@t2000/engine 0.54.2 → 0.56.0

package/README.md

@@ -1,6 +1,6 @@
 # @t2000/engine
 
-Agent engine for conversational finance — implements **Audric Intelligence** (the moat behind the Audric consumer product). Five systems work together: Agent Harness (34 tools — 23 read, 11 write), Reasoning Engine (9 guards across 3 priority tiers + 7 YAML skill recipes), Silent Profile, Chain Memory, and AdviceLog. Every action it triggers waits on Audric Passport's tap-to-confirm.
+Agent engine for conversational finance — implements **Audric Intelligence** (the moat behind the Audric consumer product). Five systems work together: Agent Harness (34 tools — 23 read, 11 write), Reasoning Engine (14 guards across 3 priority tiers + 6 YAML skill recipes), Silent Profile, Chain Memory, and AdviceLog. Every action it triggers waits on Audric Passport's tap-to-confirm.
 
 QueryEngine orchestrates LLM conversations, financial tools, user confirmations, and MCP integrations into a single async-generator loop.
 
@@ -33,6 +33,20 @@ for await (const event of engine.submitMessage('What is my balance?')) {
 }
 ```
 
+## Audric Intelligence — the 5 systems
+
+> _Not a chatbot. A financial agent._ Five systems work together to **understand** the user's money, **reason** about decisions, **act** through 34 financial tools in one conversation, **remember** what they did on-chain, and **remember what it told them**. Every action still waits on Audric Passport's tap-to-confirm.
+
+| System | One-line | Owns | Lives in |
+|---|---|---|---|
+| 🎛️ **Agent Harness** | 34 tools, one agent. | Tool registry, parallel reads, serial writes, permission gates, streaming dispatch | `engine.ts`, `tool.ts`, `orchestration.ts`, `tools/*` |
+| ⚡ **Reasoning Engine** | Thinks before it acts. | Adaptive thinking effort, 14 guards (12 pre-exec + 2 post-exec), 6 YAML skill recipes, prompt caching, preflight validation | `classify-effort.ts`, `guards.ts`, `recipes/registry.ts`, `engine.ts` `cache_control` |
+| 🧠 **Silent Profile** | Knows your finances. | Daily on-chain orientation snapshot + Claude-inferred profile, injected as `<financial_context>` block at every boot | _Audric-side_: `UserFinancialContext` + `UserFinancialProfile` Prisma models + `buildFinancialContextBlock()` |
+| 🔗 **Chain Memory** | Remembers what you do on-chain. | 7 classifiers extract `ChainFact` rows from on-chain history, hydrated as silent context | _Audric-side_: 7 classifier crons + `ChainFact` Prisma model + `buildMemoryContext()` |
+| 📓 **AdviceLog** | Remembers what it told you. | Every recommendation logged (`record_advice` audric tool); last 30 days hydrated each turn so the chat never contradicts itself | _Audric-side_: `AdviceLog` Prisma model + `record_advice` tool + `buildAdviceContext()` |
+
+The engine package owns **Agent Harness** and **Reasoning Engine** in code. The other three systems are powered by audric-side data and injected via the system prompt — see `audric/.cursor/rules/engine-context-assembly.mdc` for the host contract.
+
 ## Architecture
 
 ```
@@ -149,7 +163,18 @@ QueryEngine.submitMessage()
 > fetches. `protocol_deep_dive` retains its DefiLlama dependency (narrow scope, no
 > equivalent on BlockVision). Net: 23 reads + 11 writes = 34 tools.
 
-## Audric 2.0 Engine Features
+## Recent Upgrades — Spec 1 (Correctness) + Spec 2 (Intelligence)
+
+Two upgrades shipped on top of the 5-system base:
+
+| Spec | Versions | What it added |
+|---|---|---|
+| **Spec 1 — Correctness** | v0.41.0 → v0.50.3 | Per-yield `attemptId` (UUID v4) on every `pending_action` — stable join key from action → on-chain receipt → `TurnMetrics` row. `modifiableFields` registry — fields the user can edit on a confirm card without losing the LLM's reasoning (resume route applies `modifications`). `EngineConfig.onAutoExecuted` hook so `auto`-permission writes participate in the same telemetry as confirm-gated ones. |
+| **Spec 2 — Intelligence** | v0.47.0 → v0.54.1 | BlockVision swap — replaced 7 `defillama_*` tools with one `token_prices`; `balance_check` + `portfolio_analysis` rewired to BlockVision Indexer REST. Sticky-positive cache + retry/circuit breaker (`fetchBlockVisionWithRetry`) for graceful 429 handling. `<financial_context>` boot-time orientation injected from the daily `UserFinancialContext` snapshot (Silent Profile). `attemptId`-keyed resume (no clobbering between two pending actions in the same turn). `protocol_deep_dive` retained on DefiLlama as the lone exception. |
+
+> Local-only specs: `AUDRIC_HARNESS_CORRECTNESS_SPEC_v1.3.md`, `AUDRIC_HARNESS_INTELLIGENCE_SPEC_v1.4.1.md`. Cross-repo contracts: `t2000/.cursor/rules/agent-harness-spec.mdc` + `t2000/.cursor/rules/blockvision-resilience.mdc` + `audric/.cursor/rules/audric-transaction-flow.mdc` + `audric/.cursor/rules/write-tool-pending-action.mdc`.
+
+## Engine Features
 
 ### Streaming Tool Execution (Early Dispatch)
 
@@ -170,8 +195,8 @@ Write tool permission resolved dynamically via `resolvePermissionTier(operation,
 ### Reasoning Engine
 
 - **Adaptive thinking** — routes queries to `low`/`medium`/`high` effort based on financial complexity
-- **Guard runner** — 9 guards across 3 priority tiers (Safety > Financial > UX)
-- **Skill recipes** — YAML recipe files with longest-trigger-match-wins
+- **Guard runner** — 14 guards (12 pre-execution + 2 post-execution hints) across 3 priority tiers (Safety > Financial > UX). See `guards.ts` for the full list.
+- **Skill recipes** — 6 YAML recipes (`swap_and_save`, `safe_borrow`, `send_to_contact`, `portfolio_rebalance`, `account_report`, `emergency_withdraw`) with longest-trigger-match-wins
 - **Context compaction** — 200k limit, 85% compact trigger, LLM summarizer fallback
 - **Tool flags** — `mutating`, `requiresBalance`, `affectsHealth`, `irreversible` etc.
 - **Preflight validation** — input validation on `send_transfer`, `swap_execute`, `pay_api`, `borrow`, `save_deposit`

package/dist/index.d.ts

@@ -370,10 +370,28 @@ interface AddressPortfolio {
 }
 /**
  * One-shot wallet portfolio fetcher. BlockVision returns balances + USD
- * prices in a single call; on failure we degrade to a Sui-RPC coin fetch
- * with hardcoded stablecoin pricing. Memoised in-process for `CACHE_TTL_MS`
- * keyed by `address` so back-to-back tool calls inside the same chat
- * session (balance_check + portfolio_analysis) share the same response.
+ * prices in a single call; on failure we degrade to a Sui-RPC coin
+ * fetch with hardcoded stablecoin pricing.
+ *
+ * Caching shape (PR 1+2 — v0.55):
+ *   1. Store-level cache (Redis in prod, in-memory in CLI/tests). Read
+ *      first; if entry is fresh-for-source serve directly.
+ *   2. In-process inflight Map dedupes concurrent in-process callers
+ *      onto one promise.
+ *   3. The leader path runs under a cross-instance lock
+ *      (`bv-lock:wallet:<addr>`) so at most one Vercel instance per
+ *      address is hitting BlockVision at any moment. Followers poll
+ *      the store for the leader's write; if the leader dies they fall
+ *      through to a direct fetch as a defensive degraded path.
+ *
+ * Sticky-positive write rules mirror the DeFi half:
+ *   - `blockvision` (any total) → write unconditionally; latest BV
+ *     truth always wins.
+ *   - `sui-rpc-degraded` → write only when no fresher positive
+ *     `blockvision` entry exists in the sticky window (30 min). When
+ *     a positive sticky entry exists, return it as-is and DO NOT
+ *     overwrite — preserves the original `pricedAt` so a UI can
+ *     render "last refresh Nm ago".
  */
 declare function fetchAddressPortfolio(address: string, apiKey: string | undefined, fallbackRpcUrl?: string): Promise<AddressPortfolio>;
 /**
@@ -414,21 +432,35 @@ interface DefiSummary {
     source: 'blockvision' | 'partial' | 'partial-stale' | 'degraded';
 }
 declare function fetchAddressDefiPortfolio(address: string, apiKey: string | undefined, priceHints?: Record<string, number>): Promise<DefiSummary>;
-declare function clearPortfolioCache(): void;
 /**
- * [v1.4 — Day 2.5] Per-address invalidator. The module-level
- * `portfolioCache` carries a 60s TTL — long enough to outlive a
- * `save_deposit` / `swap_execute` / etc. plus the engine's 1.5s
- * Sui-RPC-indexer-lag delay, so without explicit invalidation
- * `runPostWriteRefresh` would re-fetch the *cached pre-write
- * snapshot*. Engine calls this from `runPostWriteRefresh` right
- * before the lag-delay so the next `fetchAddressPortfolio` for the
- * affected address is forced to hit BlockVision again.
+ * Wipe the wallet portfolio cache plus any in-flight promises.
+ *
+ * Async because the underlying store may be Redis-backed. Existing
+ * fire-and-forget callers (tests, `clearDefiCache`-shape utilities)
+ * still work since they just don't `await` — they get the same
+ * effective "schedule the clear" behavior.
+ */
+declare function clearPortfolioCache(): Promise<void>;
+/**
+ * [v1.4 — Day 2.5] Per-address invalidator.
+ *
+ * Engine `runPostWriteRefresh` calls this right before the 1.5s
+ * Sui-RPC-indexer-lag delay so the next `fetchAddressPortfolio` for
+ * the affected address is forced to hit BlockVision again instead of
+ * returning the cached pre-write snapshot.
+ *
+ * **MUST be awaited** — pre-PR-1 the cache was a synchronous Map and
+ * the `engine.ts` caller fired-and-forgot. Now the underlying store
+ * is async (Upstash in production), so without `await` the next
+ * `balance_check` races the Redis delete and can fetch the stale
+ * pre-write balance — exactly the symptom v0.54 sticky cache shipped
+ * to fix. `engine.ts` `runPostWriteRefresh` was updated to await this
+ * in the same PR.
  *
  * No-op when `address` doesn't match a cached entry — cheap to call
  * unconditionally on every write.
  */
-declare function clearPortfolioCacheFor(address: string): void;
+declare function clearPortfolioCacheFor(address: string): Promise<void>;
 declare function clearPriceMapCache(): void;
 
 /**
@@ -2504,6 +2536,196 @@ declare function getDefiCacheStore(): DefiCacheStore;
 /** Restore the default in-memory store. Used by test teardowns. */
 declare function resetDefiCacheStore(): void;
 
+/** Cache entry stored under each address key. */
+interface WalletCacheEntry {
+    data: AddressPortfolio;
+    /**
+     * Wall-clock ms when this entry was WRITTEN to the cache — used by
+     * the fetcher to compute freshness.
+     *
+     * Distinct from `data.pricedAt` (the upstream-data timestamp from
+     * BlockVision / Sui RPC). Mirrors `DefiCacheEntry.pricedAt` for
+     * cross-pattern consistency.
+     */
+    pricedAt: number;
+}
+/**
+ * Pluggable cache backend.
+ *
+ * All methods are async because Redis-backed implementations are
+ * inherently async; the in-memory impl wraps in resolved promises.
+ *
+ * Implementations MUST tolerate transport errors gracefully — `get`
+ * should return `null` (not throw) on backend failure so the fetcher
+ * falls through to a fresh BlockVision read instead of erroring the
+ * whole `balance_check` call. `set` should swallow errors (logging is
+ * fine) so a Redis hiccup doesn't break a successful read.
+ */
+interface WalletCacheStore {
+    /** Returns the cached entry, or `null` if not found / expired / backend error. */
+    get(address: string): Promise<WalletCacheEntry | null>;
+    /** Writes an entry with a TTL in seconds. Errors are swallowed (logged). */
+    set(address: string, entry: WalletCacheEntry, ttlSec: number): Promise<void>;
+    /** Removes the entry for an address. Errors are swallowed. */
+    delete(address: string): Promise<void>;
+    /** Removes all entries. Used by tests and `clearPortfolioCache()`. */
+    clear(): Promise<void>;
+}
+/**
+ * Process-local cache backed by a `Map`. Used as the default when no
+ * other store has been injected via `setWalletCacheStore()`.
+ *
+ * NOT suitable for multi-instance deployments — each Vercel function
+ * instance gets its own Map, which causes the SSOT divergence this
+ * module exists to fix. Audric replaces this at engine init with an
+ * Upstash-backed store; the CLI and MCP server keep it.
+ *
+ * Address normalization: keys are lowercased to match the
+ * `UpstashWalletCacheStore.key()` convention. Sui addresses are
+ * case-insensitive after the `0x` prefix in practice; keying by
+ * lowercase prevents accidental cache misses from `0xABC...` vs
+ * `0xabc...` callers (e.g. one route normalizes, another doesn't).
+ */
+declare class InMemoryWalletCacheStore implements WalletCacheStore {
+    private readonly store;
+    get(address: string): Promise<WalletCacheEntry | null>;
+    set(address: string, entry: WalletCacheEntry, ttlSec: number): Promise<void>;
+    delete(address: string): Promise<void>;
+    clear(): Promise<void>;
+}
+/**
+ * Swap the active wallet cache store. Call once at engine init from a
+ * runtime that wants a non-default backend (e.g. Audric injecting
+ * `UpstashWalletCacheStore`). Idempotent — calling again replaces the
+ * previous store, but does NOT migrate entries; warm cache is lost on
+ * swap. Tests use this to inject a fake/spy store and
+ * `resetWalletCacheStore()` to restore the in-memory default.
+ */
+declare function setWalletCacheStore(store: WalletCacheStore): void;
+/** Returns the currently active store. Used by `fetchAddressPortfolio`. */
+declare function getWalletCacheStore(): WalletCacheStore;
+/** Restore the default in-memory store. Used by test teardowns. */
+declare function resetWalletCacheStore(): void;
+
+/**
+ * Pluggable distributed mutex. Implementations MUST tolerate transport
+ * errors gracefully — `acquire` returning `false` (not throwing) on
+ * backend failure is fine; the caller will fall through to a direct
+ * fetch. The Upstash impl uses `SET NX EX` which is the canonical
+ * Redis distributed-lock primitive at this scale.
+ */
+interface FetchLock {
+    /**
+     * Try to acquire the lock for `key`, with an automatic expiry of
+     * `leaseSec` seconds.
+     *
+     * Returns `true` if acquired (caller is now the leader for that key,
+     * MUST call `release` when done). Returns `false` if the lock is
+     * already held by someone else (caller is a follower).
+     *
+     * Errors are swallowed (logged); on backend failure return `false`
+     * so the caller falls through to its degraded direct-fetch path.
+     */
+    acquire(key: string, leaseSec: number): Promise<boolean>;
+    /**
+     * Release the lock for `key`. Idempotent — calling on a key the
+     * caller doesn't hold is a no-op (we accept a small window of
+     * potential ABA: if our lease expired and another caller took the
+     * key, we'll harmlessly delete THEIR lock once. Production traffic
+     * patterns make this exceedingly rare; the cost is one extra
+     * fan-out, which is the same cost we'd pay anyway under contention.)
+     *
+     * Errors are swallowed (logged) — release failures are non-fatal
+     * because the lease expires on its own.
+     */
+    release(key: string): Promise<void>;
+}
+/**
+ * Process-local mutex backed by a `Map<key, expiryMs>`. Used as the
+ * default when no other lock has been injected via `setFetchLock()`.
+ *
+ * NOT suitable for multi-instance deployments — each Vercel function
+ * instance gets its own Map, which means N concurrent instances all
+ * acquire successfully and all fan out to BlockVision. Audric replaces
+ * this at engine init with `UpstashFetchLock`; the CLI keeps it
+ * (single-process, so process-local is correct).
+ */
+declare class InMemoryFetchLock implements FetchLock {
+    private readonly held;
+    acquire(key: string, leaseSec: number): Promise<boolean>;
+    release(key: string): Promise<void>;
+}
+/**
+ * Swap the active fetch lock backend. Call once at engine init from a
+ * runtime that wants a non-default backend (e.g. Audric injecting
+ * `UpstashFetchLock`). Idempotent — calling again replaces the previous
+ * lock instance, but does NOT migrate held leases (which are stored in
+ * the backend, not the lock object).
+ */
+declare function setFetchLock(lock: FetchLock): void;
+/** Returns the currently active lock. Used by `awaitOrFetch`. */
+declare function getFetchLock(): FetchLock;
+/** Restore the default in-memory lock. Used by test teardowns. */
+declare function resetFetchLock(): void;
+/** Default lease seconds — sized for worst-case BV retry budget. See header. */
+declare const DEFAULT_LEASE_SEC = 15;
+/** Default follower poll budget ms — must be < engine per-tool timeout. */
+declare const DEFAULT_POLL_BUDGET_MS = 4500;
+/** Default poll interval ms (jittered ±20%) — ~45 GETs over 4.5s. */
+declare const DEFAULT_POLL_INTERVAL_MS = 100;
+interface AwaitOrFetchOpts<T> {
+    /** Override the active lock instance (test seam). */
+    lock?: FetchLock;
+    /** Lease seconds for `lock.acquire`. Default 15. */
+    leaseSec?: number;
+    /** Total ms a follower will poll the cache before falling through. Default 4500. */
+    pollBudgetMs?: number;
+    /** Poll cadence ms (jittered ±20%). Default 100. */
+    pollIntervalMs?: number;
+    /**
+     * Optional cache reader for followers. If provided, followers poll
+     * this every `pollIntervalMs` (jittered) until it returns non-null
+     * or the budget is exhausted. If omitted, followers fall through
+     * immediately to a direct `fetcher()` call (no coalescing benefit).
+     *
+     * The follower's "is this fresh enough?" decision lives inside this
+     * callback — it should return `null` when the cache entry exists but
+     * is too stale to serve, so the poll keeps trying.
+     */
+    pollCache?: () => Promise<T | null>;
+    /** Test seam — defaults to `Math.random()`. */
+    rng?: () => number;
+    /** Test seam — defaults to `setTimeout`-backed promise. */
+    sleep?: (ms: number) => Promise<void>;
+    /** Test seam — defaults to `Date.now()`. */
+    now?: () => number;
+    /** Optional abort signal — caller cancellation halts polling. */
+    signal?: AbortSignal;
+}
+/**
+ * Cross-instance request coalescer.
+ *
+ * `key` MUST be stable across instances for the same logical fetch —
+ * e.g. `bv-lock:wallet:0xabc...` for the wallet portfolio of `0xabc...`.
+ * Different operations on the same address (wallet vs DeFi) MUST use
+ * different keys so they don't block each other.
+ *
+ * `fetcher` is the leader's work. It SHOULD write the cache as its
+ * last act, otherwise followers will time out and fall through to
+ * direct fetches (functionally correct, just no coalescing benefit).
+ *
+ * The leader is guaranteed to call `release()` even if `fetcher()`
+ * throws — propagated to the caller after release completes.
+ *
+ * Followers degrade to direct `fetcher()` on:
+ *   - lock backend failure (`acquire` threw or returned `false` due to a
+ *     transport error rather than contention)
+ *   - poll budget exhausted (leader didn't write cache, or wrote with
+ *     an old enough `pricedAt` that `pollCache` keeps returning `null`)
+ *   - no `pollCache` callback provided
+ */
+declare function awaitOrFetch<T>(key: string, fetcher: () => Promise<T>, opts?: AwaitOrFetchOpts<T>): Promise<T>;
+
 /**
  * Resolve the audric API base URL from the engine's env shim, falling
  * back to `process.env`. Returns `null` when no override is configured —
@@ -2571,6 +2793,6 @@ declare function fetchAudricHistory(address: string, opts: {
     limit?: number;
 }, env?: Record<string, string>, signal?: AbortSignal): Promise<AudricHistoryRecord[] | null>;
 
-declare const DEFAULT_SYSTEM_PROMPT = "You are Audric \u2014 a financial agent on Sui. Audric is exactly five products: Audric Passport (the trust layer \u2014 Google sign-in, non-custodial wallet, tap-to-confirm consent, sponsored gas \u2014 wraps every other product), Audric Intelligence (you \u2014 the 5-system brain: Agent Harness with 34 tools, Reasoning Engine with 9 guards and 7 skill recipes, Silent Profile, Chain Memory, AdviceLog), Audric Finance (manage money on Sui \u2014 Save via NAVI lending at 3-8% APY USDC, Credit via NAVI borrowing with health factor, Swap via Cetus aggregator across 20+ DEXs at 0.1% fee, Charts for yield/health/portfolio viz), Audric Pay (move money \u2014 send USDC, receive via payment links / invoices / QR; free, global, instant on Sui), and Audric Store (creator marketplace, ships Phase 5 \u2014 say \"coming soon\" if asked). Save, swap, borrow, repay, withdraw, charts \u2192 Audric Finance. Send, receive, payment-link, invoice, QR \u2192 Audric Pay. Your silent context (profile, memory, chain facts, advice log) shapes your replies but never surfaces as a notification \u2014 you act only when the user asks, and every write waits on their tap-to-confirm via Passport. You can also call 41 paid APIs (music, image, research, translation, weather, fulfilment) via MPP micropayments using the pay_api tool \u2014 this is an internal capability, not a promoted product, so only mention it when the user asks for something that needs it.\n\n## Response rules\n- 1-2 sentences max. No bullet lists unless asked. No preambles.\n- Never say \"Would you like me to...\", \"Sure!\", \"Great question!\", \"Absolutely!\" \u2014 just do it or say you can't.\n- Present amounts as $1,234.56 and rates as X.XX% APY.\n- Show top 3 results unless asked for more. Summarize totals in one line.\n\n## Caption rules (after tool calls)\n- **When a canvas was rendered (`render_canvas` was called, or any tool that auto-renders a card like balance_check / portfolio_analysis / savings_info / health_check / transaction_history): the canvas IS the answer.** Your chat message must NOT restate wallet, savings, debt, holdings, or net-worth numbers \u2014 they are already on screen. Add at most ONE sentence of context, advice, or next step (e.g. \"Your USDC is idle \u2014 consider depositing for ~4.5% APY\"), or say nothing.\n- **When NO canvas was rendered:** lead with the result and quote the actual numbers from the tool. One sentence.\n- **NEVER describe a position as \"no\", \"none\", \"minimal\", \"zero\", or \"inactive\" if the tool result contains a positive value for that field.** The tool result is the source of truth \u2014 never your interior summary. If the canvas shows $100 in savings, you cannot say \"no active savings\" in the caption.\n- **NEVER claim \"no DeFi positions\" when the tool result says the DeFi slice is UNAVAILABLE.** When `balance_check` displayText contains \"DeFi positions: UNAVAILABLE\" or \"DeFi data source unreachable\", the slice is unknown \u2014 say \"DeFi data is currently unavailable\" or omit the mention. Only claim \"no DeFi positions\" when the displayText explicitly omits any DeFi line (i.e. the fetch succeeded with $0 across every covered protocol).\n\n## Execution rule\nOnly offer to execute actions you have tools for. If you retrieved a quote, data, or information but have no tool to act on it, give the user the result and tell them where to execute manually \u2014 in one sentence. Never say \"Would you like me to proceed?\" unless you have a tool that can actually proceed.\n\n## Before acting\n- ALWAYS call a read tool first before any write tool \u2014 balance_check before save/send/borrow, savings_info before withdraw.\n- Show real numbers from tools \u2014 never fabricate rates, amounts, or balances.\n- When user says \"all\" or an imprecise amount, call the read tool first to get the exact number.\n\n## Tool usage\n- Use tools proactively \u2014 don't refuse requests you can handle.\n- For real-world questions (weather, search, news, prices), use pay_api. Tell the user the cost first.\n- For NAVI lending APYs, use rates_info; for VOLO liquid staking stats, use volo_stats; for spot token prices, use token_prices.\n- For protocol-level due diligence (TVL, fees, audits, safety) on Sui DeFi protocols, use protocol_deep_dive with the slug.\n- Run multiple read-only tools in parallel when you need several data points.\n- If a tool errors, say what went wrong and what to try instead. One sentence.\n\n## Savings = USDC or USDsui (critical)\n- save_deposit and borrow accept ONLY USDC or USDsui. No other token can be deposited or borrowed.\n- USDC is the canonical default. USDsui is permitted because it has a productive NAVI pool (often a higher APY than USDC). All other holdings (GOLD, SUI, USDT, USDe, ETH, NAVX, WAL) are NOT saveable.\n- When asked \"how much can I save?\":\n  - Report saveableUsdc from balance_check (the user's USDC wallet balance \u2014 canonical saveable).\n  - If the user also holds USDsui in their wallet, report that separately as \"USDsui (saveable): X.XX\". Do NOT roll the two together \u2014 the LLM must keep the per-asset distinction so the user can pick.\n- When the user says \"save 10 USDC\" \u2192 call save_deposit with asset=\"USDC\". When they say \"save 10 USDsui\" \u2192 call with asset=\"USDsui\". Never silently substitute.\n- When the user says \"save 10\" (no asset) \u2192 call balance_check first and ask which stable they want, OR pick whichever they hold more of with a one-line explanation.\n- \"Best stable to save right now?\" \u2192 call rates_info to compare USDC vs USDsui APY on NAVI; let the user pick.\n- NEVER say a non-saveable token (GOLD, SUI, USDT, etc.) is \"in savings\" or \"earning APY in savings\". Wallet holdings \u2260 savings positions, even for stables we don't accept.\n- If user wants to save a non-saveable token, tell them to swap to USDC or USDsui first. Do NOT auto-chain swap + deposit.\n- Repay symmetry: a USDsui debt MUST be repaid with USDsui (and USDC debt with USDC). When calling repay_debt, pass asset=\"USDsui\" if the borrow is USDsui. If the user asks \"repay my debt\" and savings_info shows borrows in BOTH stables, list both and ask which to repay first. If the user holds the wrong stable, tell them to swap manually \u2014 do NOT auto-chain swap + repay.\n\n## Multi-step flows\n- \"How much X for Y?\": swap_quote first, then swap_execute if user confirms.\n- \"Swap then save\": swap_execute \u2192 balance_check \u2192 save_deposit. Confirm each step.\n- \"Buy $X of token\": token_prices \u2192 calculate amount \u2192 swap_execute.\n- \"Best yield on SUI\": compare rates_info (NAVI lending) + volo_stats (vSUI liquid staking).\n- withdraw supports legacy positions: USDC, USDe, USDsui, SUI. Pass asset param to withdraw a specific token.\n- \"Deposit SUI to earn yield\": volo_stake for SUI liquid staking. save_deposit only accepts USDC or USDsui.\n- \"Is protocol X safe?\" / \"Tell me about NAVI\": protocol_deep_dive with the slug.\n- \"Full account report\" / \"account summary\" / \"give me everything\" / \"complete overview\": triggers the `account_report` recipe \u2014 when the recipe block appears, follow EVERY step including all six tool calls. Each step renders a distinct rich card; skipping a step means a missing card.\n\n## Safety\n- Never encourage risky financial behavior.\n- Warn when health factor < 1.5.\n- All amounts in USDC unless stated otherwise.";
+declare const DEFAULT_SYSTEM_PROMPT = "You are Audric \u2014 a financial agent on Sui. Audric is exactly five products: Audric Passport (the trust layer \u2014 Google sign-in, non-custodial wallet, tap-to-confirm consent, sponsored gas \u2014 wraps every other product), Audric Intelligence (you \u2014 the 5-system brain: Agent Harness with 34 tools, Reasoning Engine with 14 guards and 6 skill recipes, Silent Profile, Chain Memory, AdviceLog), Audric Finance (manage money on Sui \u2014 Save via NAVI lending at 3-8% APY USDC, Credit via NAVI borrowing with health factor, Swap via Cetus aggregator across 20+ DEXs at 0.1% fee, Charts for yield/health/portfolio viz), Audric Pay (move money \u2014 send USDC, receive via payment links / invoices / QR; free, global, instant on Sui), and Audric Store (creator marketplace, ships Phase 5 \u2014 say \"coming soon\" if asked). Save, swap, borrow, repay, withdraw, charts \u2192 Audric Finance. Send, receive, payment-link, invoice, QR \u2192 Audric Pay. Your silent context (profile, memory, chain facts, advice log) shapes your replies but never surfaces as a notification \u2014 you act only when the user asks, and every write waits on their tap-to-confirm via Passport. You can also call 41 paid APIs (music, image, research, translation, weather, fulfilment) via MPP micropayments using the pay_api tool \u2014 this is an internal capability, not a promoted product, so only mention it when the user asks for something that needs it.\n\n## Response rules\n- 1-2 sentences max. No bullet lists unless asked. No preambles.\n- Never say \"Would you like me to...\", \"Sure!\", \"Great question!\", \"Absolutely!\" \u2014 just do it or say you can't.\n- Present amounts as $1,234.56 and rates as X.XX% APY.\n- Show top 3 results unless asked for more. Summarize totals in one line.\n\n## Caption rules (after tool calls)\n- **When a canvas was rendered (`render_canvas` was called, or any tool that auto-renders a card like balance_check / portfolio_analysis / savings_info / health_check / transaction_history): the canvas IS the answer.** Your chat message must NOT restate wallet, savings, debt, holdings, or net-worth numbers \u2014 they are already on screen. Add at most ONE sentence of context, advice, or next step (e.g. \"Your USDC is idle \u2014 consider depositing for ~4.5% APY\"), or say nothing.\n- **When NO canvas was rendered:** lead with the result and quote the actual numbers from the tool. One sentence.\n- **NEVER describe a position as \"no\", \"none\", \"minimal\", \"zero\", or \"inactive\" if the tool result contains a positive value for that field.** The tool result is the source of truth \u2014 never your interior summary. If the canvas shows $100 in savings, you cannot say \"no active savings\" in the caption.\n- **NEVER claim \"no DeFi positions\" when the tool result says the DeFi slice is UNAVAILABLE.** When `balance_check` displayText contains \"DeFi positions: UNAVAILABLE\" or \"DeFi data source unreachable\", the slice is unknown \u2014 say \"DeFi data is currently unavailable\" or omit the mention. Only claim \"no DeFi positions\" when the displayText explicitly omits any DeFi line (i.e. the fetch succeeded with $0 across every covered protocol).\n\n## Execution rule\nOnly offer to execute actions you have tools for. If you retrieved a quote, data, or information but have no tool to act on it, give the user the result and tell them where to execute manually \u2014 in one sentence. Never say \"Would you like me to proceed?\" unless you have a tool that can actually proceed.\n\n## Before acting\n- ALWAYS call a read tool first before any write tool \u2014 balance_check before save/send/borrow, savings_info before withdraw.\n- Show real numbers from tools \u2014 never fabricate rates, amounts, or balances.\n- When user says \"all\" or an imprecise amount, call the read tool first to get the exact number.\n\n## Tool usage\n- Use tools proactively \u2014 don't refuse requests you can handle.\n- For real-world questions (weather, search, news, prices), use pay_api. Tell the user the cost first.\n- For NAVI lending APYs, use rates_info; for VOLO liquid staking stats, use volo_stats; for spot token prices, use token_prices.\n- For protocol-level due diligence (TVL, fees, audits, safety) on Sui DeFi protocols, use protocol_deep_dive with the slug.\n- Run multiple read-only tools in parallel when you need several data points.\n- If a tool errors, say what went wrong and what to try instead. One sentence.\n\n## Savings = USDC or USDsui (critical)\n- save_deposit and borrow accept ONLY USDC or USDsui. No other token can be deposited or borrowed.\n- USDC is the canonical default. USDsui is permitted because it has a productive NAVI pool (often a higher APY than USDC). All other holdings (GOLD, SUI, USDT, USDe, ETH, NAVX, WAL) are NOT saveable.\n- When asked \"how much can I save?\":\n  - Report saveableUsdc from balance_check (the user's USDC wallet balance \u2014 canonical saveable).\n  - If the user also holds USDsui in their wallet, report that separately as \"USDsui (saveable): X.XX\". Do NOT roll the two together \u2014 the LLM must keep the per-asset distinction so the user can pick.\n- When the user says \"save 10 USDC\" \u2192 call save_deposit with asset=\"USDC\". When they say \"save 10 USDsui\" \u2192 call with asset=\"USDsui\". Never silently substitute.\n- When the user says \"save 10\" (no asset) \u2192 call balance_check first and ask which stable they want, OR pick whichever they hold more of with a one-line explanation.\n- \"Best stable to save right now?\" \u2192 call rates_info to compare USDC vs USDsui APY on NAVI; let the user pick.\n- NEVER say a non-saveable token (GOLD, SUI, USDT, etc.) is \"in savings\" or \"earning APY in savings\". Wallet holdings \u2260 savings positions, even for stables we don't accept.\n- If user wants to save a non-saveable token, tell them to swap to USDC or USDsui first. Do NOT auto-chain swap + deposit.\n- Repay symmetry: a USDsui debt MUST be repaid with USDsui (and USDC debt with USDC). When calling repay_debt, pass asset=\"USDsui\" if the borrow is USDsui. If the user asks \"repay my debt\" and savings_info shows borrows in BOTH stables, list both and ask which to repay first. If the user holds the wrong stable, tell them to swap manually \u2014 do NOT auto-chain swap + repay.\n\n## Multi-step flows\n- \"How much X for Y?\": swap_quote first, then swap_execute if user confirms.\n- \"Swap then save\": swap_execute \u2192 balance_check \u2192 save_deposit. Confirm each step.\n- \"Buy $X of token\": token_prices \u2192 calculate amount \u2192 swap_execute.\n- \"Best yield on SUI\": compare rates_info (NAVI lending) + volo_stats (vSUI liquid staking).\n- withdraw supports legacy positions: USDC, USDe, USDsui, SUI. Pass asset param to withdraw a specific token.\n- \"Deposit SUI to earn yield\": volo_stake for SUI liquid staking. save_deposit only accepts USDC or USDsui.\n- \"Is protocol X safe?\" / \"Tell me about NAVI\": protocol_deep_dive with the slug.\n- \"Full account report\" / \"account summary\" / \"give me everything\" / \"complete overview\": triggers the `account_report` recipe \u2014 when the recipe block appears, follow EVERY step including all six tool calls. Each step renders a distinct rich card; skipping a step means a missing card.\n\n## Safety\n- Never encourage risky financial behavior.\n- Warn when health factor < 1.5.\n- All amounts in USDC unless stated otherwise.";
 
-export { type AddressPortfolio, AnthropicProvider, type AnthropicProviderConfig, type AudricHistoryRecord, type AudricPortfolioResult, type BalancePrices, type BalanceResult, BalanceTracker, type BuildToolOptions, CANVAS_TEMPLATES, type CanvasTemplate, type ChatParams, type CompactOptions, type ContentBlock, ContextBudget, type ContextBudgetConfig, type ConversationState, type ConversationStateStore, type CostSnapshot, CostTracker, type CostTrackerConfig, DEFAULT_GUARD_CONFIG, DEFAULT_PERMISSION_CONFIG, DEFAULT_SYSTEM_PROMPT, type DefiCacheEntry, type DefiCacheStore, type DefiProtocol, type DefiSummary, EarlyToolDispatcher, type EngineConfig, type EngineEvent, type GuardCheckResult, type GuardConfig, type GuardEvent, type GuardInjection, type GuardResult, type GuardRunnerState, type GuardTier, type GuardVerdict, type HealthFactorResult, InMemoryDefiCacheStore, type LLMProvider, type McpCallResult, McpClientManager, McpResponseCache, type McpServerConfig, type McpServerConnection, type McpToolAdapterConfig, type McpToolDescriptor, MemorySessionStore, type Message, NAVI_MCP_CONFIG, NAVI_MCP_URL, NAVI_SERVER_NAME, type NaviRawCoin, type NaviRawHealthFactor, type NaviRawPool, type NaviRawPosition, type NaviRawPositionsResponse, type NaviRawProtocolStats, type NaviRawRewardsResponse, type NaviReadOptions, NaviTools, type OutputConfig, PERMISSION_PRESETS, type PendingAction, type PendingActionModifiableField, type PendingReward, type PendingToolCall, type PermissionLevel, type PermissionOperation, type PermissionResponse, type PermissionRule, type PortfolioCoin, type PositionEntry, type PreflightResult, type ProtocolStats, type ProviderEvent, QueryEngine, READ_TOOLS, type RatesResult, type Recipe, type RecipePrerequisite, RecipeRegistry, type RecipeStep, type RecipeStepOnError, RetryTracker, type SSEEvent, type SavingsResult, type ServerPositionData, type SessionData, type SessionStore, type StateType, type StopReason, type SuiCoinBalance, type SystemBlock, type SystemPrompt, TOOL_FLAGS, TOOL_MODIFIABLE_FIELDS, type ThinkingConfig, type ThinkingEffort, type Tool, type ToolChoice, type ToolContext, type ToolDefinition, type ToolFlags, type ToolJsonSchema, type ToolResult, TxMutex, type UserFinancialProfile, type UserPermissionConfig, WRITE_TOOLS, type WalletCoin, activitySummaryTool, adaptAllMcpTools, adaptAllServerTools, adaptMcpTool, applyToolFlags, balanceCheckTool, borrowTool, budgetToolResult, buildCachedSystemPrompt, buildMcpTools, buildProactivenessInstructions, buildProfileContext, buildSelfEvaluationInstruction, buildStateContext, buildTool, claimRewardsTool, classifyEffort, clearPortfolioCache, clearPortfolioCacheFor, clearPriceMapCache, compactMessages, createGuardRunnerState, engineToSSE, estimateTokens, explainTxTool, extractConversationText, extractMcpText, fetchAddressDefiPortfolio, fetchAddressPortfolio, fetchAudricHistory, fetchAudricPortfolio, fetchAvailableRewards, fetchBalance, fetchHealthFactor, fetchPositions, fetchProtocolStats, fetchRates, fetchSavings, fetchTokenPrices, fetchWalletCoins, findTool, getAudricApiBase, getDefaultTools, getDefiCacheStore, getMcpManager, getModifiableFields, getToolFlags, getWalletAddress, guardArtifactPreview, guardStaleData, hasNaviMcp, healthCheckTool, loadRecipes, microcompact, mppServicesTool, parseMcpJson, parseRecipe, parseSSE, payApiTool, portfolioAnalysisTool, protocolDeepDiveTool, ratesInfoTool, registerEngineTools, renderCanvasTool, repayDebtTool, requireAgent, resetDefiCacheStore, resolvePermissionTier, resolveUsdValue, runGuards, runTools, saveContactTool, saveDepositTool, savingsInfoTool, sendTransferTool, serializeSSE, setDefiCacheStore, spendingAnalyticsTool, swapExecuteTool, swapQuoteTool, tokenPricesTool, toolNameToOperation, toolsToDefinitions, transactionHistoryTool, transformBalance, transformHealthFactor, transformPositions, transformRates, transformRewards, transformSavings, updateGuardStateAfterToolResult, validateHistory, voloStakeTool, voloStatsTool, voloUnstakeTool, webSearchTool, withdrawTool, yieldSummaryTool };
+export { type AddressPortfolio, AnthropicProvider, type AnthropicProviderConfig, type AudricHistoryRecord, type AudricPortfolioResult, type AwaitOrFetchOpts, type BalancePrices, type BalanceResult, BalanceTracker, type BuildToolOptions, CANVAS_TEMPLATES, type CanvasTemplate, type ChatParams, type CompactOptions, type ContentBlock, ContextBudget, type ContextBudgetConfig, type ConversationState, type ConversationStateStore, type CostSnapshot, CostTracker, type CostTrackerConfig, DEFAULT_GUARD_CONFIG, DEFAULT_LEASE_SEC, DEFAULT_PERMISSION_CONFIG, DEFAULT_POLL_BUDGET_MS, DEFAULT_POLL_INTERVAL_MS, DEFAULT_SYSTEM_PROMPT, type DefiCacheEntry, type DefiCacheStore, type DefiProtocol, type DefiSummary, EarlyToolDispatcher, type EngineConfig, type EngineEvent, type FetchLock, type GuardCheckResult, type GuardConfig, type GuardEvent, type GuardInjection, type GuardResult, type GuardRunnerState, type GuardTier, type GuardVerdict, type HealthFactorResult, InMemoryDefiCacheStore, InMemoryFetchLock, InMemoryWalletCacheStore, type LLMProvider, type McpCallResult, McpClientManager, McpResponseCache, type McpServerConfig, type McpServerConnection, type McpToolAdapterConfig, type McpToolDescriptor, MemorySessionStore, type Message, NAVI_MCP_CONFIG, NAVI_MCP_URL, NAVI_SERVER_NAME, type NaviRawCoin, type NaviRawHealthFactor, type NaviRawPool, type NaviRawPosition, type NaviRawPositionsResponse, type NaviRawProtocolStats, type NaviRawRewardsResponse, type NaviReadOptions, NaviTools, type OutputConfig, PERMISSION_PRESETS, type PendingAction, type PendingActionModifiableField, type PendingReward, type PendingToolCall, type PermissionLevel, type PermissionOperation, type PermissionResponse, type PermissionRule, type PortfolioCoin, type PositionEntry, type PreflightResult, type ProtocolStats, type ProviderEvent, QueryEngine, READ_TOOLS, type RatesResult, type Recipe, type RecipePrerequisite, RecipeRegistry, type RecipeStep, type RecipeStepOnError, RetryTracker, type SSEEvent, type SavingsResult, type ServerPositionData, type SessionData, type SessionStore, type StateType, type StopReason, type SuiCoinBalance, type SystemBlock, type SystemPrompt, TOOL_FLAGS, TOOL_MODIFIABLE_FIELDS, type ThinkingConfig, type ThinkingEffort, type Tool, type ToolChoice, type ToolContext, type ToolDefinition, type ToolFlags, type ToolJsonSchema, type ToolResult, TxMutex, type UserFinancialProfile, type UserPermissionConfig, WRITE_TOOLS, type WalletCacheEntry, type WalletCacheStore, type WalletCoin, activitySummaryTool, adaptAllMcpTools, adaptAllServerTools, adaptMcpTool, applyToolFlags, awaitOrFetch, balanceCheckTool, borrowTool, budgetToolResult, buildCachedSystemPrompt, buildMcpTools, buildProactivenessInstructions, buildProfileContext, buildSelfEvaluationInstruction, buildStateContext, buildTool, claimRewardsTool, classifyEffort, clearPortfolioCache, clearPortfolioCacheFor, clearPriceMapCache, compactMessages, createGuardRunnerState, engineToSSE, estimateTokens, explainTxTool, extractConversationText, extractMcpText, fetchAddressDefiPortfolio, fetchAddressPortfolio, fetchAudricHistory, fetchAudricPortfolio, fetchAvailableRewards, fetchBalance, fetchHealthFactor, fetchPositions, fetchProtocolStats, fetchRates, fetchSavings, fetchTokenPrices, fetchWalletCoins, findTool, getAudricApiBase, getDefaultTools, getDefiCacheStore, getFetchLock, getMcpManager, getModifiableFields, getToolFlags, getWalletAddress, getWalletCacheStore, guardArtifactPreview, guardStaleData, hasNaviMcp, healthCheckTool, loadRecipes, microcompact, mppServicesTool, parseMcpJson, parseRecipe, parseSSE, payApiTool, portfolioAnalysisTool, protocolDeepDiveTool, ratesInfoTool, registerEngineTools, renderCanvasTool, repayDebtTool, requireAgent, resetDefiCacheStore, resetFetchLock, resetWalletCacheStore, resolvePermissionTier, resolveUsdValue, runGuards, runTools, saveContactTool, saveDepositTool, savingsInfoTool, sendTransferTool, serializeSSE, setDefiCacheStore, setFetchLock, setWalletCacheStore, spendingAnalyticsTool, swapExecuteTool, swapQuoteTool, tokenPricesTool, toolNameToOperation, toolsToDefinitions, transactionHistoryTool, transformBalance, transformHealthFactor, transformPositions, transformRates, transformRewards, transformSavings, updateGuardStateAfterToolResult, validateHistory, voloStakeTool, voloStatsTool, voloUnstakeTool, webSearchTool, withdrawTool, yieldSummaryTool };