What problem does it solve? Developing and verifying the hermes-opencode plugin (the bridge between Hermes and a headless opencode server) is error-prone: the plugin tools live in Hermes' deferred tool catalog, the opencode v1 HTTP/SSE surface has directory-scoping and version-drift traps, and event-driven behaviors (turn-complete injection, question asks, permission routing) fail silently in non-interactive harnesses. This Skill encodes the verified contracts, live-probe recipes, and known failure shapes so you can build, upgrade, and e2e-test the plugin without rediscovering each pitfall. ## Core Features & Use Cases - End-to-end plugin driving: Invoke opencode_prompt, opencode_session_tail, opencode_session_read, opencode_question_reply, and opencode_command through the deferred catalog (tool_search → tool_describe → tool_call) against a live opencode serve on 127.0.0.1:4096. - Build-to-standard recipes: Fail-soft register(), manifest v2 with config_schema, minimum-version discipline, and hermes plugins doctor . --ci as the authoritative build check. - Failure-shape diagnostics: Distinguish paraphrased question asks, tool{} stalls, wedged que_ ids, SSE location-filter starvation, and inject silent no-ops, with REST curl probes to verify each independently. - Use Case: A delegated opencode turn is stuck busy — use the Skill's tail-reading and GET /question / GET /permission recipes (with the exact x-opencode-directory header) to determine whether a permission ask was denied or a question ask is pending, then reply or abort accordingly. ## Quick Start Ask the agent to smoke test the hermes-opencode plugin by sending a prompt through opencode_prompt, reading the session tail, and confirming completion via the live server status endpoint.
