{
  "prompt": "Review this function for race conditions",
  "cwd": "/path/to/project",
  "sandbox": "read-only",
  "approval-policy": "untrusted",
  "model": "gpt-5.2-codex"
}

{
  "threadId": "thread_abc123",
  "prompt": "Now show me the fix"
}

1. Call mcp__codex__codex with initial prompt
2. Get response with threadId
3. Call mcp__codex__codex-reply with threadId + follow-up
4. Repeat step 3 for multi-turn conversations

# Direct prompt argument
codex exec "explain this algorithm" --full-auto

# Pipe from stdin using -
codex exec - --full-auto < /tmp/question.txt

# Pipe directly (prompt reads from stdin)
cat /tmp/question.txt | codex exec - --full-auto

cat > /tmp/question.txt << 'EOF'
Problem: [Clear problem statement]

Error/Symptom: [Specific error message or behavior]

Here is the full function:
```language
[paste COMPLETE code - don't truncate]

### Step 2: Run Codex

```bash
codex exec - -o /tmp/reply.txt --full-auto < /tmp/question.txt

cat /tmp/reply.txt

# Initial analysis
codex exec "review this code for race conditions" --full-auto

# Continue with follow-up
codex exec resume --last "now fix the race conditions you found"

# Or resume specific session by ID
codex exec resume SESSION_ID "implement the suggested fix"

# schema.json
{
  "type": "object",
  "properties": {
    "bugs_found": { "type": "array", "items": { "type": "string" } },
    "severity": { "type": "string", "enum": ["low", "medium", "high", "critical"] }
  },
  "required": ["bugs_found", "severity"]
}

# Run with schema
codex exec "analyze for bugs" --output-schema ./schema.json -o result.json --full-auto

codex exec -i error_screenshot.png "what's causing this error?" --full-auto
codex exec -i diagram.png,spec.png "does implementation match spec?" --full-auto

# One-time setup: store key in Keychain
security add-generic-password -a "$USER" -s "openai-api-key" -w "sk-..."

# In ~/.zshrc
export OPENAI_API_KEY=$(security find-generic-password -a "$USER" -s "openai-api-key" -w 2>/dev/null)

# Option 1: Inline env var
CODEX_API_KEY=sk-... codex exec "task" --full-auto

# Option 2: Pipe to login (persists credentials)
echo "$OPENAI_API_KEY" | codex login --with-api-key

# Option 3: Check login status
codex login status

# 1. Create detailed question
cat > /tmp/question.txt << 'EOF'
Problem: Progressive JPEG decoder fails at block 1477 with "Invalid Huffman code"

Here is the AC refinement function:
```c
void decode_ac_refinement(bitstream *bs, int16_t *block, int start, int end) {
    // [full 50-line function here]
}

## Best Practices

1. **Provide complete code**: Don't truncate. Codex needs full context.

2. **Be specific**: "Why does Huffman decoding fail at block 1477 in AC refinement?" beats "Why does this fail?"

3. **Include specs**: Reference relevant standards (JPEG ITU-T.81, PNG RFC 2083, etc.).

4. **Verify suggestions**: Codex is helpful but not infallible—always verify against authoritative sources.

5. **Iterate with resume**: Use `codex exec resume --last` for multi-turn debugging sessions.

## Common Issues

| Problem | Solution |
|---------|----------|
| "stdin is not a terminal" | You used `codex` instead of `codex exec` |
| No output file created | Ensure `-o` path is writable |
| Timeout/blocking | Use `--full-auto` to skip approval prompts |
| "Not in a git repository" | Use `--skip-git-repo-check` or initialize git |
| Need write permissions | Use `--sandbox workspace-write` or `danger-full-access` |
| Authentication error | Set `CODEX_API_KEY` env var or pipe key to `codex login --with-api-key` |

## Model Selection

```bash
# Use specific model
codex exec -m gpt-5-codex "analyze this" --full-auto
codex exec -m gpt-5 "analyze this" --full-auto

# Use local model via Ollama
codex exec --oss "analyze this" --full-auto