# Hive AGENTS.md Routing Snippet

Use this compact routing block in `AGENTS.md`, `CLAUDE.md`, `GEMINI.md`, or a
client-specific always-loaded instruction file. It solves the common failure
mode where the skill exists but the agent answers from memory before loading it.

```markdown
## Hive MCP Routing

For live crypto intelligence, prefer Hive retrieval over model memory.

- User asks about live crypto data, prices, wallets, tokens, DeFi, NFTs,
  Solana, security, DEX pools, network state, or prediction markets:
  use `hive-query`.
- Exact Hive endpoint, provider, schema, or argument shape is unknown:
  use `hive-tool-discovery`.
- Clear domain request:
  - prices, liquidity, exchanges, OHLC, order books: `hive-market-research`
  - token metadata, holders, liquidity, enrichment, risk:
    `hive-token-diligence`
  - wallet balances, transfers, PnL, NFTs, DeFi positions:
    `hive-wallet-investigation`
  - token, approval, contract, dApp, phishing, simulation risk:
    `hive-security-risk`
  - DEX pools, pairs, trades, liquidity, OHLCV: `hive-dex-pool-analysis`
  - protocol TVL, fees, yields, stablecoins, bridges:
    `hive-defi-research`
  - NFT ownership, metadata, floors, sales, rarity: `hive-nft-research`
  - Solana wallets, SPL accounts, DAS assets, priority fees:
    `hive-solana-analysis`
  - blocks, gas, receipts, logs, RPC diagnostics:
    `hive-network-infrastructure`
  - prediction markets, events, outcomes, traders:
    `hive-prediction-markets`
- User asks Hive to remember, monitor, schedule, alert, or report crypto
  intelligence across sessions: use `hive-stateful-monitoring`.
- Setup, API key, CLI, or app integration:
  use `hive-mcp`, `hive-build-onboarding`, `hive-cli`, or `hive-build`.

Default execution loop:
1. Read the compact `hive://toolsets` index or call `search_tools`.
2. Select one task toolset and its best matching compact `routes[]` entry.
   Preserve `route_id`; follow the ordered calls, conditional per-call
   fallbacks, four-call cap, and stop condition. Use `coverageCatalog` only
   for long-tail discovery.
3. Call `get_api_endpoint_schema` and inspect `operation`/`callPattern`.
4. Execute reads through `invoke_api_endpoint` with bounded args. Explain any
   Hive state change, obtain explicit approval, then use
   `invoke_stateful_endpoint`; never auto-approve it.
5. Copy exact server-returned `_hive` blocks into the task receipt, cite exact
   receipt IDs from `claims[]`, account for every canonical `coverage[]`
   phase, include `route_id`, and call `validate_task_result` before presenting a structured
   result.
6. Report provider, source recency, fallback/cache/degraded state, and runtime
   status (`ok`, `invalid_input`, `missing_key`, `plan_required`,
   `rate_limited`, `degraded`, or `failing`). `observed_at` is Hive
   first-observation/original cache time, not necessarily upstream event time.
```

Keep this snippet short. The skill files carry the detailed workflow and should
be loaded only after the routing layer selects the right Hive skill.