Claude codex bridge

Claude Codex Bridge

Use this workflow for bidirectional Claude-Codex review loops.

• Claude execution -> Codex audit -> Claude repair feedback.
• Codex execution -> Claude audit -> Codex repair feedback.

Default to automatic return after audit unless the user explicitly asks for audit-only mode.

Bridge delivery defaults to cli mode through the helper scripts. In cli mode, the target CLI performs the audit and prints the result back into this same conversation; do not use desktop App automation.

CLI runs are tracked under .agent-bridge/runs/<runId>/. Use .agent-bridge/runs/latest.json to inspect the latest run. A completed run has status: success and cli_done.json; a failed run has status: error and bridge_error.json. If a run still says status: started, it is in progress or was interrupted; do not treat that as a successful audit.

CLI preconditions: Codex CLI must be runnable for Codex audits. The bridge tries AGENT_BRIDGE_CODEX_BIN, then codex from PATH, then /Applications/Codex.app/Contents/Resources/codex. claude --version plus claude auth status must work for Claude audits. If the Claude auth preflight fails, tell the user to run Claude Code login/status rather than treating it as an App handoff failure.

If a CLI is missing or auth fails, run:

node "$HOME/.agent-bridge/bridge.mjs" doctor

Use the doctor's suggested fix. Common fixes are:

npm install -g @openai/codex
npm install -g @anthropic-ai/claude-code
claude auth login
claude auth status

If Codex Desktop is installed, the bridge can use /Applications/Codex.app/Contents/Resources/codex without an npm Codex install.

Use --delivery current, --delivery ax, or --delivery new only when the user explicitly asks to operate the desktop App UI. In ax or current mode, treat bridge success as macOS paste/submit automation completion. If the bridge reports a focus error or non-zero exit, do not tell the user the feedback was delivered; report the error and ask them to bring the target app conversation to the front before retrying.

Audit Claude From Codex

When the user asks Codex to audit Claude, find the handoff content in this order:

1. Use any file path the user gives.
2. Otherwise use the newest .agent-bridge/inbox/*audit-claude-to-codex.md file in the current workspace.
3. If no file exists, ask the user to run the Claude handoff bridge first.

Read the entire handoff file before judging the result.

Output:

Findings
- [P0/P1/P2/P3] Concrete issue with file/command evidence when available.

Missing Verification
- Tests, commands, screenshots, or checks Claude should still run.

Suggested Reply To Claude
<a concise message the user can send back to Claude>

Default CLI path: read the newest handoff, produce the audit, and leave the result in .agent-bridge/runs/<runId>/audit.md plus .agent-bridge/outbox/latest-codex-audit.md. The command output is already the return path to this conversation.

Only when using App delivery modes, write Suggested Reply To Claude to .agent-bridge/outbox/latest-reply-to-claude.md, then run:

node "$HOME/plugins/claude-codex-bridge/scripts/send-file-to-claude.mjs" .agent-bridge/outbox/latest-reply-to-claude.md

If the command exits successfully in an App delivery mode, say the bridge sent paste/submit automation to Claude. If it fails, include the error and do not claim Claude received the feedback.

Send Codex To Claude For Audit

When the user asks Claude to audit Codex/GPT-5.5 execution, create a handoff with:

• Original user goal.
• What Codex changed, including file paths.
• Commands run and verification results.
• Anything uncertain, skipped, or risky.
• A direct request for Claude to audit correctness, completeness, safety, and missing verification, then return Suggested Reply To Codex.

Send the handoff through the default CLI path:

node "$HOME/plugins/claude-codex-bridge/scripts/send-to-claude-for-audit.mjs"

Pass the full handoff on stdin. Claude CLI is expected to audit and return the feedback as command output in this same Codex conversation.

Shared Audit Rules

For either direction, prioritize behavioral bugs, incomplete work, risky commands, missing validation, and places where the executing agent claimed success without evidence. If no issues are found, say so clearly and still mention residual risk.

Use this audit shape:

Findings
- [P0/P1/P2/P3] Concrete issue with file/command evidence when available.

Missing Verification
- Tests, commands, screenshots, or checks Claude should still run.

Suggested Reply To <Agent>
<a concise message the user can send back to the executing agent>

If the user says "only audit", "do not forward", or equivalent, stop after showing the audit.