Say what you want done. Get a verified result at the boundary you requested.
Railyard helps Codex and Claude Code choose suitable models, reasoning effort, and workflows without turning every edit into a delivery ceremony. Routine work uses native tools and bounded subagents. A useful Compound Engineering skill activates when the work calls for it; fleet and account orchestration remain explicit choices.
- 🚦 Plain-language work. Native execution for routine changes, selected CE workflows for planning, debugging, review, and PR delivery when useful.
- 🧠 Deliberate allocation. Choose model and reasoning effort together. In Claude Code, Opus 5.5 at medium effort handles substantive subagent work, with Fable 5.1 for frontier-hard or long autonomous work, Sonnet 5.5 for well-scoped edits, bug fixes, and docs, and Haiku 4.5 for read-only search. In Codex, GPT-6.1 Sol at medium is the baseline, with Sol 6.1 at high for review, Luna for routine checked work, and Astra only for a demonstrated gap. An omitted model inherits the parent's, and unsupported selections are disclosed.
- 🔍 One completion owner. CE owns review settlement and CI watching for a selected PR workflow. An explicit Deliver implementation/fix request runs through merge, required release or deployment, and consumer verification; explicit narrower requests keep their stated endpoint.
- 🖥️ Explicit fleet work. Orchestrates across machines when requested, and only dispatches to hosts its sibling roundhouse has verified are ready.
claude plugin marketplace add novotnyllc/marketplace
claude plugin install railyard@novotnyllc
# then just say: "set up railyard"(Codex: codex plugin marketplace add novotnyllc/marketplace and
codex plugin add railyard --marketplace novotnyllc.)
Read the public documentation — what it can do, how it works, and what a delivery actually looks like.
Freight doesn't move itself — the yard receives it, sorts it, assembles the train, clears the departure, and confirms arrival:
| Track | Skills |
|---|---|
| Setup & health | setup — inspect and configure relevant prerequisites; doctor — diagnose and fix drift |
| Allocation | model-routing — native model and reasoning-effort guidance for Claude Code and Codex |
| Semantic decisions | jev — model/effort selection, workflow choice, evidence selection, work priority, and review triage, consulted only when a choice is genuinely open and matters; requires TYPESAFE_API_KEY |
| Delivery | deliver — select native execution or one useful CE workflow and complete the authorized scope |
| Orchestration & placement | orchestrate — explicit fleet/account work, cross-project coordination, and supported remote placement |
| Specialist review | thermos, thermo-nuclear-review, thermo-nuclear-code-quality-review — when their perspective is useful; oracle — a manual GPT-6 Pro second opinion through the ChatGPT browser |
| Optional audit | audit — explain a run's decisions, outcomes, and resource use when requested or useful |
| Runtime repair | cleanup-codex — inspect or repair a concrete process problem |
Railyard uses the external
Compound Engineering
plugin for selected ce-* workflows. Use ce-commit-push-pr when creating a
PR or pushing user-requested commits to an existing PR. The selected CE
workflow owns review, feedback resolution, CI, and its watch loop; Railyard
does not add a second watcher or mandatory reviewer.
LFG remains available for an explicitly selected full workflow. Routine work does not require it, a delivery contract, a retrospective, or process cleanup. Startup performs no dependency bootstrap. Install CE through the plugin manager when needed:
claude plugin marketplace add EveryInc/compound-engineering-plugin
claude plugin install compound-engineering@compound-engineering-pluginPonytail and Superpowers are not Railyard prerequisites. Native subagents are the default for bounded parallel work; user-visible tasks are created only when the user explicitly asks for them.
TypeSafe Jev selects among eligible model/effort
pairs, available workflows, relevant evidence, ready work, and next review
investigations. Railyard consults it only when such a choice is genuinely open
between eligible options and matters for cost or outcome, skipping explicit
choices, clear defaults, single options, and small dispatches, and reuses an
answer already reached in the session. An auth or rate-limit failure stops
further consultation for the rest of the session. TYPESAFE_API_KEY must be
present, or Jev is inactive; no extra installation is needed. Use --offline
to disable the helper; missing credentials, service errors, and uncertain
judgments fall back to ordinary Railyard reasoning.
Jev receives only the context and candidates explicitly supplied to its helper. It leaves permissions, model choice, execution, and CE review settlement with their existing owners. See the Jev skill for examples and the request contract.
Machine and infrastructure administration lives in public fleet docs. The public product story, scenarios, and guides live at novotnyllc.github.io/railyard. Charter and boundaries: AGENTS.md.
MIT — see LICENSE.
Portions are adapted from steipete/oracle and cursor/plugins (both MIT),
with steipete/agent-scripts (MIT) reviewed alongside. Their copyright
notices are preserved in LICENSE and the incorporations are
itemized in THIRD-PARTY-NOTICES.md.