External Coding Agents

External AI Coding Agents

Delegate coding work to third-party AI coding agents from within Hermes. These tools (Claude Code, Codex, OpenCode) are autonomous CLI agents that can read files, write code, run shell commands, and manage git workflows — each with its own strengths.

When to Delegate vs Do It Yourself

| Situation | Approach | |-----------|----------| | Bug fix in a known file | Do it yourself with patch/write_file — faster, no overhead | | Complex multi-file refactor | Delegate — the external agent has persistent context across files | | Task needs a fresh perspective | Delegate — the external agent will spot things you might miss | | CI/CD automation | Delegate — use print mode for clean automation | | PR review of someone else's code | Delegate — the external agent can see the full diff in context | | Research/reconnaissance | Delegate — agent can browse and explore independently | | Simple one-line change | Do it yourself |

Common Orchestration Patterns

All three coding agents use the same core patterns:

Print Mode (Non-Interactive — Preferred)

Cleanest integration. No PTY, no dialog handling, structured output:

terminal(command=" -p 'Add error handling to all API calls in src/'", workdir="/path/to/project", timeout=120)

PTY Background Mode (Interactive)

For iterative work needing multi-turn conversation. Use background=true, pty=true and process tool to monitor:

terminal(command="", workdir="~/project", background=true, pty=true)
process(action="submit", session_id="", data="Implement OAuth refresh flow")
process(action="poll", session_id="")
process(action="log", session_id="")
process(action="kill", session_id="")

Get the Diff Back

After delegation completes, always inspect what changed:

git diff
git diff --stat

Agent Decision Guide

| Need | Best Agent | Why | |------|-----------|-----| | Best overall coding ability | Claude Code | Anthropic's most capable model, v2.2+, rich feature set | | Cheapest / open-source preference | OpenCode | Provider-agnostic (OpenRouter, Anthropic, OpenAI, local models) | | OpenAI ecosystem | Codex | OpenAI's autonomous agent, OAuth integration | | PR review with inline comments | Claude Code | Built-in --from-pr flag, formal review workflow | | Batch multi-issue fixing | Codex or OpenCode | Lightweight, parallel worktree pattern | | Long-running autonomous session | Any | Use background=true + process tool for monitoring |

Common Pitfalls

1. PTY is required for interactive sessions — all three agents are TUI apps; skip pty=true and they hang
2. Git repo required — Codex and OpenCode refuse to run outside a git directory. Use mktemp -d && git init for scratch work
3. Always inspect the diff after delegation — external agents can make surprising decisions. Use git diff --stat to verify scope
4. Keep tasks focused — narrow prompts produce better results than vague requests
5. Set timeouts — external agents can loop indefinitely. Use --max-turns (Claude Code) or timeout in terminal calls
6. Cleanup background sessions — kill tmux sessions and worktree directories when done
7. Parallel worktrees need isolation — don't share a single working directory across parallel agents

Agent-Specific References

Detailed reference docs for each agent:

• references/claude-code.md — Full Claude Code orchestration guide (flags, auth, hooks, MCP, slash commands, 745 lines)
• references/codex.md — Codex CLI usage (flags, background mode, worktrees)
• references/opencode.md — OpenCode CLI usage (binary resolution, parallel patterns, TUI keybindings)

Getting Started

Which Agent is Installed?

which claude codex opencode 2>/dev/null || echo "check npm list -g"

Quick One-Shot (any agent)

# Claude Code
terminal(command="claude -p 'Add input validation to all endpoint handlers' --max-turns 10", workdir="~/project", timeout=120)

# Codex
terminal(command="codex exec 'Add input validation to all endpoint handlers'", workdir="~/project", pty=true)

# OpenCode
terminal(command="opencode run 'Add input validation to all endpoint handlers'", workdir="~/project")

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: S3YED
• Source: S3YED/appie-kit
• License: MIT
• Homepage: https://weblyfe.ai

Install and usage instructions live in the source repository linked above.