kud/ask-opencode

Available models

Call the list_models tool from the opencode MCP server to get the current list of allowed models at runtime.

Step 1 — Resolve the prompt and model

If the user invoked the skill with an argument (e.g. /ask-opencode <question>), use that argument verbatim as the prompt.

If no argument was given, ask the user what they would like to send to opencode before proceeding.

If the user specifies a model (e.g. /ask-opencode --model github-copilot/gpt-4o <question>), use that model. Otherwise use the MCP server's default (call list_models if unsure what's available).

Step 1.5 — Inject repo context if relevant

If the prompt references "this repo", "this project", "here", "audit", or similar context-dependent language, gather the following before sending and prepend it to the prompt:

• Current working directory (absolute path)
• Git repo root name and current branch (git rev-parse --show-toplevel, git branch --show-current)
• Top-level file/directory listing
• Contents of key config files if present (e.g. package.json, CLAUDE.md, README.md — truncated to ~50 lines each)

Prepend this as a fenced block labelled ## Repo Context before the user's prompt so opencode has accurate grounding.

If the prompt is a general question with no repo reference, skip this step entirely.

Step 2 — Send the prompt

Call the query tool from the opencode MCP server with the resolved prompt (including any injected context) and model.

Step 3 — Present the response

Display the response under a clearly labelled heading:

### opencode (<model>)

<output>

Do not paraphrase, edit, or interpret the output — show it exactly as returned.

Step 4 — Offer your own perspective (optional)

After presenting the opencode response, offer to share your own take or highlight any meaningful differences. Only if it adds value — if the user did not ask for a comparison, keep this to a single brief sentence.