MCP prompts return workflow instructions for the client to follow. Requesting a prompt does not execute its suggested tools or approve an operation. The six existing prompt names and input schemas are unchanged.
| Prompt | Arguments | Guidance |
|---|---|---|
opencode-code-review |
sessionId |
Confirm project scope, inspect diffs, and provide evidence-based review feedback |
opencode-debug |
issue, optional context |
Inspect setup, choose a model, investigate, then propose or implement an authorized fix |
opencode-project-setup |
none | Read project context and key files; use a read-only OpenCode request when low-level file tools are not advertised |
opencode-implement |
description, optional requirements |
Dispatch implementation through run/fire, retain identifiers, resolve explicit input, review changes and tests |
opencode-best-practices |
none | Select tools/models, interpret async states, recover jobs, handle input, and understand scope and lifecycle |
opencode-session-summary |
sessionId |
Check current state, inspect bounded history and changes, and summarize completed and remaining work |
The implementation and debugging prompts retain the returned jobId, sessionId, and messageId. They use opencode_job_get, opencode_check, or opencode_wait to follow existing work. A timeout does not instruct the client to submit a duplicate prompt.
When work reports input_required, the prompt asks the client to present the pending permission or question and forward the user's explicit response. It does not recommend bypassing the request with blanket permission approval.
Use the absolute directory on the OpenCode server for relevant tool calls. These prompt schemas do not contain a directory argument; the client must use the project directory established in the conversation. Do not assume that a previous tool call changed the server's default project.
Provider/model IDs come from discovery or both configured environment defaults. The essential profile supports the common workflow tools, while specialist file/API tools require the full profile. Prompts should only invoke capabilities actually advertised to the client.
See examples for concrete tool calls and configuration for profiles and lifecycle behavior.