Switchyard

Switchyard is a local Python CLI that makes an existing Claude Code + Codex handoff workflow executable: one command advances the right agent through the right step, writes durable artifacts, and stops at defined gates. It does not invent a new SDLC or run an uncontrolled agent swarm.

Quick Start

1. Prerequisites

• Python 3.13+ (python --version)
• Node.js (LTS) — both agent CLIs ship via npm (node --version)
• Codex CLI — npm install -g @openai/codex, then codex login
• Claude Code CLI — npm install -g @anthropic-ai/claude-code, then claude login

The standalone CLIs are separate from the VS Code extensions; Switchyard shells out to the CLIs. Confirm both run non-interactively before using Switchyard:

codex exec "say hi"
claude -p "say hi"

2. Install

From the repo root:

pip install -e .

This registers the switchyard command on your PATH. The install is editable, so changes to the source take effect immediately — you only re-run this on a new machine or virtualenv.

3. Run it in any repo

cd <target-repo>
switchyard run "<task>"

• The current directory is the target repo by default; use --repo <path> to point at a repo from elsewhere.
• Cap review rounds with --max-rounds N (default 3).
• Artifacts are written under the target repo's .switchyard/ directory — add .switchyard/ to that repo's .gitignore if you don't want them tracked.
• The run stops at a finished plan (final-plan.md) or a blocked-report.md. It does not modify code, run tests, or commit.

What Switchyard Does

switchyard run executes a bounded planning loop:

task packet
-> Claude authors a plan
-> Codex reviews it against the repo and decides (approved / needs_revision / blocked)
needs_revision -> Claude refines -> Codex reviews again
-> until approved, blocked, or max rounds
-> final-plan.md (approved) or blocked-report.md

Roles differ by phase:

• Planning (implemented): Claude authors and refines the plan; Codex reviews it against the repo and owns the approve/revise/block decision.
• Implementation (future, Phase 3+): Codex implements the approved plan; Claude reviews the diff.

Files are the source of truth: the full audit trail is .switchyard/runs/<run-id>/, and the latest artifacts are mirrored to .switchyard/handoff/active/ (cleared at the start of each run).

Commands

Planning loop (implemented):

switchyard run "<task>" [--repo <target-repo>] [--max-rounds N]
-> .switchyard/runs/<run-id>/
metadata.json
01-task-packet.md
02-claude-plan-round-1.md
03-codex-review-round-1.md
... # more rounds only if needs_revision
final-plan.md | blocked-report.md

Adapter spike (Phase 0 probe):

switchyard spike-adapters "<problem statement>"
-> runs/<run-id>/
00-task-packet.md
01-codex-plan.md
02-claude-review.md
adapter-notes.md

Neither command implements code, runs tests, commits, pushes, or opens PRs.

Current Status

Implemented and verified: the Phase 0 adapter spike and the bounded planning loop (Phases 1 + 2a + 2b). The next build target is Phase 3 — Codex implements an approved final-plan.md on a Switchyard branch with a git boundary and test execution — followed by Phase 4, Claude final diff review.

Documentation Map

• design/switchyard_mvp_addendum.md narrows the MVP into phases and holds the ratified loop model.
• design/active-checklist.md tracks the current implementation queue and next work.
• design/implementation_skeleton.md records the initial Phase 0 package skeleton.
• design/reference/task-packet.md, design/reference/run-artifacts.md, design/reference/workflow-states.md, design/reference/configuration.md, and design/reference/guardrails-and-reporting.md hold durable workflow reference details.
• design/reference/cli/codex-cli-reference.md and design/reference/cli/claude-cli-reference.md capture CLI reference findings.
• design/phase-0-cli-probe-findings.md records Phase 0 adapter probe results.
• design/archive/switchyard_design_draft.md preserves the original source design draft.
• design/manual-instructions/codex-instructions-reference.md and design/manual-instructions/claude-instructions-reference.md preserve the manual agent rules.

Design Principle

Switchyard keeps the core boring and reliable:

files as source of truth
metadata as workflow state
agent lanes behind adapters
bounded loops
explicit stop conditions
git as containment
human approval for risky transitions

That stable core can later support API-token adapters, concurrent runs, dashboards, PR automation, or additional model lanes without discarding the initial CLI work.