llm-router

<!-- llm_router-rules-version: 1 -->
# LLM Router — VS Code / GitHub Copilot Routing Rules

> These rules apply when using llm-router MCP tools inside VS Code with GitHub Copilot.
> VS Code loads MCP servers from .vscode/mcp.json (workspace) or the user mcp.json.
> Use `llm_auto` for routing + cross-session savings tracking.

---

## How to Route

VS Code loads MCP servers on startup. Once llm-router is registered, call tools by name:

```
Rule: for research/generate/code/analyze tasks → call llm_auto(prompt=<the task>)
```

Add to `.github/copilot-instructions.md` for workspace-level routing guidance:

```
When a task requires live web search, call the llm_research MCP tool.
For code generation or complex analysis, call llm_code or llm_analyze.
For auto-routing with savings tracking, call llm_auto.
```

---

## Task Type → Tool Mapping

| Task | Tool | Notes |
|---|---|---|
| Simple factual question | `llm_query` | Gemini Flash / Groq |
| Research / current events | `llm_research` | Perplexity (web-grounded) |
| Writing / content | `llm_generate` | Gemini Flash / Haiku |
| Deep analysis | `llm_analyze` | GPT-4o / Gemini Pro |
| Code generation | `llm_code` | Ollama → Codex → o3 |
| Unknown / auto-detect | `llm_auto` | Classifies + routes |

---

## Savings Visibility

Run `llm_savings` in Copilot Chat to see cross-session totals.

---

## Token-Efficient Responses

Skip preamble. Lead with result. Fragments fine when meaning is clear.
No trailing summaries. ≥3 items → bullets. Never restate the user's request.


<!-- llm_router-rules-version: 2 -->
# LLM Router — VS Code / GitHub Copilot Routing Rules

> LLM Router routes each task to the cheapest capable model before Copilot generates
> its own response. Call the appropriate tool BEFORE answering.
>
> **Pull routing:** Copilot uses pull routing — YOU must call the tool. The tool
> returns the response; relay it verbatim. Requires VS Code ≥ 1.99 in agent mode.

## Task Type → Tool Mapping

| Task | Tool | Notes |
|---|---|---|
| Simple factual question | `llm(task="query")` | Ollama → Gemini Flash → GPT-4o-mini |
| Research / current events | `llm(task="research")` | Perplexity (web-grounded) |
| Writing / content | `llm(task="generate")` | Gemini Flash → Haiku |
| Deep analysis / review | `llm(task="analyze")` | Ollama → GPT-4o |
| Code generation / refactor | `llm(task="code")` | Ollama → Codex → o3 |
| Deep reasoning / proofs | `llm(task="analyze")` | Extended-thinking model |
| Unknown / auto-detect | `llm_auto` | Classifies + routes automatically |

## Rules

1. Call the matching tool **before** generating your own response.
2. Return the tool's output to the user **as-is** (don't paraphrase).
3. Only use native Copilot intelligence for file navigation, terminal commands,
   or when all LLM Router MCP servers are unavailable.
4. If `llm(task="code")` or similar is unavailable, proceed normally and note it.

## Tips

- Use `llm(task="research")` for anything time-sensitive — Ollama has a training cutoff
- Use `llm_auto` when unsure which tool to call
- Run `llm_router_status` to see cross-session totals

## Token-Efficient Responses

Skip preamble. Lead with result. Fragments fine when meaning is clear.
No trailing summaries. ≥3 items → bullets. Never restate the user's request.