ask-claude-code

name: ask-claude-code description: Delegate a self-contained coding or analysis subtask to Claude Code (Anthropic's agent) and get the result back — use to ask Claude Code to investigate, edit, or build something, to run a task through Claude Code via ACP, to continue a conversation with Claude Code or resume a Claude Code session across calls, or to get a second opinion from Claude on code.

ask-claude-code

Bridge that lets opencode hand a task to Claude Code and receive the answer. opencode cannot speak ACP (Agent Client Protocol) itself, so this skill wraps a small CLI that acts as the ACP client driving the claude-agent-acp agent.

opencode (you)  ──runs──▶  acp-client.mjs (ACP client)  ──ACP──▶  claude-agent-acp (Claude Code)  ──result──▶  back to you

Invoke

Run the CLI with the task as the prompt. Always pass --cwd "$(pwd)" so Claude Code edits the files in the current opencode project, not somewhere else.

__ACP_CLI__ --cwd "$(pwd)" "<task>"

The final answer prints to stdout. Exit code 0 means success; non-zero means error.

Examples

Ask a question (read-only investigation):

__ACP_CLI__ --cwd "$(pwd)" \
  "Explain how auth tokens are refreshed in this codebase and where the logic lives."

Code task with live progress on stderr:

__ACP_CLI__ --cwd "$(pwd)" --stream \
  "Add input validation to the POST /users handler and update its tests."

Machine-readable output (parse the JSON):

__ACP_CLI__ --cwd "$(pwd)" --json \
  "Audit this repo for hardcoded secrets and list each finding." 
# stdout: { "text": "...", "stopReason": "...", "toolCalls": [...] }

For a long task text, omit the positional prompt and pipe it on STDIN instead.

Multi-turn conversations (persistent sessions)

By default each call is an ephemeral one-shot — Claude Code does NOT remember earlier calls. To hold a multi-turn conversation or keep context across separate invocations, pass a STABLE --session <name> (one name per ongoing task/thread, e.g. --session refactor-auth). The first call with a name creates and saves the session; later calls with the same name RESUME it, so Claude Code remembers the earlier context. The --cwd is remembered per session.

Start a session:

__ACP_CLI__ --cwd "$(pwd)" --session refactor-auth \
  "Walk the auth module and outline a plan to split the token logic out of session.js."

Follow-up call that relies on remembered context:

__ACP_CLI__ --cwd "$(pwd)" --session refactor-auth \
  "Good. Now implement step 1 of that plan."

List stored sessions (name, id, cwd, last used, turn count):

__ACP_CLI__ --list-sessions

Every prompting run prints which session was used to STDERR:

[acp] session=refactor-auth id=<sessionId> (resumed)

With --json, the same info is in the output object as sessionName, sessionId, and resumed (alongside text, stopReason, toolCalls). Read either to report which session you used.

• Reuse the same --session <name> to continue a thread; pass --new to force-start a fresh session under that name (replacing the old thread) when you want to drop prior context.
• No --session at all is fine for one-off questions — the result just isn't saved.

Options

• --cwd <path> — session working dir (default: current dir). Pass the project root. Remembered per named session.
• --session <name> — use a persistent named session. First use creates+saves it; later uses with the same name resume it. Omit for an ephemeral one-shot (not saved).
• --new — force-start a fresh session for the given --session <name>, replacing the old thread.
• --list-sessions — list stored sessions (name, id, cwd, last used, turn count); with --json, raw JSON. Exits without prompting.
• --delete-session <name> — forget a stored session.
• --yolo — auto-approve every tool permission with allow_always. Without it, each tool call is approved once (allow_once).
• --stream — echo Claude's thoughts and tool calls to stderr for progress visibility.
• --json — emit { "text", "stopReason", "toolCalls", "sessionId", "sessionName", "resumed" } instead of plain text.
• --timeout <seconds> — default 600.
• --agent-cmd <cmd> — agent binary, default claude-agent-acp.

Sessions are persisted at ~/.config/opencode-claude-acp/sessions.json (override with OPENCODE_CLAUDE_ACP_STORE) and resumed via ACP session/load.

When to use

• Offload a self-contained coding, refactor, or analysis task to Claude Code.
• Get a second opinion from Claude on a design or a tricky bug.

When not to use

• Trivial edits or lookups you can do yourself — don't round-trip them.
• Be aware: this consumes Claude / Anthropic credits and can take a while (tasks may run up to the timeout).

File edits and --yolo

By default each tool permission is approved once per call. Pass --yolo to auto-approve everything (allow_always) so Claude Code can read and modify files unattended. Only suggest --yolo when the user explicitly wants Claude Code to make changes on its own; otherwise leave it off so edits surface for approval.

Prerequisites

• @agentclientprotocol/claude-agent-acp available (the claude-agent-acp binary on PATH, or installed as a dependency of this tool). Install globally with npm i -g @agentclientprotocol/claude-agent-acp if missing.
• Claude authenticated locally. If you hit auth errors, the user should run claude once to log in.