codex-cc

## Quick Start ```bash python scripts/claude_bridge.py --cd "/path/to/project" --PROMPT "Analyze auth flow" ``` **Output:** JSON with `success`, `SESSION_ID`, `agent_messages`, `stream_file` (the raw Claude JSONL stream), and optional `all_messages`, `stderr`, or `error`. ## Runtime Contract - Default execution uses `claude -p` and does **not** force `--safe-mode`, `--bare`, `--disable-slash-commands`, or `--strict-mcp-config`. - Claude Code therefore keeps its normal loading path for trusted-workspace customizations such as `CLAUDE.md`, skills, plugins, MCP servers, custom commands, and rules. - `claude -p` skips the interactive trust dialog and silently ignores invalid settings files, so use this only in workspaces you already trust and whose `.claude` settings already validate. - The bridge requests `stream-json`, persists each record immediately, and only treats Claude's final `result` record as a completed answer. Intermediate `assistant` records never mask a failed or incomplete turn. ## Parameters ```text usage: claude_bridge.py [-h] --PROMPT PROMPT --cd CD [--SESSION_ID SESSION_ID] [--model MODEL] [--permission-mode {,acceptEdits,auto,bypassPermissions,manual,dontAsk,plan}] [--dangerously-skip-permissions] [--timeout TIMEOUT] [--stream-file STREAM_FILE] [--return-all-messages] {mcp,plugin} ... options: --PROMPT PROMPT Instruct...