A CI/CD pipeline comprehension and migration tool. Open any of 17 platforms' pipeline definitions — GitLab CI, GitHub Actions, Jenkins, Azure, CircleCI, drone, bitbucket, buildkite, Tekton, Argo, Travis, woodpecker, AWS CodeBuild/CodePipeline, Google Cloud Build, Earthly, dagger — get back the same neutral Hub-IR, and from there inspect, render, document, run locally, migrate to another platform, or compose with reusable recipes.
Built on a Triple Graph Grammar kernel
(seesaw-tgg): each
platform's rule set is generated from a declarative construct
catalog and compiled into both directions, so parsing and emitting
can never drift — every transformation is rule-based, bidirectional,
and round-trip-verified.
17 surface syntaxes one neutral graph any projection
┌──────────────────────┐ ┌──────────────────────┐ ┌─────────────────────────┐
│ .gitlab-ci.yml │ forward │ Hub-IR │ │ inspect (JSON) │
│ workflow.yml │ TGG │ hub:pipeline │ ───▶ │ render (SVG / runbook) │
│ Jenkinsfile │ ──────▶ │ ├─ hub:job … │ │ capabilities profile │
│ azure-pipelines.yml │ │ │ └─ hub:step … │ │ run locally (Docker) │
│ .drone.yml │ ◀────── │ └─ provenance │ │ export (md/html/doc) │
│ Earthfile … │ backward│ (byte spans) │ └─────────────────────────┘
└──────────────────────┘ TGG └──────────────────────┘ │
▲ │ re-key / synthesize │
└─────────────── migrate ◀────────┘ (all 272 ordered pairs) ◀──┘
The same Hub-IR feeds the CLI (pipewright), the Qt6 desktop
app, and a C-ABI (pipeline-ffi) for third-party hosts. The
rule sets come from catalog/ — per-platform construct inventories
plus one neutral field mapping, from which the bidirectional TGG
rules are generated.
# Inspect any CI definition — auto-detects the platform (17 supported)
pipewright inspect ./.gitlab-ci.yml | jq '.pipeline.jobs[].name'
# What would run, in what order? (no Docker needed)
pipewright plan ./.gitlab-ci.yml
# Run it locally in Docker, streaming output (container-shell platforms)
# Read-only mount by default; --rw-copy / --rw opt into writes.
pipewright run ./.gitlab-ci.yml --job test
# How portable is it?
pipewright capabilities ./.gitlab-ci.yml
# Migrate — any platform to any platform, even across structural families
pipewright migrate ./.gitlab-ci.yml --to github > workflow.yml
pipewright migrate ./.drone.yml --to azure > azure-pipelines.yml
# Generate a human-readable runbook (en/de)
pipewright render ./.gitlab-ci.yml --format md --locale de
# Compose a pipeline from reusable recipes
pipewright compose --to gitlab recipes/rust-ci.recipe.ymlOr the desktop UI — jobs list, editable DAG diagram, capability profile, migration with friction report, recipe browser/composer, local runs, exportable runbook:
| About | Vision, the TGG kernel, the construct catalog, Hub-IR, capability and friction concepts |
| Install | Build from source, dependencies, packaging |
| Quickstart | Five-minute CLI tour with real outputs |
| User manual | Full reference, organised by role: every UI tab, every CLI subcommand, recipes, the FFI surface |
| Evidence | The cross-platform interop matrix in docs/interop-matrix.md |
catalog/ The declarative heart: per-platform construct
inventories + ir.toml field mapping + generators
that emit the per-platform TGG rule sets
crates/ 11 Rust crates, one workspace
├── pipeline-cst YAML CST with byte-accurate rebuilt source
├── pipeline-jenkinsfile-cst Jenkinsfile (Groovy DSL) → same Document shape
├── pipeline-earthfile-cst Earthfile → same Document shape
├── pipeline-tgg-seeder Catalog-driven CST → typed-graph seeding + emit
├── pipeline-forward forward / re-emit / migrate / edit entry points
├── pipeline-hub-ir Hub-IR read API (jobs, steps, attrs, provenance)
├── pipeline-render Model lift + SVG diagram + localized runbook prose
├── pipeline-recipe Port-typed recipe registry, compose & apply
├── pipeline-cli The `pipewright` binary (13 subcommands)
├── pipeline-ffi C-ABI bridge with cbindgen header
└── chaos-generator Schema-walking pipeline generator (test harness)
ui/qt6/ Qt6/QML desktop app (consumes pipeline-ffi)
recipes/ Standard recipe library (rust-ci, go-ci, …)
tests/cross_corpus/ Real-config corpus, 2 fixtures × 17 platforms
docs/ user/, interop-matrix
.github/workflows/ CI (lint, test, gate-sample, supply-chain) + release
| Area | State |
|---|---|
| Platforms (parse / forward / backward) | 17, all bidirectional |
| Round-trip verification | real-config corpus 59/59 · chaos stress 17×30 seeds, 0 failures · byte-identical CST round-trip (full sweep is a nightly CI gate; a fast sample gates every push) |
| Cross-platform interop | all 272 ordered pairs faithful + stable (matrix) |
| Local Docker runner | live for the 11 container-shell platforms — mounts the repo (read-only by default; --rw-copy/--rw opt into writes), passes env, evaluates rules:if/when, starts services: as sidecars; streamed logs. k8s-CRD / service-orchestration / code-defined platforms are translate-only (the tool says so, doesn't pretend) |
| Migration friction report | derived (re-parse + capability diff), not declared — surfaced in the UI and on the CLI |
| Recipes | standard library + user sources (dir/git), port-aware apply/compose |
| Qt6 desktop UI | 7 tabs incl. editable DAG; en/de; headless smoke test (nightly CI gate) |
| Workspace tests | 300 passing + heavy gates on schedule |
| CI | GitHub Actions — fmt + clippy, tests, round-trip sample gate, MSRV, supply-chain audit on every push; full gates + UI smoke nightly |
Pipewright was developed with heavy AI pair-programming. Every claim in this README is backed by a test or a gate that runs in CI — the round-trip numbers, the platform coverage, the run behaviour. The TGG rule sets are generated from the declarative catalog, not hand-written. If something here doesn't match what the code does, that's a bug; please open an issue.
Apache-2.0. See LICENSE.