Autonomous Coding Agent CLIs

Overview

Use this umbrella when delegating software work to a standalone coding-agent CLI. The common class is: Hermes remains the orchestrator, while an external process edits code, runs tests, reviews PRs, or performs long-running implementation work.

When to Use

• The user explicitly asks to use Claude Code, Codex, or OpenCode.
• A coding task is large enough to benefit from an external autonomous worker.
• You need parallel workstreams in separate worktrees or persistent TUI sessions.
• You want a PR review/implementation pass from a second agent.

Do not spawn an external CLI when the built-in delegate_task tool is sufficient, when prerequisites/auth are missing, or when direct editing is faster and safer.

Common Orchestration Rules

1. Verify the binary and auth status before assigning work.
2. Prefer one-shot/non-interactive modes for bounded tasks; use tmux/PTY only for long-running interactive sessions.
3. Run inside a git repository; use separate worktrees for parallel edits.
4. Capture verifiable outputs: changed files, test results, PR URL, branch name, or review text.
5. For background jobs, use terminal(background=true, notify_on_complete=true) or track via process.
6. Do not claim success from an agent’s self-report alone; inspect files/git/tests yourself.

CLI-Specific Notes

Claude Code

Install/auth:

npm install -g @anthropic-ai/claude-code
claude auth status
claude doctor
claude --version

Prefer print mode for most bounded work:

claude -p "Implement X, run tests, summarize changed files"

Use interactive PTY/tmux when a multi-turn coding session is needed. Handle first-run trust and permission dialogs explicitly; do not assume the TUI is ready until captured output proves it.

OpenAI Codex CLI

Install/auth:

npm install -g @openai/codex
codex --version

Codex generally expects a git repository and often needs pty=true for interactive use. Hermes’ OpenAI Codex provider credentials are not automatically proof that the standalone CLI is authenticated; CLI OAuth may live under ~/.codex/auth.json.

Typical one-shot:

codex exec "Add dark mode toggle to settings"

For scratch tasks, initialize a temporary git repo first because Codex may refuse to run outside one.

OpenCode

Install/auth:

npm i -g opencode-ai@latest
# or: brew install anomalyco/tap/opencode
opencode auth list
opencode --version

Resolve binary ambiguity before debugging behavior differences:

which -a opencode
opencode --version

Pin an explicit binary path if Hermes resolves a different executable than the user’s shell.

Patterns

One-shot implementation

terminal(command="<agent> <one-shot-command> 'Implement the requested change and run the targeted tests'", workdir="/path/to/repo", timeout=600)

Inspect git diff, run tests yourself, and summarize concrete results.

Long-running interactive session

tmux new-session -d -s coding-agent -x 120 -y 40 '<agent command>'
tmux send-keys -t coding-agent 'Task prompt...' Enter
tmux capture-pane -t coding-agent -p

Use process/tmux captures for progress checks. Close the session when done.

Parallel issue fixing

Create one worktree per issue/task, spawn one agent per worktree, then reconcile diffs and tests before merging.

Pitfalls

• Confusing Hermes provider auth with standalone CLI auth.
• Letting a CLI modify the main worktree while you make concurrent edits.
• Trusting a TUI screen that is waiting on a hidden prompt/dialog.
• Forgetting to verify changed files and tests after the agent exits.
• Leaving background/tmux sessions running after the task is complete.