Claude from codex

Claude From Codex

When To Use

Use this skill when the user wants Codex to involve local Claude Code as a second model or parallel worker. Good fits include:

• Independent investigation or review while Codex keeps working.
• A second opinion on a plan, diff, failure, or design choice.
• A bounded implementation slice where the user explicitly allows Claude to write.
• Long-running analysis that should continue in the background.

Do not use this skill to configure Claude auth, provider routing, cc-switch, or ~/.claude/settings.json. Those are external user setup. Run setup diagnostics and tell the user what must be fixed outside Codex.

Choose The Path

Use the simplest path that fits the task.

• Basic path, no JavaScript required: call the user's claude CLI directly for simple foreground review, second opinion, or investigation.
• Advanced path, Node required: use scripts/claude-companion.mjs for setup diagnostics, Codex-vs-terminal environment debugging, background jobs, cancellation, state pruning, or stricter env filtering.

Rules

• Pass the task prompt on stdin, never as a command-line argument.
• Include explicit context in the prompt: goal, relevant paths, constraints, acceptance criteria, and whether this is read-only or write-capable.
• Default to read-biased delegation. Use write-capable Claude work only when the current user turn explicitly asks for it.
• Treat Claude's output as input to reconcile. Inspect any factual claims or reported changes before integrating them.
• For write-capable runs, inspect the full workspace diff, including git metadata and unexpected files, before continuing.
• If direct claude behaves differently inside Codex than in a normal terminal, switch to the advanced setup diagnostics instead of editing Claude settings.

Basic Commands

These commands require only the user's Claude Code CLI. They do not require Node or the companion script.

Check auth visibility from the current Codex process:

claude --setting-sources user,project,local auth status --json

Foreground read-biased run, with prompt on stdin:

printf '%s\n' "$PROMPT" | claude -p --output-format json --setting-sources user,project,local --permission-mode plan --allowedTools Read,Grep,Glob --disable-slash-commands --no-session-persistence

Use write-capable direct Claude only after explicit user request and normal Claude permission handling:

printf '%s\n' "$PROMPT" | claude -p --output-format json --setting-sources user,project,local --permission-mode default --disable-slash-commands --no-session-persistence

If claude is not on PATH inside Codex, try the common user-local binary directly:

printf '%s\n' "$PROMPT" | "$HOME/.local/bin/claude" -p --output-format json --setting-sources user,project,local --permission-mode plan --allowedTools Read,Grep,Glob --disable-slash-commands --no-session-persistence

Advanced Commands

Use these only when Node is available and the task benefits from diagnostics, env filtering, or background job management. From this skill folder, the companion is:

node scripts/claude-companion.mjs setup

When invoking from another working directory, use the absolute installed path:

node "$HOME/.agents/skills/claude-from-codex/scripts/claude-companion.mjs" setup --json

Foreground read-biased run through the companion:

printf '%s\n' "$PROMPT" | node "$HOME/.agents/skills/claude-from-codex/scripts/claude-companion.mjs" run --json

Foreground write-capable run, only after explicit user request:

printf '%s\n' "$PROMPT" | node "$HOME/.agents/skills/claude-from-codex/scripts/claude-companion.mjs" run --write --json

Background run:

printf '%s\n' "$PROMPT" | node "$HOME/.agents/skills/claude-from-codex/scripts/claude-companion.mjs" run --background --json

Background management:

node "$HOME/.agents/skills/claude-from-codex/scripts/claude-companion.mjs" status --json
node "$HOME/.agents/skills/claude-from-codex/scripts/claude-companion.mjs" status <job-id> --json
node "$HOME/.agents/skills/claude-from-codex/scripts/claude-companion.mjs" result <job-id> --json
node "$HOME/.agents/skills/claude-from-codex/scripts/claude-companion.mjs" cancel <job-id> --json
node "$HOME/.agents/skills/claude-from-codex/scripts/claude-companion.mjs" prune --older-than 7d --json

Prompt Shape

When delegating, write a self-contained prompt. Claude does not receive the full Codex transcript automatically.

Goal: ...
Workspace: ...
Relevant files:
- ...
Mode: read-biased investigation | write-capable implementation
Constraints:
- ...
Acceptance criteria:
- ...
Return:
- Findings, decisions, changed files if any, and remaining risks.

Result Handling

For read-biased runs, summarize Claude's findings and verify any factual claims against the workspace when they matter.

For write-capable runs, require the companion result to show mode, exit status, changed files, and the diff-inspection reminder. Then inspect the local diff before continuing.

Boundaries

This v1 skill is a plain Codex skill folder. It does not require a Codex plugin or marketplace login. Future plugin packaging, richer session sync, Windows background cancellation, and enforced sandboxing are deferred.