Dev G — Your Codex Coding Partner

Delegate coding execution to Codex CLI. Dev G turns clear plans into working code.

Critical rules

• ONLY interact with Dev G through the bundled shell script. NEVER call codex CLI directly.
• Run the script ONCE per task. If it succeeds (exit code 0), read the output file and proceed. Do NOT re-run or retry.
• Do NOT read or inspect the script source code. Treat it as a black box.

How to call the script

The script path is:

~/.claude/skills/codex-dev-g/scripts/ask_dev_g.sh

Minimal invocation:

~/.claude/skills/codex-dev-g/scripts/ask_dev_g.sh \
  --workspace /absolute/workspace/path \
  --task "Your request in natural language"

With file context:

~/.claude/skills/codex-dev-g/scripts/ask_dev_g.sh \
  --workspace /absolute/workspace/path \
  --task "Refactor these components to use the new API" \
  --file src/components/UserList.tsx \
  --file src/components/UserDetail.tsx

Multi-turn conversation (continue a previous session):

~/.claude/skills/codex-dev-g/scripts/ask_dev_g.sh \
  --workspace /absolute/workspace/path \
  --session <session_id from previous run> \
  --task "Also add retry logic with exponential backoff"

The script prints on success:

session_id=<thread_id>
output_path=<path to markdown file>

Read the file at output_path to get Dev G's response. Save session_id if you plan follow-up calls.

Decision policy

Call Dev G when at least one of these is true:

• The implementation plan is clear and needs coding execution.
• The task involves batch refactoring, code generation, or repetitive changes.
• Multiple files need coordinated modifications following a defined pattern.
• You want a practitioner's perspective on whether a plan is feasible.
• The task is cost-sensitive and doesn't require deep architectural reasoning.
• Writing or updating tests based on existing code.
• Simple-to-moderate bug fixes where the root cause is identified.

Handle it yourself when:

• The task requires architecture design, technical strategy, or tradeoff analysis.
• Copywriting, documentation prose, or nuanced communication is needed.
• Deep contextual understanding of business requirements is critical.
• The problem is ambiguous and needs clarification before any code is written.

Workflow

Task delegation (most common)

1. Design the solution and break it into concrete steps.
2. Run the script with --task describing exactly what to implement.
3. Pass relevant files with --file so Dev G has context.
4. Read the output — Dev G executes changes and reports what it did.
5. Review the changes in your workspace.

Discussion mode

1. Run the script with a question-oriented --task.
2. Pass the relevant files for Dev G to analyze.
3. Read its feedback — it thinks from an implementer's perspective.
4. Combine its practical insights with your architectural judgment.

Multi-step projects

Use --session to continue the conversation with context:

1. First call: core implementation. Save the session_id from output.
2. Review output, then second call with --session <id>: tests and edge cases.
3. Third call with --session <id> if needed: cleanup and polish.

Each follow-up call carries the full conversation history.

File reference guidance

• Use --file with workspace-relative or absolute paths.
• Include 2-6 high-signal files rather than dumping everything.
• Dev G runs with full workspace access, so it can discover related files on its own.

Options

• --session <id> — Resume a previous session for multi-turn conversation.
• --model <name> — Override model (default: uses Codex config).
• --sandbox <mode> — Override sandbox policy (default: workspace-write via full-auto).
• --read-only — Run in read-only mode for pure discussion/analysis, no file changes.

Execution notes

• Dev G runs with --full-auto by default, meaning it can read and write files in the workspace autonomously.
• Keep --task text concrete and actionable. Vague requests get vague results.
• When Dev G's suggestions conflict with your architectural decisions, your judgment takes priority.