Qvts orchestration

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 model targets it): qvts -p "<repo-root>" --json "<natural-language locate task>" (Plugin/Windows fallback: node "${CLAUDE_PLUGIN_ROOT}/vts-bridge.mjs" --json "<task>" or pwsh -File "${CLAUDE_PLUGIN_ROOT}/qvts.ps1" -Json "<task>".) → stdout JSON {task, answer, trace}. Trust the answer's file:line; read bodies yourself with read_symbol. Keep the warm daemon up (VTS_AUTO_DAEMON=1) so delegated locates return in ~seconds, not a cold spawn.

• Do it yourself with the vs-search MCP tools ONLY when:

  • the delegated answer came back empty / "no match" / TOOL ERROR (fall back, then retry the search yourself)
  • it's a trivial peek at a file you JUST edited (already in context)
  • the task needs multi-step reasoning, design, review, or edits (a small local model is weak at these) </delegation_rules>

<delegation_protocol>

1. Treat the returned file:line as fact; if you need a declaration's body, read it directly (read_symbol/Read).
2. If the answer is empty / "no match" / contains TOOL ERROR, retry the search yourself via vs-search.
3. Writes/edits are Claude's job (or an explicit apply) — delegation is read-only.
4. UE/C++ symbol & reference tools need a clangd compile DB (one-time vts_admin gen_compile_db); without it, delegate only search_text / find_files style locates, which need no clangd. </delegation_protocol>

Delegate READING too — not just locating

The same token rule applies to INGESTING text, not only searching for it. When you'd otherwise Read a big file (or a whole module) just to survey/summarize/assess it — e.g. "which handlers spew large output", "what does this file do" — DON'T pull the raw bytes into your context. Delegate the READ; only a compact brief returns:

• qvts digest "<file>" --focus "<question>" → shortest faithful brief (the local model reads the file).
• qvts digest-dir "<dir>" --focus "<question>" → per-file briefs + an overview for a whole module.
• qvts triage-diff [--staged] → a git diff → {summary, hotspots[], open[]} so you open only flagged files. This is a CONTEXT-SAVING SUMMARY (the local model reads the diff), NOT an impact analysis. To judge a change's BLAST RADIUS + a deterministic risk band (LOW/MED/HIGH, computed from the LSP call graph + git co-change coupling), call vs-token-safer's detect_changes MCP tool directly — it is exact and reproducible where triage-diff is an LLM summary. Split of labor: triage-diff decides WHICH files to open; detect_changes judges HOW RISKY the change is ("is this safe to commit", "what does this touch", review). detect_changes is a multi-step LSP+git+risk composition, not a single locate — never route it to the local model (it can't drive it); Claude calls it directly.
• qvts vcs <p4|git> <read-only sub> [args] [--focus "..."] → run a big version-control query (e.g. qvts vcs p4 opened, qvts vcs git status, qvts vcs git log -20) and get back a short summary instead of the raw dump. Read-only subcommands only (mutating p4/git ops are refused). Don't run p4 opened / git status / git log in Bash just to read them — the full list lands in your context; route them here.

All are content-cached and credit the savings ledger. Read a file directly only when you need its exact bytes to EDIT it, or it's small / just-edited (already in context). This is the "delegate reading, not just searching" axis — the most common missed save is Claude reading a 1–10k-line source file to answer a question a 3-second local digest could.

Prereq: setup has run (setup-macos.sh on macOS/Linux, setup.ps1 on Windows) — the configured model (gemma4-vts by default) loaded (ollama ps → 100% GPU) and qvts.config.json written. See ORCHESTRATION.md / USAGE.md.