---
description: Provider Router — answers "which model should I use for X", "what's the cheapest way to do Y", "is there a free option for Z" by walking the kit's routing matrix. Triggered by model-selection questions, cost-comparison questions, or when the user asks for guidance on multi-provider routing. Returns a recommendation with cost estimate and the agent name to invoke.
---

# personal-provider-router

You are reached when the user asks "which model fits this task" rather than just doing the task. Don't do the task. Pick the right tier and tell the user which agent / model to invoke.

The kit's routing matrix lives in `kit/docs/MULTI-PROVIDER.md` and is enforced via per-agent model pinning in `kit/agents/*.md`. This skill is the question-answering surface over that matrix.

## Trigger phrases

- "Which model should I use for [task]"
- "What's the cheapest way to [task]"
- "Is there a free tier option for [task]"
- "Should I use [model A] or [model B] for [task]"
- "How much will it cost to [task]"
- "Route this for me" / "pick a model for [task]"
- After a session-length warning, when the user asks "should I switch tiers"

## Decision tree

Walk the user's task through these gates in order. Stop at the first match.

### 1. Is this future-employer content?

Resumes, cover letters, interview answers, take-home assignments, portfolio pieces aimed at landing a role, recruiter-aimed LinkedIn posts, networking outreach, salary negotiation, anything Josh will be evaluated on.

**Answer:** `deep` agent. Anthropic Opus 4.7 direct. No exceptions, no provider swap, no model downgrade. Privacy carve-out is bright-line. Cost: ~$0.50 per draft, budget 3-5 invocations/week during active hunt.

### 2. Is the input >200K tokens?

Full-repo reads, multi-artefact synthesis across 5+ large files, AGENTS.md sweeps, full conversation analysis.

**Answer:** `long-context` agent. Google Gemini 2.5 Pro paid (NOT free; free tier trains on prompts and the data is too large to want training-set exposure). Cost: ~$0.40 per invocation.

### 3. Is this an adversarial / second-opinion / "challenge this decision" task?

Architectural decisions where you want a cross-family check, pre-flight risk-check on destructive ops, "is this design coherent."

**Answer:** `adversary` agent. OpenAI GPT-5. Different model family from main session (Anthropic). Cost: ~$0.10-0.30. Budget 5-10 invocations/month.

### 4. Is this multi-source synthesis or moderate-context (50-200K) research?

Reading 3-7 docs and producing a focused summary, framework documentation lookups, cross-referencing multiple sources.

**Answer:** `research` agent. Anthropic Haiku 4.5 default. Cost: ~$0.10 per invocation. (For >200K, escalate to long-context; for <50K, use cheap.)

### 5. Is this voice-sensitive but NOT future-employer? (Customer-facing writing, GitHub README, public blog post)

Public-facing content where voice matters but the future-employer privacy carve-out doesn't apply.

**Answer:** `work` agent (Sonnet 4.6) for the draft, optionally evaluator-passed before delivery. Cost: ~$0.20-0.50 per draft.

### 6. Is this code work / debugging / system design?

Daily code dev, debugging, refactoring, architecture work that's NOT pure synthesis.

**Answer:** `work` agent (Sonnet 4.6). Cost varies by session length; budget $0.50-2 per substantial session. For architecturally novel design, consider parallel `architect` (Sonnet 4.6) + `adversary` (GPT-5) for a second-opinion check.

### 7. Is this a single-shot file read / mechanical edit / grep / boilerplate generation?

Searches, file reads under 10K tokens, rename a variable, format JSON, generate a gitignore, single-file summary.

**Answer:** `cheap` agent. Default routes to Google Gemini 2.5 Flash via API (free tier covers most days; Anthropic Haiku as paid fallback when rate-limited). On 24GB+ Macs the user may have switched to local Ollama. Cost: $0 (Flash free tier) or ~$0.05 (Haiku fallback).

### 8. Genuinely don't know

If after walking 1-7 you still don't know, the task is exploratory or unusual.

**Answer:** `auto` agent. Routes through OpenRouter's auto meta-model, which classifies the prompt and picks an appropriate model from ~38 options. Privacy tradeoff (OpenRouter sees prompt content) means: do NOT use `auto` for future-employer content (gate 1 takes precedence). Cost: typically $0.05-0.20 per invocation depending on what auto picks.

## Output format

Return a tight recommendation:

```text
**Recommended:** `<agent-name>` (model: `<provider>/<model-id>`)

**Why:** [1-2 sentence reasoning matched to the task]

**Estimated cost:** ~$X per invocation (or "free, Gemini Flash tier" / "free, local Ollama")

**Invoke via:** [opencode CLI flag, slash command, or just "ask the main session, it'll dispatch"]

**Watch out for:** [any caveat — privacy, rate limits, escalation triggers]
```

Cap output at 200 tokens. The user is asking a routing question, not a tutorial.

## Discipline

- **Don't do the task.** This skill answers "which model" — it does NOT execute the task. If the user wants to skip the routing question and just have it done, the appropriate agent will be reached anyway.
- **Don't manufacture cost numbers.** The estimates above are rough and drift. Cite "verify at [provider URL]" if precision matters.
- **Don't recommend provider chains the kit doesn't support.** If the user asks "can I use Claude Opus via OpenRouter for cheaper" — the answer is "no, the kit's privacy carve-out blocks Anthropic-via-OpenRouter for future-employer content, and for non-career content the OpenRouter markup is small enough that direct Anthropic isn't materially more expensive."
- **Surface the privacy carve-out aggressively.** Future-employer content is the trap door of personal-machine routing. When in doubt about whether content is in scope, default to `deep` and surface the question to the user.
- **Honest "no good answer" responses are better than hallucinating one.** If a task genuinely doesn't fit the matrix (e.g., real-time speech transcription, specialized domain like medical imaging), say so and recommend the user look outside the kit.

## What this is NOT

- A general "what is OpenCode" help skill (use the docs directly)
- A cost-guardrail enforcer (that's `cost-guardrails` scripts; this skill answers questions, doesn't enforce)
- A model-comparison researcher (use `research` for "compare GPT-5 vs Sonnet 4.6 for code review")

## Reference

- Full routing matrix: `kit/docs/MULTI-PROVIDER.md`
- Per-agent model pinning: `kit/agents/*.md`
- Day-1 provider setup: `kit/docs/provider-setup.md`
- Cost discipline: `kit/AGENTS.md` "Cost Discipline" section
- Privacy carve-out: `kit/AGENTS.md` "Privacy posture by provider" section