-only session; borrow creates an expiring lease.
Startup verifies the exact desired set before the harness discovers it, and a
useful trial becomes durable only through explicit adoption.
Two minutes in—from inside your harness
# install from the public catalog in Codex
$ codex plugin marketplace add lyra-forge/marketplace
$ codex plugin add spindle@lyra-forge
› $spindle:spindle set up Spindle for this repository
# or install from the same catalog in Claude Code
$ claude plugin marketplace add lyra-forge/marketplace
$ claude plugin install spindle@lyra-forge
› /spindle:spindle set up Spindle for this repository
# after setup and a new session, use the project-local operator
› $spindle inspect a review skill for this migration.
Spindle reads live context · previews mutations · preserves foreign stateThe small operator does not duplicate the product manual. It reads a compact, machine-generated context contract containing current custody, conflicts, hook evidence, authorization boundaries, and exact next commands. The marketplace plugin carries a deterministic launcher for its exact pinned Spindle source, so no separate Python package install is required; Python 3.11 or newer is required. Claude and Codex use separate manifests and install commands, while sharing the operator and CLI implementation. Harness discovery and hook execution remain separate gates. The control loop is the product: intent → temporary authority → exact composition → startup verification → per-agent realization and activation → bounded evaluation → adopt, update, distill, release, or retire. See the lifecycle for the complete chain.
The shape of the idea
Why it exists →
The case against install-and-hope: ambient conflicts, stale instructions, unclear authority, context cost, and maintenance nobody owns.
The lifecycle →
Inspect, try, borrow, bootstrap, activate, evaluate, adopt, update, distill, roll back, and retire with exact custody.
How it works →
The record chain, ownership model, startup control loop, agent-local realizations, and composition pipeline.
Does it actually help? →
Availability, activation, routing, authorization, behavior, and adapter conformance are separate gates. Minimalism is deletion under evidence.
Use it →
Install the harness operator, try a candidate, manage startup evidence, adopt and update safely, then author distributions when you need them.
What it is, precisely
Spindle is a small, dependency-light Python control plane for behavioral dependencies. It keeps source, package, intent, temporary lease, durable adoption, surface binding, agent realization, activation evidence, and evaluation claims distinct. Packages and distributions remain ordinary pip-installable projects, but installation alone grants neither runtime nor maintenance authority.
The public base ships a minimal reference distribution under
examples/spindle-sample. Everything private — an organization's
own packages, project registries, task sinks, marketplace providers — lives in
separate adapter packages that depend on Spindle, never inside its core.
A skill should earn ongoing custody in the exact environment that will carry it.Temporary first; durable only with evidence.