Spiral Stake — mcp read-data backbone (Step 1a)

The composition authority + read API for Spiral Stake strategy data. It ports the app's client-side data composition server-side and serves it back as JSON, so three consumers run off one core:

1. The app — the complete read surface it currently fetches, for a like-for-like Step-2 swap.
2. Agents — the frozen /strategies contract.
3. Partners (later) — a clean, versioned, self-documenting REST surface (OpenAPI).

Read-only. No contract writes, no tx building, no signing, no private keys. It composes numbers and serves JSON. The MCP tool layer, execution/build_*, and partner auth are later steps — the seams are pre-cut (core + http) but not built.

Architecture

src/
  core/        pure composition — no I/O, unit-testable
    leverage.ts        ← VERBATIM copy of v2-client/src/utils/leverage.ts (byte-diff gated)
    apy.ts             collateral-APY resolution (Pendle/DeFiLlama/Royco/StableWatch/on-chain)
    leverageApy.ts     30/60/90d leveraged-APY windows + lumpy-token smoothing
    compose.ts         Market assembly from warm raw (port of FlashLeverage.createInstance)
    strategy.ts        Market → frozen /strategies contract (ladder, freshness, spiralHints)
    exitLiquidity.ts   exit-liquidity tier (the single verdict) + thresholds
  sources/     one adapter per upstream (fetch + parse). native fetch + viem; no axios/wagmi
  cache/       two-layer cache — RAW store (last-good) + warming policy (CONTRACT cadences)
  warmer/      background scheduler (stale-while-revalidate); reads NEVER fetch
  http/        thin hono REST layer (the only thing that knows about requests)
  data/        static market config (copied from the app) + readMarkets port
  config/      env (validated) + structured logger with correlation ids

Data flow: warmer fetches upstreams on the CONTRACT cadence → writes raw to the cache (last-good on failure) → core composes strategies from warm raw only → http serves them. A stale/failed upstream serves last-good with a visible stale age per field-group — never dropped, never coerced to 0.

Run

cp .env.example .env      # fill in RPC + API keys (secrets via env only)
npm install
npm run dev               # tsx watch; serves on :$PORT (default 8787)
npm test                  # golden + parity gates
npm run typecheck         # tsc --noEmit
npm run openapi           # (re)emit openapi.json
node --import tsx scripts/smoke.ts   # live: prime + print one composed strategy

/health (liveness + cache diagnostics) and /ready (503 until the warmer primes the required data) are available immediately on boot.

Endpoints — app-read inventory → endpoint mapping

Every data read the app performs in v2-client/src/api-services/* maps to an endpoint here (positions stay on the dashboard backend and are not reimplemented):

Cross-cutting: /v1 versioning, CORS limited to CORS_ORIGINS, a consistent error envelope ({ error: { code, message, correlationId } }), a correlation id per request (echoed as x-correlation-id), structured JSON logs, and an OpenAPI 3.1 spec at /openapi.json (+ committed openapi.json).

Parity gates

1. leverage.ts golden-vector — test/golden/leverage.golden.test.ts freezes fixed inputs → exact outputs; leverage.parity.test.ts asserts the file is a byte-for-byte copy of the app's leverage.ts. This locks every APY/LTV/ladder number to the app.
2. /strategies composition — test/parity/strategies.compose.test.ts seeds the raw cache with a controlled fixture and asserts the composed contract field-by-field (sourcing, LTV math, ladder via the verbatim leverage.ts, freshness cadences, null-vs-absent, spiralHints isolation). Runs offline in CI.
3. Live golden-set + app-surface parity — scripts/capture-parity.ts primes the warmer against real upstreams and writes the composed /strategies (+ apy/borrow histories) for a fixed market set to test/fixtures/. The parity diff vs the app is produced by feeding the same fixed inputs to the app's client-side composition and comparing (the app logs its composed markets); the PR commits the fixtures and the empty diff. This is the gate that lets us offload TVL safely.

Frozen decisions (parity-critical)

• bignumber.js pinned to the app's 9.3.1; no global BigNumber.config change → toFixed(2) / rounding (ROUND_HALF_UP) identical.
• On-chain reads via viem replace wagmi: getCollateralValueInLoanToken (multicall3), stUSDS str(), spUSDG vsr() (chain 4663). The app's (intentionally lossy) BigInt(liqLtv) from the JSON number is matched exactly, not "fixed".
• maxLeverage = calcLeverage(maxLtv); defaultLeverage = calcLeverage(safeLtv), safeLtv = maxLtv% − 0.75 — exactly as the app derives them. Ladder = integer steps 1x … ⌊maxLeverage⌋ (LTV =(1 − 1/lev)·100) then the exact max.
• Exit slippage is baked into collateralTokens.json by the app's weekly refresh script; its file mtime is the exitLiquidity.asOf. The tier (the only verdict) lives in spiralHints with its thresholds; raw slippagePct stays under exitLiquidity, null (no route) preserved.