name	claude-channel-cli
description	Use when Codex needs to ask a user's live Claude Code session for a response through the local Claude Channel CLI or check whether the channel is reachable without using API-mode Claude access.

Claude Channel CLI

Use the bundled Claude Channel CLI MCP tools to communicate with the user's live Claude Code session. This preserves Claude Code's native harness and lets the user monitor the same session in their Claude Code window.

Workflow

List available Claude targets with list_claude_targets. If there is one target, use it. If there are several, show the numbered candidates and ask which visible Claude Code window to use.
Check the selected target with status_claude_channel, passing target when one was selected.
Use ask_claude, passing target when needed.
For review-sized, complex, or high-effort Claude Code asks, expect replies to take several minutes and sometimes close to the 30-minute default timeout. Do not treat a quiet Claude Code session as failed before the configured timeout. Ask the user before using a longer timeout.
If a tool returns multiple_claude_targets, ask the user to choose from candidates, then retry with that candidate's endpoint_id.
If the MCP tools are unavailable in a development checkout, fall back to the claude-channel CLI.

Rules

Do not use this skill for normal web/API Claude access. It is only for a local live Claude Code session.
Use ask_claude for every Claude Code request. The channel sender receives a response only through the reply-required completion path.
Keep Claude-facing prompts separate from Codex handling instructions. Send Claude only what Claude should answer.
If the user says "verbatim", "exactly", "send this block", or gives a fenced/quoted prompt, send only that exact block as message. Preserve wording, ordering, and whitespace.
For long or multiline exact prompts, pass the block directly as MCP message. In CLI fallback, stream only that block to ask-file -; never include Codex-only handling instructions in file/stdin payloads.
When content comes from another tool or plugin, summarize it only if the user asked for a summary. If the user asks to send it exactly, pass the selected text exactly.
Keep messages explicit about their source, e.g. start with From Codex:.
After ask_claude returns, use the structured answer field when answer_truncated is false. If answer_truncated is true, use answer_preview only for orientation and read the full response from answer_file before acting on it. Do not dump raw JSON unless debugging.
CLI fallback ask and ask-file print Claude's answer text by default. Use --output json only when the full response envelope is needed.
For very large CLI fallback reviews, redirect answer text to a visible file instead of relying on terminal output capture.
When multiple live targets exist, ask the user to choose from candidates; never guess from branch names, task names, terminal titles, or transcript names.
Prefer endpoint ids for retries and scripts. Numeric list indexes are acceptable only for immediate human CLI fallback.
If status_claude_channel reports the channel is not reachable, tell the user the channel is not running and ask them to start Claude Code in the receiver project with claude-channel start. claude-channel setup is only a readiness and stale-registration check.

CLI Fallback

Use this only when the Codex MCP tools are not installed or not available in the current thread:

claude-channel list
claude-channel status
claude-channel setup
claude-channel start
claude-channel ask --to ep_ABC234 "From Codex: review this plan and return your answer with complete_channel_request."
printf '%s\n' "$prompt" | claude-channel ask-file --to ep_ABC234 -
printf '%s\n' "$prompt" | claude-channel ask-file --output json --to ep_ABC234 - > claude-review.json

Claude Channel CLI

Workflow

List available Claude targets with list_claude_targets. If there is one target, use it. If there are several, show the numbered candidates and ask which visible Claude Code window to use.

Check the selected target with status_claude_channel, passing target when one was selected.

Use ask_claude, passing target when needed.

For review-sized, complex, or high-effort Claude Code asks, expect replies to take several minutes and sometimes close to the 30-minute default timeout. Do not treat a quiet Claude Code session as failed before the configured timeout. Ask the user before using a longer timeout.

If a tool returns multiple_claude_targets, ask the user to choose from candidates, then retry with that candidate's endpoint_id.

If the MCP tools are unavailable in a development checkout, fall back to the claude-channel CLI.

Rules

Do not use this skill for normal web/API Claude access. It is only for a local live Claude Code session.

Use ask_claude for every Claude Code request. The channel sender receives a response only through the reply-required completion path.

Keep Claude-facing prompts separate from Codex handling instructions. Send Claude only what Claude should answer.

If the user says "verbatim", "exactly", "send this block", or gives a fenced/quoted prompt, send only that exact block as message. Preserve wording, ordering, and whitespace.

For long or multiline exact prompts, pass the block directly as MCP message. In CLI fallback, stream only that block to ask-file -; never include Codex-only handling instructions in file/stdin payloads.

When content comes from another tool or plugin, summarize it only if the user asked for a summary. If the user asks to send it exactly, pass the selected text exactly.

Keep messages explicit about their source, e.g. start with From Codex:.

After ask_claude returns, use the structured answer field when answer_truncated is false. If answer_truncated is true, use answer_preview only for orientation and read the full response from answer_file before acting on it. Do not dump raw JSON unless debugging.

CLI fallback ask and ask-file print Claude's answer text by default. Use --output json only when the full response envelope is needed.

For very large CLI fallback reviews, redirect answer text to a visible file instead of relying on terminal output capture.

When multiple live targets exist, ask the user to choose from candidates; never guess from branch names, task names, terminal titles, or transcript names.

Prefer endpoint ids for retries and scripts. Numeric list indexes are acceptable only for immediate human CLI fallback.

If status_claude_channel reports the channel is not reachable, tell the user the channel is not running and ask them to start Claude Code in the receiver project with claude-channel start. claude-channel setup is only a readiness and stale-registration check.

CLI Fallback

Use this only when the Codex MCP tools are not installed or not available in the current thread:

claude-channel list claude-channel status claude-channel setup claude-channel start claude-channel ask --to ep_ABC234 "From Codex: review this plan and return your answer with complete_channel_request." printf '%s\n' "$prompt" | claude-channel ask-file --to ep_ABC234 - printf '%s\n' "$prompt" | claude-channel ask-file --output json --to ep_ABC234 - > claude-review.json

claude-channel-cli