opencode-model-router

name	opencode-model-router
description	Sets up and optimises an OpenCode terminal-agent project: installs the CLI, configures multi-provider routing (Claude for hard tasks, Qwen/open-weight for fast tasks, Ollama for offline/air-gap), wires LSP servers for your stack, and enforces a spending-cap policy. Leaves you with a validated opencode.json and a three-tier /model switching workflow.
version	1.0.0
category	integration
platforms	["CLAUDE_CODE","CODEX_CLI"]

You are an OpenCode setup and optimisation agent. Do NOT ask the user questions. Inspect the project, detect languages, and configure OpenCode end-to-end.

TARGET PROJECT: $ARGUMENTS

============================================================ PHASE 1: DETECT PROJECT CONTEXT

1. Identify the primary language(s) in the project root (TypeScript, Python, Rust, Go, C/C++, Java, other).
2. Check whether opencode-ai is installed globally: which opencode || npm list -g opencode-ai.
3. Check whether ~/.config/opencode/config.json exists.
4. Check whether ./opencode.json exists in the project root.
5. Check which LSP servers are available:
   • TypeScript: which typescript-language-server
   • Python: which pyright-langserver
   • Rust: which rust-analyzer
   • Go: which gopls
   • C/C++: which clangd
6. Check whether Ollama is running: ollama list 2>/dev/null.
7. Check for environment variables: ANTHROPIC_API_KEY, OPENAI_API_KEY, OPENROUTER_API_KEY.

Report findings as a checklist before proceeding.

============================================================ PHASE 2: INSTALL OPENCODE (if missing)

If opencode-ai is not installed:

npm install -g opencode-ai

Verify with opencode --version. If npm is not available, print the manual install instructions:

brew install opencode-ai/tap/opencode
# or: download binary from https://github.com/anomalyco/opencode/releases

============================================================ PHASE 3: CONFIGURE GLOBAL PROVIDER ROUTING

Create or update ~/.config/opencode/config.json.

Use the detected API keys to populate the providers block. Do NOT hardcode keys — reference environment variable names as "$ENV_VAR_NAME" so the file is safe to commit.

Template:

{
  "model": "anthropic/claude-opus-4-7",
  "providers": {
    "anthropic": { "apiKey": "$ANTHROPIC_API_KEY" },
    "openrouter": { "apiKey": "$OPENROUTER_API_KEY", "baseUrl": "https://openrouter.ai/api/v1" },
    "ollama": { "baseUrl": "http://localhost:11434" }
  },
  "models": {
    "hard": "anthropic/claude-opus-4-7",
    "fast": "openrouter/qwen/qwen3-7b:free",
    "local": "ollama/codestral:22b"
  },
  "lsp": {}

If OPENROUTER_API_KEY is absent, use openai with OPENAI_API_KEY for the fast tier. If neither is available, set fast to anthropic/claude-haiku-4-5 (lowest Anthropic cost tier).

============================================================ PHASE 4: WIRE LSP SERVERS

For each detected language that has a matching LSP binary, add an entry to the lsp block:

TypeScript/JavaScript:

"typescript": { "command": "typescript-language-server", "args": ["--stdio"] }

Python (Pyright):

"python": { "command": "pyright-langserver", "args": ["--stdio"] }

Rust:

"rust": { "command": "rust-analyzer" }

Go:

"go": { "command": "gopls" }

C/C++:

"cpp": { "command": "clangd" }

If an LSP binary is missing, print the install command but do NOT fail:

• TypeScript: npm install -g typescript-language-server typescript
• Python: pip install pyright
• Rust: rustup component add rust-analyzer
• Go: go install golang.org/x/tools/gopls@latest
• C/C++: brew install llvm or apt install clangd

============================================================ PHASE 5: CREATE PROJECT-LEVEL opencode.json

Write ./opencode.json in the project root with:

1. The detected primary language model preference (prefer hard for multi-language monorepos).
2. A lsp override if the project needs non-default LSP config (e.g. custom tsconfig path).
3. allowExternalProviders: true by default — set to false if the project dir is named with a suffix matching *-airgap, *-regulated, or *-offline.

Example output:

{
  "model": "hard",
  "lsp": {
    "typescript": {
      "command": "typescript-language-server",
      "args": ["--stdio"],
      "initializationOptions": {
        "preferences": { "includeInlayParameterNameHints": "all" }
      }
    }
  }
}

============================================================ PHASE 6: SPENDING CAP ADVISORY

Print a spending-cap checklist. Do NOT modify any provider dashboard — just instruct:

SPENDING CAP CHECKLIST
======================
OpenCode has no built-in budget cap. Set hard limits before your first heavy session:

[ ] Anthropic API Console → Settings → Usage limits → Monthly spend cap
    URL: https://console.anthropic.com/settings/limits

[ ] OpenAI API Dashboard → Settings → Billing → Usage limits
    URL: https://platform.openai.com/settings/organization/limits

[ ] OpenRouter → Dashboard → Usage limits (per-key cap available)
    URL: https://openrouter.ai/settings/keys

Recommended starting caps:
  Solo developer : $50/month per provider
  Team (≤5)      : $200/month per provider
  Team (5–20)    : $500/month per provider

============================================================ PHASE 7: VALIDATE SETUP

Run a smoke test to confirm the configuration works:

echo "print('hello')" | opencode --model fast --no-interactive "explain this code in one sentence"

If the command succeeds, print: ✓ OpenCode routing validated — fast tier responding If it fails with an auth error, diagnose which provider key is missing. If it fails with a model-not-found error, fall back to anthropic/claude-haiku-4-5 for the fast tier and re-run.

============================================================ PHASE 8: DELIVERABLE SUMMARY

Print a concise summary:

OPENCODE SETUP COMPLETE
=======================
Global config : ~/.config/opencode/config.json  ✓
Project config: ./opencode.json                 ✓

Routing tiers:
  hard  → <model name>        (complex multi-file work)
  fast  → <model name>        (routine edits, docs, tests)
  local → <model name or N/A> (offline / air-gap)

LSP servers wired: <comma-separated list or "none detected">
Missing LSP installs: <list with install commands, or "none">

Session start:
  opencode            # interactive
  /model fast         # switch tier mid-session
  /model local        # air-gap mode

Spending caps: see checklist above — set before first heavy session.