Archived. This branch is the last state of the 0.1.x compiler, which lowered a markdown spec into GitHub Agentic Workflows. It is kept for the record and receives no changes. From 0.2.0 the in-lockstep name is a different product, the framework on main; ADR 0001 records why. The in-lockstep-exec distribution and the pipeline-exec image this line published are withdrawn.

Compiles pipeline definitions — commands, agents, guardrails, skills, contexts, profiles — into GitHub Agentic Workflows.

The markdown spec stays the single source of truth. Lockstep lowers it onto GitHub-native primitives: orchestration becomes plain Actions YAML (jobs, needs:, matrix fan-out), and each agent becomes a gh-aw agentic workflow. A drift gate recompiles on every pull request, so committed output can never silently diverge from the spec.

spec (commands/ agents/ guardrails/ skills/ contexts/ profiles/ mcp/ + pipeline.yaml + overlays/)
        │  lockstep compile
        ▼
.github/workflows/<command>.yml      orchestrators (plain Actions)
.github/workflows/aw-<agent>.md      agentic workflow sources
.github/workflows/shared/*.md        flattened prompt layers
        │  gh aw compile
        ▼
.github/workflows/aw-<agent>.lock.yml   what actually runs

Getting started walks through building a real pipeline against a live public API, and explains each part of the framework as you meet it — including what happens once the output is hosted on GitHub: how changes get reviewed, where output is stored, and how reports survive long enough to show a trend.

uv tool install in-lockstep         # the compiler

The runtime, in-lockstep-exec, is installed by compiled pipelines rather than by you — jobs run it as the container image, and only a job that materializes inherited definitions installs the package.

lockstep compile                    # generate workflows
lockstep compile --check            # drift gate: verify committed output matches the spec
lockstep compile --semantic-diff    # report security and cost surface deltas
lockstep show-surface               # every GitHub-target decision in one document

A compiled pipeline references three things, and only the first is a dev dependency:

actions/ and pipeline-exec share this repository with the compiler because the compiler emits references to both as literal text: tests/test_contract.py parses every emitted invocation against the real CLI and every input against the real action, so a renamed flag fails a build rather than a scheduled run. A pipeline points at wherever you published them:

capabilities:
  actions: github.com/<owner>/<repo>@v1.0.0    # resolved to a commit by `lockstep pin`
  exec-image: quay.io/<owner>/pipeline-exec    # any registry; resolved to a digest

Both are published, and the examples here pin them for real:

capabilities:
  actions: github.com/in-lockstep/lockstep/actions@actions-v0.1.1   # -> aad2f112…
  exec-image: ghcr.io/in-lockstep/pipeline-exec                     # -> sha256:70de3f80…

lockstep doctor reports no findings on them, where it used to report DOC015 on every one. The basic test fixture still pins placeholders deliberately, because something has to keep exercising what the compiler does when a capability is unpinned.

Extending the framework covers the two extension points — third-party builtins in pipeline-exec, and your own composite actions — worked through a pipeline that fixes bugs from an issue tracker and opens pull requests.

Implementing an issue by review builds a pipeline that turns an issue into a pull request, and lets reviewers revise the plan or the code with ordinary PR comments plus a slash command.

Publishing a report to GitHub Pages builds a triage-report pipeline, and uses it to look closely at how context, guardrails and skills shape what an agent produces.

Reviewing pull requests on request builds a /review security intent bot — one review per aspect, revised in place, and silent when nothing has changed.

Adding a pipeline to a repository you already have covers adoption into an existing project with its own CI, and the security model for pull requests from forks.

What goes where is the rule that keeps application knowledge out of the framework: which tier code belongs to, and which of the three prompt layers a piece of prose belongs in.

Inheriting a pipeline from another repository walks through one team owning the standards and many repositories inheriting them: a consuming repository writes four files, gets guardrails it cannot weaken, and reviews an upstream change as a diff of the prompt text.

Sharing pipelines across an organization is the design behind that: which layers travel, what a consumer may tune, and how an inherited pipeline opens its own bump pull request when upstream moves.

What an eval case promises is the grading contract: which expectations a machine settles, which need a model, and why a case carrying a rubric is never reported as passed.

What the people who would adopt this need next is the backlog derived from five adopter personas rather than from the design: what gates a public launch, what a security lead has to write themselves today, and what the run ledger measures that a director is not asking about.

This repository compiles its own drift gate covers self-hosting: why the gate installs the checkout rather than a release, why the hand-written CI workflow stays hand-written, and the defects dogfooding surfaced.

lockstep init --name=my-pipeline    # scaffold a working pipeline
lockstep pin                        # resolve capability tags to commits
lockstep compile                    # generate the workflows
lockstep compile --check            # drift gate: committed output must match the spec
lockstep lint                       # is the spec well built?
lockstep doctor                     # will GitHub accept it?
lockstep show-surface               # every target decision in one document
lockstep eject <file>               # take ownership of one generated file

All seven phases of the design are implemented. See docs/status.md for what each covers and what remains open — chiefly round-trip evals across both backends, which need pipeline-framework.

make check      # format, lint, typecheck, test