OpenCode Code Relay Plugin

OpenCode plugin that automates the Code Relay protocol: cross-session and cross-repo state handoff for AI coding agents. It turns the manual HANDOFF / CHECKPOINT routine into single commands plus a session hook, so context survives across sessions and repositories without the agent starting from zero.

Concept based on yan5xu/code-relay. Licensed MIT.

Why

AI coding agents lose context across sessions and across repositories. Code Relay solves this with a file-based handoff protocol, but it relies on the developer remembering to write HANDOFF docs. This plugin automates capture, cross-repo awareness, and restoration.

Install

# from npm
opencode plugin add opencode-code-relay-plugin
# or locally (during development)
opencode plugin add .

Commands

• relay init — create relay.json at the workspace root and capture project context.
• relay save — snapshot the current repo to .code-relay/state.json, update the root state, run Spec Sync, and render RESUME.md.
• relay close — Closing Ritual: consistency check → delta snapshot → next_step.md → atomic commit with a session-end tag.
• relay switch <repo> — move the active baton to another repo and load its specs/ as context.
• relay resume — print the current repo's RESUME.md.
• relay status — condensed view of relay.json.

How it works

• Relay Manager — relay.json at the workspace root holds project_context, active_baton, and a per-repo map. Per-repo snapshots live in .code-relay/state.json.
• Spec Sync — deterministic diff of specs/**/*.spec.md (no LLM) compresses intent into the root state and powers the relay close consistency check.
• Context Injector — renders a Resume Prompt from state + handover templates, injected automatically via the experimental.session.compacting hook and on demand via relay resume.

State model (hybrid)

• relay.json — machine-readable global state (workspace root).
• <repo>/.code-relay/state.json — per-repo snapshot.
• <repo>/.code-relay/RESUME.md — rendered, agent-readable resume prompt.
• <repo>/.code-relay/next_step.md — next session's first action.

Non-goals

LLM-based spec summarization, GitHub Collaboration Mode, multi-agent swarm orchestration, and a GUI are deferred.

License

MIT — see LICENSE.