Codex Claude Bridge

Codex Claude Bridge lets Codex call Claude as a secondary verifier, reviewer, or risk critic while keeping Codex as the primary operator.

It avoids Claude's interactive terminal UI and always calls Claude through non-interactive claude -p --output-format json.

What It Installs

• codex-claude-bridge command
• A Codex skill named codex-claude-bridge
• A small AGENTS.md rule block that tells Codex when to call Claude

Requirements

• Codex CLI or Codex App
• Claude Code CLI authenticated on the same machine
• Python 3
• Bash

Optional:

• Report-compatible CLI for report logging, configured through environment variables

One-Line Install For Codex Users

If you already have access to this repository, paste this into Codex App or Codex CLI:

Clone this repository into a temporary directory, run ./install.sh from the repository root, then run codex-claude-bridge doctor --json and summarize the result.

Or run the same install directly from a shell, replacing the placeholder with the Git clone URL shown by GitHub:

repo_url="<repository-git-url>" && tmpdir="$(mktemp -d)" && git clone "$repo_url" "$tmpdir/codex-claude-bridge" && cd "$tmpdir/codex-claude-bridge" && ./install.sh && "${CODEX_HOME:-$HOME/.codex}/bin/codex-claude-bridge" doctor --json

Install

Clone this repository, then run:

./install.sh

Run a health check:

codex-claude-bridge doctor --json

If your shell cannot find the command, either call it directly from your Codex home:

$HOME/.codex/bin/codex-claude-bridge doctor --json

or add the installer-created local bin path to PATH:

export PATH="$HOME/.local/bin:$PATH"

Usage

Ask Claude for a bounded second opinion:

codex-claude-bridge ask "Check this plan for missing risks."

Ask for a code review:

codex-claude-bridge review --cwd /path/to/repo "Review the current diff."

Ask for a high-risk trading-system review:

codex-claude-bridge trade-review --cwd /path/to/repo "Review risk controls and failure modes."

Check status:

codex-claude-bridge status --json

Read the latest result:

codex-claude-bridge last --result-only

Model Routing

By default, --model auto --effort auto chooses a model and reasoning effort based on the task:

• Small consults may use Haiku.
• Code reviews may use Sonnet.
• Trading risk, high-risk architecture, or deep delegation may use Opus with max effort.

High-risk trade-review runs may take several minutes. If a caller passes a very short --timeout-sec, the bridge raises it to a safer local floor of 900 seconds by default. Override that floor with:

export CODEX_CLAUDE_BRIDGE_TRADE_REVIEW_MIN_TIMEOUT_SEC=1200

You can override this:

codex-claude-bridge ask --model sonnet --effort high "Question"

Logging

Local JSONL logs are written to:

$CODEX_HOME/log/claude-bridge.jsonl

or, when CODEX_HOME is unset:

$HOME/.codex/log/claude-bridge.jsonl

Report logging is opt-in. Set both:

export CODEX_CLAUDE_BRIDGE_REPORT_CMD=/path/to/report-command
export CODEX_CLAUDE_BRIDGE_REPORT_REPO=/path/to/report-repo

Safety Boundary

This project reduces accidental local data leakage by:

• Avoiding interactive Claude terminal mode
• Keeping Codex as final decision maker
• Logging only bounded prompt previews and Claude results
• Treating Claude output as advisory evidence
• Disallowing file edits in review mode
• Using timeouts for Claude and optional report logging

It does not make any machine, model, or repository impossible to reverse engineer or compromise. Do not place API keys, trading credentials, private data, or production secrets in prompts, logs, documentation, or repository files.

Uninstall

./uninstall.sh

Local logs are intentionally left in place.