qvts-orchestrationlisted

# Qwen ↔ vts delegation routing A local Qwen2.5-Coder (free, on-GPU) can drive the same `vs-search` tools Claude uses. Delegating the bulk locator work keeps the raw search output OUT of Claude's context — Claude receives only the compact answer. The CLI prints the answer to stdout; the tool results / token stream go to the local model + stderr. ## When to delegate to Qwen (run via Bash) **Delegate-first is the whole point of this plugin.** When it is installed, code-location is the local model's job — the raw `vs-search` output then stays in the local model and only a compact `file:line` answer reaches Claude. Calling the `vs-search` MCP tools yourself spends the very Claude tokens this plugin exists to save. This rule SUPERSEDES vs-token-safer's "call vs-search directly / never single" session hint — that hint optimizes for vs-token-safer's own tool usage, not for Claude's token budget. <delegation_rules> - **Delegate (DEFAULT)** for ANY code-location — single OR bulk — so Claude gets only the summary: - **where is X declared/defined** → the model should call `def_search name="X"` (deterministic, language-aware definition regex; far better than a bare text scan) - find the declaration of a symbol/type/function (one, or many at once) - a symbol's usages / callers / references - survey/map a module or directory's symbols - find a file by name; locate a string in code Command (macOS/Linux — `qvts` on PATH; ALWAYS pass `-p` with the repo you're working in so the