菜单
Use when creating, editing, or refactoring custom OpenCode agent definitions (`.md` or `.md.tmpl` files in an agents directory, including chezmoi source forms). Triggers on any edit to an agent definition file. Do NOT use for AGENTS.md, SKILL.md, or slash commands.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Use when operating on the Linear issue tracker via the `linear` CLI: creating, viewing, searching, querying, or updating issues, projects, or documents; adding or editing issue comments; managing relations, links, labels, and teams; assigning, labeling, or transitioning issues. Triggers on phrases like "create a Linear issue", "file a ticket", "update ENG-###", "move to Ready For Dev", "search for issues", "find issues about X", "add a comment to the Linear issue", "link a PR to the issue", "block ENG-42", or any task naming Linear, a Linear issue key (`ENG-`, `OPS-`, etc.), a Linear project, or a Linear document. Do NOT use for GitHub Issues, Jira, or other trackers.
Use when creating, editing, refactoring, or reviewing AGENTS.md files at any level (repo, package, or global). Triggers on phrases like "update AGENTS.md", "add a rule to AGENTS.md", "global directives", or any edit to a file named AGENTS.md (including chezmoi source forms). Do NOT use for SKILL.md, agent definitions, or slash commands.
Use when creating, editing, refactoring, or reviewing OpenCode custom slash commands; writing or revising command frontmatter, arguments, agent routing, or permissions. Triggers on phrases like "new slash command", "add a /command", "edit the command file", or any edit to files in a commands directory (including chezmoi source forms). Do NOT use for AGENTS.md, SKILL.md, or agent definitions.
Use when creating, editing, refactoring, or reviewing SKILL.md files and skill directories. Triggers on any edit to a SKILL.md file or skill directory (including chezmoi source forms). Do NOT use for AGENTS.md, agent definitions, or slash commands.
Use when reading, posting, or managing PR review comments via the `gh-review` tool: viewing PR comments with filtering, leaving inline comments on specific diff lines, starting or deleting pending reviews, replying to review threads, editing or removing individual review comments. Triggers on phrases like "review this PR", "check for comments", "leave a comment on line X", "add review feedback", "start a pending review", "reply to the comment", "edit the comment", "remove that comment", "delete that comment", "fix my comment", "move comment to line X", or any task involving PR comments and code review. Do NOT use for merging, approving via `gh pr review --approve`, or non-review PR operations.
Use when creating, editing, refactoring, or reviewing modularized Python CLI scripts managed by uv with pyproject.toml; scaffolding new script projects; adding commands or subcommands to Click-based CLIs; writing pytest tests for CLI tools; configuring hatchling builds or dependency groups; creating wrapper shell scripts for uv-managed projects. Triggers on phrases like "new python script", "add a CLI command", "scaffold a script project", "python CLI", "click command", or any work in a scripts/ directory containing pyproject.toml with hatchling. Do NOT use for single-file scripts, Jupyter notebooks, web applications, or Django/Flask projects.
| name | subagent-authoring |
| description | Use when creating, editing, or refactoring custom OpenCode agent definitions (`.md` or `.md.tmpl` files in an agents directory, including chezmoi source forms). Triggers on any edit to an agent definition file. Do NOT use for AGENTS.md, SKILL.md, or slash commands. |
Conventions for OpenCode agent definitions. Omissions intentional. All authored content MUST follow the Authoring rules in global AGENTS.md.
Specialized personas with models, prompts, tools, and permissions. Primary agents are user-facing; subagents are invoked via the Task tool.
Markdown files in .opencode/agents/<name>.md (project) or ~/.config/opencode/agents/<name>.md
(global). JSON alternative exists under agent in opencode.json with a prompt field; prefer
markdown.
name (string): agent identifier (defaults to filename)description (string): drives caller delegation (required; see below)mode (string): primary, subagent, or all (default: all)model (string): provider/model identifiervariant (string): reasoning variant (see below)prompt (string): system prompt (or @./path.txt to include file)hidden (boolean): hide from @ menu (subagent only)Tool-level and skill-level access; values allow, deny, ask. Bash and task permissions support
glob patterns; last matching rule wins.
permission:
edit: deny
bash:
"*": ask
"git diff": allow
"git log*": allow
webfetch: deny
skill:
"*": allow
"internal-*": deny
task:
"*": deny
"my-subagent-*": allow
The deprecated tools field converts to permissions internally; prefer permission.
Same trigger-coverage discipline as skill descriptions (see skill-authoring). Routing
reinforcement lives in the caller's AGENTS.md, not the subagent's own file.
Subagent descriptions additionally MUST specify:
Primary agent descriptions are looser; surfaced in TUI for humans, not programmatic routing.
steps (integer): max agentic iterations (maxSteps is deprecated)disable (boolean): set true to disabletemperature (number): LLM randomness (0.0-1.0)top_p (number): response diversity (0.0-1.0)color (string): #hex or theme name (primary, accent, error)low, medium, high, max (adaptive)high, max (fixed token budget)none, minimal, low, medium, high, xhighlow, highVariant takes highest priority in the options merge chain. Unknown frontmatter fields pass through
as provider-specific model options.
Set hidden: true on subagents that should only be called by other agents.
Cover (not all required): workflow/prerequisites, domain ownership, hard constraints (RFC 2119; see
agents-authoring), verification commands, output format, when stuck. Do not restate constraints
already enforced by permissions.
Include tool command signatures the agent needs to act without discovery calls. Standalone subagents MUST have reference inline; when a companion skill exists, reference by name instead.
The test: can the agent do its job from the prompt alone? If not, add the missing reference.