Codeagent Multi-Backend Orchestration Skill
Run AI code tasks across Codex, Claude, Gemini, and OpenCode backends through one codeagent-wrapper command, with parallel multi-agent orchestration built in.
Data flow is transparent; worktree isolation, read-only agent presets, and explicit allowed/disallowed tool enumeration help safety. Major deductions: --skip-permissions exists and CODEAGENT_SKIP_PERMISSIONS defaults to true, CODEX_BYPASS_SANDBOX defaults to true — dangerous defaults that bypass user confirmation and sandboxing; no rollback guidance. Mid-range score.
Internally consistent docs with exit-code table, progress-frame handling, failure output format, topological dependency sorting with cycle detection; CI runs Go tests for codeagent-wrapper. Capped at 10 per static rules: tests cover the wrapper, not this skill's key paths, and failure feedback is unverified by execution.
Scenarios, boundaries (read-only vs develop agents), input/output formats and triggers are well described. Deducted for full dependence on four overseas CLIs (Codex/Claude/Gemini/OpenCode) with poor mainland-China reachability and no Chinese-language support.
Clear layered structure, AGPL-3.0 license, version 6.7.0 in package.. Deducted: no skill-level changelog, no version mapping, no known-limitations section, unverified publisher, maintenance responsibility only implicit.
Multi-backend orchestration, parallel tasks, session resume and structured JSON output are fully described and plausible. Deducted: static review cannot confirm outputs are directly usable, no representative-output evidence, and users must pre-install four external CLIs, raising cost.
Real CI (three-OS matrix) plus a committed Go test suite covering the skill's core dependency codeagent-wrapper provide partial corroboration. Capped at 5 per static rules; no independent reproduction of the skill path itself.
- CODEAGENT_SKIP_PERMISSIONS and CODEX_BYPASS_SANDBOX default to true: deployment silently bypasses permission prompts and sandboxing. Explicitly set them to false and audit any --skip-permissions usage.
- Core function depends on four overseas CLIs (Codex/Claude/Gemini/OpenCode); effectively unusable from mainland-China networks without a proxy or substitute backends.
- Static review only — no code was executed: orchestration effectiveness, failure feedback and output quality are unverified; pilot in an isolated repository before sensitive or production use.
- develop-type agents directly modify code and run tests; always pair with --worktree so changes are recoverable.
What does this skill do, and when should you use it?
The codeagent skill teaches Claude Code how to invoke the codeagent-wrapper CLI to dispatch coding tasks to four pluggable backends: Codex, Claude, Gemini, and OpenCode. It bundles agent presets (code-explorer, code-architect, code-reviewer, develop), auto-injects language best-practice skills based on project files, supports git worktree isolation, and runs parallel tasks with dependency-based topological sorting. Execution defaults to background, producing structured JSON reports and resumable session IDs. It ships inside the stellarlinkco/myclaude repository under the AGPL-3.0 license as one of eleven bundled skills.
Reads marker files in the working directory (go.mod, package., etc.) to auto-select skills for prompt injection; builds and executes codeagent-wrapper commands specifying backend, model, reasoning effort, and agent preset; reads task configs from stdin in --parallel mode, sorts them with Kahn's algorithm, and runs independent tasks concurrently; executes write tasks inside isolated git worktrees (branch do/{task_id}); produces per-task pass/fail reports with log paths and structured JSON output, returning a SESSION_ID for follow-up runs.
- Developers who want to mix backends in one project: hand algorithm-heavy work to Codex, docs and prompt engineering to Claude, UI prototyping to Gemini.
- Engineers leading large refactors: run read-only code-explorer and code-architect passes first, then apply changes via develop in an isolated worktree.
- Team leads running multi-phase pipelines (explore → design → implement → review) in parallel, with dependencies scheduling task order automatically.
- Maintainers of polyglot monorepos who rely on auto-detection to inject Go/Rust/Python/frontend best practices per project.
- Automation pipelines needing auditability, using --output to emit structured JSON results plus per-task logs.
What are this skill's strengths and limitations?
- One entry point unifies four backends, letting you match model to task type.
- Built-in read-only/writable agent presets and auto skill injection cut down on hand-written prompts.
- Parallel execution includes dependency topological sorting and circular-dependency detection.
- Worktree isolation and structured JSON output make it easy to wire into existing automation.
- Session resume support and per-task logs simplify debugging failed runs.
- Requires the external codeagent-wrapper binary plus at least one backend CLI, raising setup cost.
- The repo ships no test suite or standalone evidence for this skill; the README mostly describes the whole module collection.
- AGPL-3.0 is viral for closed-source commercial integration; commercial use requires a separate license.
- Some behavior assumes Claude Code workflow conventions (background execution, Bash tool parameters) and needs adjustment in other clients.
- The source provides no version number, stability guarantees, or cost comparison across backends.
How do you install this skill?
Run the interactive installer: npx github:stellarlinkco/myclaude (use --list to see installable items, --install-dir for a custom directory, --force to overwrite). After install, codeagent-wrapper lives at ~/.claude/bin/ and this skill at ~/.claude/skills/codeagent/. The source does not document a supported way to install this single skill in isolation from the collection.
How do you use this skill?
Single task: codeagent-wrapper --agent develop --skills golang-base-practices - . <<'EOF' …task… EOF (run via the Bash tool in the background). Parallel: codeagent-wrapper --parallel <<'EOF', defining tasks with ---TASK---/---CONTENT--- blocks containing id, agent, dependencies, etc. Long tasks emit [codeagent-progress] frames — do not kill them; resume with resume <session_id>. Configure via ~/.codeagent/config.(yaml||toml) or CODEAGENT_-prefixed environment variables.