OmO Multi-Agent Orchestrator
Trigger with /omo and dispatch code analysis, bug investigation, and fixes to the minimal right mix of agents — no fixed pipeline.
Structurally least-privilege for read-only agents (explore, librarian, oracle) with explicit tool restrictions and transparent Context Pack data flow. Deductions: 'yolo': true for develop/oracle/librarian configs and the README FAQ advising Codex approval_policy='never' bypass execution confirmation; 2-hour unattended execution has no explicit user-confirmation gate; publisher identity unverified.
Input contracts are consistent across SKILL.md, README and reference files; examples are complete, so the happy path is plausible. Deductions: no tests covering this skill's own orchestration logic (CI tests codeagent-wrapper only); no documented failure feedback when codeagent-wrapper or the four backend CLIs are missing or fail; behavior depends on prompt-following by models and is not controlled.
Trigger is precise (/omo explicit command), routing signals and skipping heuristics are concrete, and non-fit ranges are described. Deductions: core function depends entirely on codeagent-wrapper plus overseas CLIs (Claude/Codex/Gemini/OpenCode) with no mainland-China reachability discussion; no Chinese-language support notes; boundaries such as concurrent-agent failure or oversized context are undeclared.
Good layered docs (SKILL.md → README → per-agent references), version (5.6.1), author, AGPL-3.0 license, and repo-level maintenance signals (CI, release workflow). Deductions: no changelog for this skill itself; codeagent-wrapper prerequisite not inlined in the skill docs; maintenance ownership (stellarlinkco vs cexll) not fully clear.
The claimed capability (risk-signal routing to a minimal agent set) is logically sound with arguable marginal value over single-agent use. Deductions: static review cannot verify actual multi-backend orchestration output quality; results depend entirely on external CLIs/models; direct-usability evidence is absent; cost (multi-model long runs, 2h timeout) vs benefit is not discussed.
All claims are traceable to repository source files and mutually corroborated (consistent input contracts across references) — auditable primary material. Deductions: no third-party execution evidence, no test suite or run records covering this skill's key paths; the librarian's permalink-evidence mechanism is unverified in execution.
- Configs contain 'yolo': true and the FAQ recommends approval_policy='never'; implementation agents may modify code and run commands without confirmation — tighten approval settings before use.
- The skill depends entirely on codeagent-wrapper and overseas CLIs (Claude/Codex/Gemini/OpenCode); it may be unreachable from mainland-China networks without a proxy.
- This is a static review with no execution; verify multi-agent output quality and cost (2h timeout, multi-model calls) yourself at small scale.
- Repository CI tests cover only codeagent-wrapper; the omo orchestration logic itself has no test coverage.
- Implementation agents have write and test-execution permissions; use in an isolated environment or under version control to enable rollback.
What does this skill do, and when should you use it?
OmO is a skill in the myclaude repository implementing a multi-agent orchestration workflow. It separates the orchestrator role (named Sisyphus) from executor agents: the orchestrator never writes code itself, instead choosing the minimal agent set based on task type and risk signals and passing full context between stages. Available agents include explore (code location), librarian (external docs lookup), oracle (high-risk review), develop (backend implementation), frontend-ui-ux-engineer (UI work), and document-writer, all invoked via codeagent-wrapper shell heredocs across Codex, Claude, Gemini, and OpenCode backends. It explicitly rejects a mechanical explore→oracle→develop conveyor belt in favor of signal-based routing and skipping.
Reads the /omo trigger and the user's original request; decides which agents to invoke based on risk signals (e.g., calling develop directly when the location is known and the change is low-risk); passes each agent an original request plus a context pack (prior agent outputs, known constraints), a current task, and acceptance criteria via codeagent-wrapper heredocs; executes in the shell tool with a 2-hour timeout; can run explore and librarian in parallel; produces location findings, root-cause analyses, implementation plans, or actual code changes per stage.
- A developer with a type error at a known file and line wants a minimal direct fix from the implementation agent without a full pipeline.
- A bug of unknown location needs explore to find the root cause, then develop to implement the fix and run relevant tests.
- Adding a feature that combines internal code patterns with external library docs, with explore and librarian researched in parallel and oracle reviewing risk before implementation.
- Understanding how a function works without changing code — explore alone returns signature, logic, and call-chain analysis.
- UI-only or docs-only changes routed to frontend-ui-ux-engineer or document-writer respectively.
What are this skill's strengths and limitations?
- Routing-first design skips unneeded agents based on risk signals, avoiding the waste of a fixed pipeline.
- Hard constraints — orchestrator never writes code, no self-serve grep for complex exploration — keep responsibilities clean and force full context passing.
- Multi-backend support (Codex/Claude/Gemini/OpenCode) avoids lock-in to a single execution engine.
- SKILL.md ships many concrete invocation examples plus an anti-example, reducing misuse.
- Strong dependency on codeagent-wrapper plus at least one backend CLI; environment setup is non-trivial.
- Per-invocation timeout is 2 hours; long tasks can be costly, and the source provides no cost information.
- It is one of 11 skills in a bundle — real-world results depend on wrapper and backend stability; the repo has no topics and the source shows no test suite or benchmark data.
- AGPL-3.0 license: closed-source commercial use requires contacting the author for a commercial license.
How do you install this skill?
Run npx github:stellarlinkco/myclaude for the interactive installer (installs to ~/.claude by default); or run npx github:stellarlinkco/myclaude --list to see installable items and pick individually; use --install-dir for a custom directory or --force to overwrite. Ensure codeagent-wrapper and at least one backend CLI (Codex/Claude/Gemini/OpenCode) are available afterwards. Module enablement is controlled via the modules.omo flag in config.; finer-grained details are undocumented in the source.
How do you use this skill?
In Claude Code, type a /omo command followed by your task, e.g.: /omo fix this type error at src/foo.ts:123 or /omo analyze this bug and fix it. The orchestrator routes by signal and invokes the appropriate agents. The source does not document a fallback behavior when codeagent-wrapper is missing.