codex

Codex: High-Reasoning AI Assistant for Claude Code

---

DEFAULT MODEL: Task-Based Model Selection with Read-Only Default

Codex uses task-based model selection. Sandbox is read-only by default - only use workspace-write when user explicitly requests file editing.

Task Type	Model	Sandbox (default)	Sandbox (explicit edit)
Code-related tasks	gpt-5.2-codex	read-only	workspace-write
General tasks	gpt-5.2	read-only	workspace-write

• Code-related tasks: Use gpt-5.2-codex - optimized for agentic coding (56.4% SWE-Bench Pro)
• General tasks: Use gpt-5.2 - high-reasoning general model
• Sandbox default: Always read-only unless user explicitly requests editing
• Explicit editing: Only when user says "edit", "modify", "write changes", etc., use workspace-write
• Always use -c model_reasoning_effort=xhigh for maximum capability

# Code task (read-only default)
codex exec -m gpt-5.2-codex -s read-only \
  -c model_reasoning_effort=xhigh \
  "analyze this function implementation"

# General task (read-only default)
codex exec -m gpt-5.2 -s read-only \
  -c model_reasoning_effort=xhigh \
  "explain this architecture"

# Code task with explicit edit request
codex exec -m gpt-5.2-codex -s workspace-write \
  -c model_reasoning_effort=xhigh \
  "edit this file to add the feature"

# General task with explicit edit request
codex exec -m gpt-5.2 -s workspace-write \
  -c model_reasoning_effort=xhigh \
  "modify the documentation file"

Model Fallback Chain

If the primary model is unavailable, fallback gracefully:

1. Code tasks: gpt-5.2-codex → gpt-5.2 → gpt-5.1-codex-max
2. General tasks: gpt-5.2 → gpt-5.1 → gpt-5.1-codex-max
3. Reasoning effort: xhigh → high → medium

---

CRITICAL: Always Use codex exec

MUST USE: codex exec for ALL Codex CLI invocations in Claude Code.

NEVER USE: codex (interactive mode) - will fail with "stdout is not a terminal" ALWAYS USE: codex exec (non-interactive mode)

Examples:

• codex exec -m gpt-5.2 "prompt" (CORRECT)
• codex -m gpt-5.2 "prompt" (WRONG - will fail)
• codex exec resume --last (CORRECT)
• codex resume --last (WRONG - will fail)

Why? Claude Code's bash environment is non-terminal/non-interactive. Only codex exec works in this environment.

---

IMPORTANT: Interactive vs Exec Mode Flags

Some Codex CLI flags are ONLY available in interactive mode, NOT in codex exec.

Flag	Interactive codex	codex exec	Alternative for exec
--search	✅ Available	❌ NOT available	--enable web_search_request
-a/--ask-for-approval	✅ Available	❌ NOT available	--full-auto or -c approval_policy=...
--add-dir	✅ Available	✅ Available	N/A
--full-auto	✅ Available	✅ Available	N/A

For web search in exec mode:

# CORRECT - works in codex exec
codex exec --enable web_search_request "research topic"

# WRONG - --search only works in interactive mode
codex --search "research topic"

For approval control in exec mode:

# CORRECT - works in codex exec
codex exec --full-auto "task"
codex exec -c approval_policy=on-request "task"

# WRONG - -a only works in interactive mode
codex -a on-request "task"

---

Trigger Examples

This skill activates when users say phrases like:

• "Use codex to analyze this architecture"
• "Ask codex about this design decision"
• "Run codex on this problem"
• "Call codex for help with this implementation"
• "I need GPT-5 reasoning for this task"
• "Get OpenAI's high-reasoning model on this"
• "Continue with codex" or "Resume the codex session"
• "Codex, help me with..." or simply "Codex"

When to Use This Skill

This skill should be invoked when:

• User explicitly mentions "Codex" or requests Codex assistance
• User needs help with complex coding tasks, algorithms, or architecture
• User requests "high reasoning" or "advanced implementation" help
• User needs complex problem-solving or architectural design
• User wants to continue a previous Codex conversation

How It Works

Detecting New Codex Requests

When a user makes a request, first determine the task type (code vs general), then determine sandbox based on explicit edit request:

Step 1: Determine Task Type (Model Selection)

• Code-related tasks: Use gpt-5.2-codex - for implementation, refactoring, code analysis, debugging, etc.
• General tasks: Use gpt-5.2 - for architecture design, explanations, reviews, documentation, etc.

Step 2: Determine Sandbox (Edit Permission)

• Default: read-only - safe for all tasks unless user explicitly requests editing
• Explicit edit request: workspace-write - ONLY when user explicitly says to edit/modify/write files

Code-related task examples:

• Read-only: "Analyze this function", "Review this implementation", "Debug this code"
• With editing: "Edit this file to fix the bug", "Modify the function", "Refactor and save"

General task examples:

• Read-only: "Design a queue data structure", "Explain this algorithm", "Review the architecture"
• With editing: "Update the documentation file", "Modify the README"

⚠️ Important: The key distinction for sandbox is whether the user explicitly asks for file modifications. Use workspace-write ONLY when user says "edit", "modify", "write changes", "save", etc.

Bash CLI Command Structure

IMPORTANT: Always use codex exec for non-interactive execution. Claude Code's bash environment is non-terminal, so the interactive codex command will fail with "stdout is not a terminal" error.

Code Task (Read-Only Default)

codex exec -m gpt-5.2-codex -s read-only \
  -c model_reasoning_effort=xhigh \
  --enable web_search_request \
  "<code-related prompt>"

General Task (Read-Only Default)

codex exec -m gpt-5.2 -s read-only \
  -c model_reasoning_effort=xhigh \
  --enable web_search_request \
  "<general prompt>"

Code Task with Explicit Edit Request

codex exec -m gpt-5.2-codex -s workspace-write \
  -c model_reasoning_effort=xhigh \
  --enable web_search_request \
  "<edit code prompt>"

General Task with Explicit Edit Request

codex exec -m gpt-5.2 -s workspace-write \
  -c model_reasoning_effort=xhigh \
  --enable web_search_request \
  "<edit general files prompt>"

Why codex exec?

• Non-interactive mode required for automation and Claude Code integration
• Produces clean output suitable for parsing
• Works in non-TTY environments (like Claude Code's bash)

Model Selection Logic

Step 1: Choose Model Based on Task Type

Use gpt-5.2-codex for code-related tasks:

• Implementation, refactoring, code analysis
• Debugging, fixing bugs, optimization
• Any task involving code understanding or modification

Use gpt-5.2 for general tasks:

• Architecture and system design
• Explanations, documentation, reviews
• Planning, strategy, general reasoning

Step 2: Choose Sandbox Based on Edit Intent

Use read-only (DEFAULT):

• Analysis, review, explanation tasks
• ANY task where user does NOT explicitly request file editing

Use workspace-write (ONLY when explicitly requested):

• User explicitly says "edit this file", "modify the code", "write changes"
• User explicitly asks to "make edits" or "save the changes"
• User explicitly requests "refactor and save" or "implement and write"

Fallback Models: gpt-5.1-codex-max and gpt-5.1 are available if primary models are unavailable. See fallback chain in DEFAULT MODEL section.

Default Configuration

All Codex invocations use these defaults unless user specifies otherwise:

Parameter	Default Value	CLI Flag	Notes
Model (code tasks)	gpt-5.2-codex	-m gpt-5.2-codex	For code-related tasks
Model (general tasks)	gpt-5.2	-m gpt-5.2	For general tasks
Sandbox (default)	read-only	-s read-only	Safe default for ALL tasks
Sandbox (explicit edit)	workspace-write	-s workspace-write	Only when user explicitly requests editing
Reasoning Effort	xhigh	-c model_reasoning_effort=xhigh	Maximum reasoning capability
Verbosity	medium	-c model_verbosity=medium	Balanced output detail
Web Search	enabled	--enable web_search_request	Access to up-to-date information

CLI Flags Reference

Codex CLI Version: 0.72.0+ (requires 0.72.0+ for gpt-5.2-codex and xhigh)

Flag	Values	Description
-m, --model	gpt-5.2-codex, gpt-5.2, gpt-5.1-codex-max, gpt-5.1	Model selection
-s, --sandbox	read-only, workspace-write, danger-full-access	Sandbox mode
-c, --config	key=value	Config overrides (e.g., model_reasoning_effort=high)
-C, --cd	directory path	Working directory
-p, --profile	profile name	Use config profile
--enable	feature name	Enable a feature (e.g., web_search_request)
--disable	feature name	Disable a feature
-i, --image	file path(s)	Attach image(s) to initial prompt
--add-dir	directory path	Additional writable directory (repeatable)
--full-auto	flag	Convenience for workspace-write sandbox with on-request approval
--oss	flag	Use local open source model provider
--local-provider	lmstudio, ollama	Specify local provider (with --oss)
--skip-git-repo-check	flag	Allow running outside Git repository
--output-schema	file path	JSON Schema file for response shape
--color	always, never, auto	Color settings for output
--json	flag	Print events as JSONL
-o, --output-last-message	file path	Save last message to file
`

---

Content truncated.