Codex Exec - Non-Interactive Codex CLI

codex exec (alias: codex e) runs Codex in non-interactive mode, suitable for automation, scripting, and CI/CD pipelines.

Preferred Defaults

Unless otherwise specified, use these settings:

codex exec --full-auto -m gpt-5.3-codex -c model_reasoning_effort=high "your prompt"

• --full-auto - Low-friction sandboxed automatic execution
• -m gpt-5.3-codex - Preferred model
• -c model_reasoning_effort=high - High reasoning level

Basic Usage

# With inline prompt
codex exec "your prompt here"

# Read prompt from stdin
echo "your prompt" | codex exec -

# Read prompt from file
codex exec - < prompt.txt

Key Options

Model Selection

-m, --model <MODEL>        # Specify model (e.g., "gpt-5.3-codex", "o3")

Sandbox Modes

-s, --sandbox <MODE>

Execution Modes

--full-auto                # Combines -a on-request with --sandbox workspace-write
--dangerously-bypass-approvals-and-sandbox  # Skip all prompts (DANGEROUS - only for externally sandboxed environments)

Working Directory

-C, --cd <DIR>            # Set working root directory
--add-dir <DIR>           # Additional writable directories
--skip-git-repo-check     # Allow running outside git repos

Output Control

--json                    # Output events as JSONL to stdout
-o, --output-last-message <FILE>  # Write agent's final message to file
--output-schema <FILE>    # JSON Schema file for structured response
--color <always|never|auto>  # Color output control

Configuration

-c, --config <key=value>  # Override config.toml values
-p, --profile <PROFILE>   # Use named profile from config.toml
--enable <FEATURE>        # Enable feature flag
--disable <FEATURE>       # Disable feature flag

Config override examples:

• -c model="o3"
• -c 'sandbox_permissions=["disk-full-read-access"]'
• -c shell_environment_policy.inherit=all

Images

-i, --image <FILE>...     # Attach image(s) to prompt

Web Search

--search                  # Enable live web search tool

Subcommands

Resume a Session

codex exec resume [SESSION_ID] [PROMPT]
codex exec resume --last [PROMPT]

Code Review

codex exec review [OPTIONS] [PROMPT]

JSON Output Format

When using --json, events are emitted as JSONL with these types:

The item.completed event includes an item object with:

• id - Item identifier
• type - One of: reasoning, agent_message, tool-related types
• text - Content of the item

Common Patterns

Default Invocation

codex exec --full-auto -m gpt-5.3-codex -c model_reasoning_effort=high "your task"

With Output Capture

codex exec --full-auto -m gpt-5.3-codex -c model_reasoning_effort=high -o result.txt "your task"

Structured Output with Schema

codex exec --output-schema schema.json "generate data matching schema"

Code Review Against Main Branch

codex exec review --base main

Review Uncommitted Changes

codex exec review --uncommitted

With Additional Config Overrides

codex exec --full-auto -m gpt-5.3-codex -c model_reasoning_effort=high -c 'sandbox_permissions=["disk-full-read-access"]' "your task"

Notes

• Prompts can be passed as arguments or via stdin (use - to read from stdin)
• Running outside a git repository requires --skip-git-repo-check
• The --full-auto flag is a convenience alias for low-friction sandboxed execution
• Use --dangerously-bypass-approvals-and-sandbox only in externally sandboxed environments