qwen-agent

qwen-agent is a skill for delegating menial, well-scoped coding tasks to a cheaper Qwen-backed subagent that runs inside a headless Claude Code instance invoked through the `claude-9arm` shell alias. The goal is cost control: keep expensive Claude reasoning for work that needs it and offload mechanical jobs — bulk renames, formatting, boilerplate, find-replace, grep-style search and summarization, log/file condensing, docstring and comment scaffolding, and running builds/linters/tests to report pass/fail.

The skill's core guidance is operational. Because the subagent inherits zero context from the current conversation, every delegated prompt must be fully standalone with absolute paths, explicit inputs/outputs, and acceptance criteria. It explains running the subagent headless with `-p` and scoping tool access with `--allowedTools Bash Read Edit Write Glob Grep` so the job finishes unattended, plus a lower-risk `--permission-mode acceptEdits` for edit-only work. It stresses sizing tasks to Qwen's 128k context window, splitting large jobs into bounded per-file/per-directory chunks, using absolute paths (since `cd` resets between Bash calls) or `--add-dir`, and a return contract covering text output, JSON parsing, and background/parallel runs. A workflow checklist ends with the essential rule: verify the cheaper model's output yourself before reporting success.

It targets developers and agent operators who want to reduce token spend on low-judgment work. The skill is explicit about what NOT to delegate — architecture, design, debugging judgment, security-sensitive edits, or anything needing conversation context. The main consideration is that it runs an autonomous subagent with pre-approved Bash and file-write tools, and optionally suggests a settings rule to suppress per-call permission prompts, which reduces human oversight of a less reliable model.