"""
OpenAI Codex CLI backend.
Implements the AgentCLI protocol for OpenAI's Codex CLI.
This is a basic implementation - full integration will be done in later tasks.
Codex CLI reference: https://github.com/openai/codex
"""
from typing import Literal
from .base import AgentCLI
from ..utils.env_vars import get_env_with_fallback
# Built-in default command.
_DEFAULT_COMMAND = "codex"
# Environment variables for command override (takes highest precedence).
_ENV_VAR = "MANIPLE_CODEX_COMMAND"
_ENV_VAR_FALLBACK = "CLAUDE_TEAM_CODEX_COMMAND"
def get_codex_command() -> str:
"""
Get the Codex CLI command with precedence: env var > config > default.
Resolution order:
1. MANIPLE_CODEX_COMMAND environment variable (for override)
2. Config file commands.codex setting
3. Built-in default "codex"
Returns:
The command to use for Codex CLI
"""
# Environment variable takes highest precedence (for override).
env_val = get_env_with_fallback(_ENV_VAR, _ENV_VAR_FALLBACK)
if env_val:
return env_val
# Try config file next.
# Import here to avoid circular imports and lazy-load config.
try:
from ..config import ConfigError, load_config
config = load_config()
except ConfigError:
return _DEFAULT_COMMAND
if config.commands.codex:
return config.commands.codex
# Fall back to built-in default.
return _DEFAULT_COMMAND
class CodexCLI(AgentCLI):
"""
OpenAI Codex CLI implementation.
Note: This is a basic structure. Full Codex integration (ready detection,
idle detection, etc.) will be implemented in later tasks (cic-f7w.3+).
Codex CLI characteristics:
- Uses `codex` command
- Has --dangerously-bypass-approvals-and-sandbox flag for non-interactive mode
- No known Stop hook equivalent (may need JSONL streaming or timeouts)
"""
@property
def engine_id(self) -> str:
"""Return 'codex' as the engine identifier."""
return "codex"
def command(self) -> str:
"""
Return the Codex CLI command.
Resolution order:
1. MANIPLE_CODEX_COMMAND environment variable (for override)
2. Config file commands.codex setting
3. Built-in default "codex"
"""
return get_codex_command()
def build_args(
self,
*,
dangerously_skip_permissions: bool = False,
settings_file: str | None = None,
) -> list[str]:
"""
Build Codex CLI arguments for interactive mode.
Args:
dangerously_skip_permissions: Maps to --dangerously-bypass-approvals-and-sandbox for Codex
settings_file: Ignored - Codex doesn't support settings injection
Returns:
List of CLI arguments for interactive mode
"""
args: list[str] = []
# Codex uses --dangerously-bypass-approvals-and-sandbox for autonomous operation.
if dangerously_skip_permissions:
args.append("--dangerously-bypass-approvals-and-sandbox")
# Note: settings_file is ignored - Codex doesn't support this
# Idle detection uses session file polling instead
return args
def ready_patterns(self) -> list[str]:
"""
Return patterns indicating Codex CLI is ready for input.
Codex in interactive mode shows status bar when ready.
Updated for Codex CLI v0.80.0 behavior.
"""
return [
"context left", # Status bar shows "100% context left"
"for shortcuts", # Status bar shows "? for shortcuts"
"What can I help you with?", # Legacy prompt (older versions)
"codex>", # Alternative prompt pattern
"»", # Codex uses this prompt symbol
"Waiting for messages", # Happy codex wrapper
"Codex Agent Running", # Happy codex status bar
]
def idle_detection_method(self) -> Literal["stop_hook", "jsonl_stream", "none"]:
"""
Codex idle detection method.
Codex writes session files to ~/.codex/sessions/YYYY/MM/DD/.
The idle_detection module polls these files for agent_message
events which indicate the agent has finished responding.
"""
return "jsonl_stream"
def supports_settings_file(self) -> bool:
"""
Codex doesn't support --settings for hook injection.
Alternative completion detection methods will be needed.
"""
return False
# Singleton instance for convenience
codex_cli = CodexCLI()
