One-line install

Install it as a Hermes plugin:

hermes plugins install https://github.com/xuyang-liu16/hermes-code-bridge --enable

Then use it inside Hermes:

/code-bridge Use Codex to do a read-only review of the current repository diff.

The plugin registers /code-bridge, which tells Hermes to load and follow the plugin-provided hermes-code-bridge:hermes-code-bridge skill for the request.

If you only want the skill file without the plugin wrapper:

hermes skills install https://raw.githubusercontent.com/xuyang-liu16/hermes-code-bridge/main/skills/hermes-code-bridge/SKILL.md --name hermes-code-bridge

What it does

Hermes Code Bridge connects Hermes Agent to local terminal-based coding agents such as Codex, Kimi Code, Claude Code, OpenCode, Gemini CLI, and other coding CLIs.

It teaches Hermes how to:

• discover installed coding CLIs;
• choose the right backend, working directory, and session;
• reuse existing CLI-agent sessions instead of creating unnecessary new ones;
• write structured dispatch prompts with role, task, constraints, success criteria, and report format;
• run the real local CLI, not a fake substitute;
• monitor background execution through terminal, tmux, or process logs;
• collect raw evidence, artifacts, diffs, and verification results;
• report what happened, what passed, what failed, and what remains risky.

The bridge pattern:

User request
  -> Hermes plans and routes the task
  -> Codex / Kimi Code / Claude Code / OpenCode executes it locally
  -> Hermes monitors, verifies, and reports evidence

Hermes is the coordinator. Your local coding CLIs are the execution backends.

Why use it?

Modern developers often use several coding agents at once. One tool might be better for implementation, another for review, another for debugging, and another for long-context research. Without a workflow, it is easy to lose session context, send vague prompts, forget verification, or accidentally pretend that one agent did work that another agent actually did.

Hermes Code Bridge gives Hermes a disciplined bridge workflow:

Supported backends

The skill includes command patterns and safety notes for:

• Codex
• Kimi Code
• Claude Code
• OpenCode
• Gemini CLI
• generic terminal-based coding assistants
• tmux-based interactive sessions
• optional multi-agent workspace tools such as CCB (claude_codex_bridge)

The command recipes are intentionally written as patterns because CLI flags change over time. Hermes should still check <command> --help when using a new version.

Usage

Load by plugin slash command

/code-bridge Use Kimi Code to implement the smallest change that fixes this bug. Reuse the existing project session if available, do not refactor unrelated files, run relevant tests, and report evidence.

The plugin registers /code-bridge, which tells Hermes to load and follow the hermes-code-bridge skill for the request.

Load as a normal skill

/skill hermes-code-bridge

or start Hermes with it preloaded:

hermes -s hermes-code-bridge

Example prompts

Ask Codex for a read-only review:

Use Codex to do a read-only review of the current repository diff. Do not modify files. Report correctness, security, and maintainability risks with evidence.

Ask Claude Code to review a change:

Use Claude Code as a read-only reviewer for the latest diff. Do not modify files. List blockers, non-blockers, test gaps, and exact evidence from the diff.

Ask OpenCode to inspect a project:

Use OpenCode to inspect this repository structure and suggest where a new feature should be implemented. Read-only only; do not create or edit files.

Coordinate implementation and review:

Use one local coding agent to implement the change and a different one to review it. Confirm the backend/session plan before dispatch. Use separate worktrees if both agents need to edit files.

Plugin vs Skill

This repository ships both:

The plugin is intentionally lightweight. The real workflow lives in SKILL.md, so users can install it either way.

Safety model

Hermes Code Bridge is strict about attribution and evidence:

• If the user asks for Codex, Hermes should run Codex.
• If the user asks for Kimi Code, Hermes should run Kimi Code.
• Hermes should not use a Hermes subagent or Python script and pretend it was a different local coding CLI.
• Hermes should not edit coding-agent session databases or internal history files.
• Hermes should not bypass sandbox or approval prompts unless the user explicitly accepts that risk.
• Hermes should verify artifacts and test results before reporting success.

Relationship to CCB and tmux workspaces

Hermes Code Bridge is not a replacement for full multi-agent workspace tools.

Tools such as CCB (claude_codex_bridge) provide visible tmux workspaces, configured agent panes, sidebars, worktrees, and inter-agent communication routes. Hermes Code Bridge is lighter: it is a Hermes skill/plugin that helps Hermes drive whatever local coding CLIs are already installed.

They can work together. If CCB is installed, Hermes can treat it as another bridge backend: inspect the CCB config, attach to the workspace, send prompts to the correct pane, and capture output.

Repository layout

hermes-code-bridge/
  README.md
  README_zh.md
  LICENSE
  plugin.yaml
  __init__.py
  after-install.md
  skills/
    hermes-code-bridge/
      SKILL.md

Privacy

This repository is designed to be public. The skill uses placeholders such as <PROJECT_DIR>, <SESSION_ID>, <PROMPT>, and <TEST_COMMAND> instead of personal paths, private project names, real session IDs, or credentials.

Before publishing your own fork, run a privacy scan:

grep -RInE "(/Users/|/home/|API_KEY|TOKEN|SECRET|PRIVATE|@)" . 2>/dev/null || true

Review any matches manually. Some placeholders may be intentional; real secrets or personal data should be removed.

License

MIT