opencode-dev

Author OpenCode (SST's open-source coding agent, opencode.ai — NOT OpenAI Codex) agents, custom commands, and project rules as markdown/JSON config. Use when creating or editing files under `.opencode/agents/`, `.opencode/commands/`, or `AGENTS.md`; setting an agent's `mode` (primary/subagent/all), `model`, `permission`, `steps`, or `temperature`; templating commands with `$ARGUMENTS`, `$1`, shell injection `` !`cmd` ``, or `@file` references; wiring the `instructions` config key; or debugging why a `tools:` field, `maxSteps`, a `/docs/modes/` link, or a `CLAUDE.md` fallback isn't behaving. Triggers: "opencode subagent", "opencode custom command", "AGENTS.md", "opencode agent frontmatter".

Build a Claude Code plugin that delegates coding work to OpenCode (SST's open-source agent, opencode.ai) — the OpenCode analog of openai/codex-plugin-cc. Maps the codex-plugin-cc template (thin bash forwarders → a Node `.mjs` companion holding a persistent connection to the OpenCode daemon → a detached background-job store with status/result/cancel) onto OpenCode's three drive paths. Use when building a CC→OpenCode delegation/handoff plugin, porting codex-plugin-cc to OpenCode, or choosing between `opencode acp`, `opencode serve` + `@opencode-ai/sdk` (`createOpencodeClient`), and `opencode run --format json --attach --dangerously-skip-permissions` as the transport for delegated tasks.

Author OpenCode plugins and hooks — JS/TS modules that hook into OpenCode's lifecycle to block tools, mutate args, inject env, react to events, or register custom tools. Use when building or debugging anything under `.opencode/plugins/`, importing `@opencode-ai/plugin`, writing a `tool.execute.before` / `tool.execute.after` / `permission.ask` / `shell.env` / `command.execute.before` handler, subscribing to OpenCode events via the `event` handler (`session.idle`, `command.executed`), registering a custom tool from a plugin, or looking for an OpenCode "stop hook" or a settings.json hook table. Distinct from OpenAI Codex.

Author OpenCode enterprise governance — the `experimental.policies` provider allow/deny layer and the `permission` per-tool gate — for `opencode.json` / `.opencode/`. Use when configuring which LLM providers OpenCode may use, migrating off `disabled_providers`/`enabled_providers`, locking down a fleet, or setting tool permissions (`allow`/`ask`/`deny`) for `bash`, `edit`, `read`, `external_directory`, `doom_loop`, etc. Trigger on mentions of OpenCode policies, `experimental.policies`, `provider.use`, `permission` config, OpenCode `.env` access, doom-loop / external-directory guards, per-agent permission overrides, or "restrict OpenCode to Anthropic only".

Drive OpenCode (the SST coding agent, opencode.ai) programmatically — the @opencode-ai/sdk JS/TS client, the github.com/sst/opencode-sdk-go Go client, and the `opencode serve` HTTP/OpenAPI server it talks to. Use when building an integration on top of OpenCode, calling createOpencode / createOpencodeClient, running `opencode serve`, hitting the OpenAPI spec at /doc, listing or prompting sessions over HTTP, wiring OPENCODE_SERVER_PASSWORD auth, requesting structured (json_schema) output, or using opencode.NewClient / Session.List / the Field param wrappers in Go. NOT OpenAI Codex; NOT the opencode.ai/docs/go Zen subscription.

1 hook across 1 event