claude-code

name	claude-code
description	Delegate heavy coding tasks to Claude Code (Anthropic's autonomous coding CLI). Use for large multi-file refactors, building entire features or modules, complex debugging that needs extensive exploration, and substantial code reviews. For small edits, single-file changes, and quick fixes, use Vesta's own Read/Edit/Bash tools directly.

Claude Code - CLI: claude

The claude CLI is a separate autonomous coding agent with a coding-tuned system prompt and the right default tools.

When you invoke it, shell out in print mode (-p). It runs to completion in one call and only the final result comes back, so Vesta's own context stays clean.

When to use this skill

Reach for claude when the task is genuinely big:

• Building a whole feature or new module
• Multi-file refactors or "do this across the repo"
• Complex debugging that needs extensive exploration before a fix
• Substantial code reviews or security reviews of large diffs
• Anything that would otherwise burn many turns of Vesta's own loop reading and editing files

For everything smaller (single-file edits, quick fixes, typos, renames, reading code, small additions to existing functions), do it yourself with Read/Edit/Bash: the subprocess and autonomous-turn overhead exceeds Claude Code's value on modest work.

Basic call

claude -p "<task>" \
  --output-format json \
  --max-turns 20 \
  --permission-mode bypassPermissions \
  --allowedTools "Read,Edit,Write,Bash,Grep,Glob"

JSON response shape:

{
  "result": "<final text Claude Code returned>",
  "session_id": "75e2167f-...",
  "total_cost_usd": 0.078,
  "num_turns": 7,
  "subtype": "success"
}

Parse result and report it. Capture session_id for follow-ups.

• Use a generous bash timeout when invoking it: coding work can take minutes, so set 5+ minute timeouts for non-trivial tasks.
• Pass the user's exact intent, not your paraphrase: Claude Code's prompt is tuned for natural-language coding tasks; don't pre-digest.
• Don't nest claude calls inside the task string: if the task itself involves running claude, you've over-decomposed; just describe the goal.

Multi-turn (follow-ups)

Sessions live on Claude Code's side. No Vesta storage needed: the session_id is just a string you hold in conversation context and pass back.

# Turn 1: get session_id from the JSON
claude -p "Refactor auth module to use JWT" --output-format json --max-turns 20 ...

# Turn 2: continue the same session
claude -p "Now add tests" --resume <session_id> --output-format json --max-turns 10 ...

# Or continue the most-recent session in this cwd
claude -p "Anything you missed?" --continue --output-format json --max-turns 5 ...

# Fork: new id, keeps history
claude -p "Try a different approach" --resume <session_id> --fork-session ...

When the user follows up ("now also fix X"), reuse the previous session_id so Claude Code keeps full context. For an unrelated coding task, start fresh.

Worktrees

For non-trivial coding work, run inside a git worktree. The user always wants new work isolated on its own branch.

cd /path/to/repo
git worktree add .claude/worktrees/<slug> -b <branch-name>
cd .claude/worktrees/<slug>
claude -p "<task>" --output-format json --max-turns 20 ...

Report the worktree path and branch back to the user along with the result.

Useful flags

• --allowedTools <list> - comma-separated whitelist ("Read,Edit,Bash"). Scope to what the task needs.
• --max-turns <n> - cost cap; always set it. A confused subagent can otherwise loop indefinitely on your dollar. 5-10 for small tasks, 20-30 for refactors.
• --max-budget-usd <n> - dollar cap. Minimum ~$0.05.
• --permission-mode bypassPermissions - always use it. Without it the CLI prompts for permission and hangs forever in non-interactive mode.
• --output-format json - single result blob with metadata; always use it. You need session_id for follow-ups and result to report back; plain-text mode loses both.
• --model sonnet | opus | haiku - override default. haiku for cheap one-shots, opus for hard reasoning.
• --add-dir <path> - grant access to additional directories beyond cwd.
• --append-system-prompt-file <path> - add extra instructions on top of Claude Code's default prompt.
• --continue / --resume <id> - resume sessions (see above).
• --fork-session - branch off a session with a new id.