franalgaba__grimoire/AGENTS.md

description, globs, alwaysApply

description	globs	alwaysApply
Grimoire - A Portable Execution Language for Onchain Strategies	*.ts, *.tsx, *.html, *.css, *.js, *.jsx, *.spell, package.json	false

Development Rules

First Message

If the user did not give a concrete task in their first message, read README.md, then ask which area to work on (compiler/runtime, venues, CLI, docs, skills). Based on the answer, read the relevant docs or package readmes.

Agent Portability (LLM-agnostic)

• Avoid tool-specific assumptions. Use standard shell commands and Bun scripts.
• Prefer repo-local context. Document workflows in this repo instead of relying on conversation memory.
• Use relative paths.
• Provide manual fallbacks if a workflow is automated.
• Keep secrets out of prompts; use environment variables.
• If a task needs new dependencies, list them and pin versions.

Code Quality

• No any unless absolutely necessary.
• Avoid Bun-only APIs in packages/core, packages/venues, and packages/cli. If you must use Bun APIs, add a Node fallback.
• Keep protocol SDKs out of @grimoirelabs/core (core stays protocol-agnostic).
• Prefer deterministic, side-effect-free logic in compiler/runtime.
• Ask before removing functionality or changing DSL semantics.

Commands

• For code changes, run bun run validate unless the user says otherwise.
• Docs-only changes do not require tests.
• Onchain tests are expensive: only run when requested. Use --dry-run first.
• Never commit unless the user asks.

Git Rules

• Do not use destructive commands (git reset --hard, git checkout ., git clean -fd) unless explicitly asked.
• Stage only files you changed in this session.

GitHub Issues

When reading issues, read all comments. Use:

gh issue view <number> --json title,body,comments,labels,state

PR Workflow

Work in a feature branch if requested. Do not open PRs unless asked.

Tools

• Bun for install/test/build: bun install, bun test, bun run ...
• Biome for lint/format: bun run lint, bun run format
• TypeScript: bun run typecheck
• Skills validation: bunx skills-ref validate <skill-dir> (for example skills/grimoire, skills/grimoire-aave)
• Changesets: bunx changeset

Style

• Keep answers concise and technical.
• No emojis in commits, issues, or code comments.

Changelog

Location: CHANGELOG.md (root). Format follows Keep a Changelog.

• Add new entries only under ## Unreleased.
• Do not edit already-released version sections.
• Use Changesets for normal release notes. Only edit CHANGELOG.md when asked.

Releasing

• Initial 0.1.0 publish is manual (see docs/how-to/publish.md).
• After that, use Changesets + CI. Do not bump versions manually.

Project Overview

Grimoire is a DSL for defining and executing onchain DeFi strategies. Strategies are written in .spell files using a brace-delimited syntax ({ ... }).

Project Structure

packages/
  core/                      # Core compiler + runtime (protocol-agnostic)
    src/
      compiler/              # Spell -> IR compilation
        grimoire/            # Tokenizer, parser, transformer
        expression-parser.ts
        ir-generator.ts
        validator.ts
      runtime/               # IR execution engine
        interpreter.ts
        context.ts
        state-store.ts       # StateStore interface + RunRecord types
        sqlite-state-store.ts # SQLite-based state persistence
        steps/               # Step handlers (compute, conditional, loop, etc.)
      venues/                # Adapter registry + types (no SDKs)
      types/                 # TypeScript type definitions
      builders/              # Fluent API for building spells
  venues/                    # Official adapters (SDK integrations)
    src/
      adapters/              # Protocol-specific adapter implementations
        aave-v3.ts
        across.ts
        hyperliquid.ts
        morpho-blue/         # Multi-file adapter (adapter, markets, preflight)
        pendle/              # Multi-file adapter (adapter, api, convert, helpers)
        polymarket/          # Multi-file adapter (adapter, client, normalize)
        uniswap-v3.ts
        uniswap-v4/          # Multi-file adapter (adapter, encoding, constants, permit2)
      shared/                # Cross-adapter utilities (bigint, bps, gas, erc20, constraints, token-registry)
      cli/                   # per-venue CLIs
  cli/                       # Grimoire CLI
  sdk/                       # (WIP) SDK for external integrations
spells/                      # Example spell files
skills/                      # Agent skills (core CLI + per-venue metadata workflows)
docs/                        # Diataxis docs

Grimoire Syntax

Spells use a brace-delimited syntax ({ ... }):

spell YieldOptimizer {
  version: "1.0.0"
  description: "Optimizes yield across lending venues"
  assets: [USDC, USDT, DAI]

  params: {
    min_amount: 100
    amount: 100000
  }

  limits: {
    max_allocation_per_venue: 50%
  }

  venues: {
    aave_v3: @aave_v3
    morpho_blue: @morpho_blue
    uniswap_v3: @uniswap_v3
  }

  skills: {
    dex: {
      type: swap
      adapters: [uniswap_v3]
      default_constraints: {
        max_slippage: 50
      }
    }
  }

  advisors: {
    risk: {
      model: anthropic:sonnet
      timeout: 30
      fallback: true
    }
  }

  state: {
    persistent: {
      counter: 0
    }
    ephemeral: {
      temp: 0
    }
  }

  on hourly: {
    for asset in assets {
      current_balance = balance(asset)

      if current_balance > params.min_amount {
        if **gas costs justify the move** via risk {
          atomic {
            aave_v3.withdraw(asset, params.amount)
            morpho_blue.lend(asset, params.amount)
          }
        }
      }
    }
  }
}

Syntax Reference

Feature	Syntax	Example
Spell declaration	spell Name {	spell YieldOptimizer {
Arrays	[item1, item2]	assets: [USDC, DAI]
Venue refs	@name	@aave_v3
Venue groups	name: [@v1, @v2]	lending: [@aave_v3, @morpho_blue]
Skills	skills: {	skills: { ... }
Advisors	advisors: {	advisors: { ... }
Percentages	N%	50% (converts to 0.5)
Triggers	on trigger: {	on hourly: {, on daily: {, on manual: {
For loops	for x in y {	for asset in assets {
Repeat loops	repeat N {	repeat 3 {
Loop until	loop until cond max N {	loop until done max 10 {
If/elif/else	if cond {	if x > 0 {
Advisory (AI)	advise advisor: "prompt"	decision = advise risk: "is this safe" { output: { type: boolean }, fallback: true }
Advise	x = advise advisor:	decision = advise risk: "..."
Atomic blocks	atomic {	Transaction grouping
Try/catch	try {	try { ... } catch *: {
Parallel	parallel ... {	parallel join=all: {
Pipeline	`expr	map: {`
Block/Do	block / do	block add(a,b) {
Comments	# comment	# Calculate rates
Method calls	obj.method(args)	venue.deposit(asset, amount)
Assignment	x = expr	rate = apy(aave_v3, asset)
Using skill	using name	swap(...) using dex
Constraints	with k=v or with (k=v, ...)	with max_slippage=50 or multi-line with (\n slippage=50,\n deadline=300,\n)
Logical ops	and, or, not	if a > 0 and b < 10 {
Emit events	emit name(k=v)	emit done(value=42)
Halt execution	halt "reason"	halt "insufficient balance"
Wait	wait N	wait 3600 (seconds)
Trailing commas	allowed in all lists	[USDC, ETH,], f(a, b,), with (k=v,)
Multi-line objects	{ key: val } across lines	{ / a: 1, / b: 2, / }

Example Spells

• spells/uniswap-swap-execute.spell
• spells/aave-supply.spell
• spells/morpho-blue-lend.spell
• spells/across-bridge.spell
• spells/hyperliquid-perps.spell

Compiler Pipeline

Source (.spell) -> Tokenizer -> Parser -> AST -> Transformer -> SpellSource -> IR Generator -> SpellIR -> Type Checker -> Validator

Key files:

• packages/core/src/compiler/grimoire/tokenizer.ts - Brace-delimited lexer, emits LBRACE/RBRACE tokens
• packages/core/src/compiler/grimoire/parser.ts - Recursive descent parser
• packages/core/src/compiler/grimoire/ast.ts - AST node type definitions
• packages/core/src/compiler/grimoire/transformer.ts - AST -> SpellSource conversion
• packages/core/src/compiler/ir-generator.ts - SpellSource -> SpellIR (executable format)
• packages/core/src/compiler/type-checker.ts - Compile-time type checker (errors block compilation)
• packages/core/src/compiler/validator.ts - IR validation (structural checks)

Execution Model (Preview / Commit)

Grimoire uses an embedded runtime with a mandatory preview / commit model for any action that moves value (onchain or offchain). The runtime is the same whether invoked from the CLI or embedded in an agent.

1. Preview -- Full execution against simulated or forked state. Produces a typed receipt listing every value-moving action, its parameters, and expected outcomes.
2. Commit -- Settlement using an approved receipt. The runtime re-checks drift bounds before submitting transactions. If bounds are violated, commit rejects and requires a fresh preview.

Local computation, state writes, and advisory calls do not cross the irreversibility boundary and therefore do not require the preview/commit cycle.

This model is defined by the Value-Flow Mandate Runtime. The same preview/commit semantics apply to both CLI and embedded runtime usage.

Venues and Adapters

Core is SDK-free. Protocol integrations live in @grimoirelabs/venues and are injected at execution time.

Supported adapters:

• aave_v3 (AaveKit) - lending/borrowing on Ethereum + Base
• uniswap_v3 - swaps via SwapRouter02
• uniswap_v4 - swaps via Universal Router + Permit2
• morpho_blue - isolated lending markets
• hyperliquid (offchain) - spot + perps via API
• across (bridge) - cross-chain bridging
• pendle - hosted SDK convert routes
• polymarket (offchain) - prediction market CLOB execution on Polygon

Adapters can return multi-transaction plans to handle approvals. Offchain venues implement executeAction.

Aave V3 Amount Format

The @aave/client SDK uses human-readable BigDecimal amounts, not raw token units:

• supply/borrow: value: "0.1" (human-readable)
• withdraw/repay: value: { exact: "100000" } (raw units in exact wrapper)

The adapter handles this conversion internally using token decimals (USDC=6, WETH=18, etc.).

Morpho Blue Default Markets

The default adapter ships with well-known Base (chain 8453) markets:

• cbBTC/USDC (86% LLTV)
• WETH/USDC (86% LLTV)

When no collateral is specified in a spell, the first matching market by loan token is selected. To target a specific market, specify the collateral token in the action.

Across Bridge Minimums

Across enforces minimum bridge amounts per token. For test spells:

• USDC: minimum about 1.00 (1000000 raw)
• WETH: minimum about 0.002 ETH

State Persistence

Spell state survives across runs via a SQLite-backed store. The execute() function stays pure; persistence is handled by the CLI or caller.

Architecture

CLI:  state = await store.load(spellId)
CLI:  result = await execute({ ..., persistentState: state })
CLI:  await store.save(spellId, result.finalState)
CLI:  await store.addRun(spellId, createRunRecord(result))
CLI:  await store.saveLedger(spellId, result.runId, result.ledgerEvents)

Key files

• packages/core/src/runtime/state-store.ts - StateStore interface, RunRecord type, createRunRecord() helper
• packages/core/src/runtime/sqlite-state-store.ts - SqliteStateStore class using bun:sqlite with a Node fallback
• packages/cli/src/commands/state-helpers.ts - withStatePersistence() wrapper for CLI commands

StateStore interface

interface StateStore {
  load(spellId: string): Promise<Record<string, unknown> | null>;
  save(spellId: string, state: Record<string, unknown>): Promise<void>;
  addRun(spellId: string, run: RunRecord): Promise<void>;
  getRunById(runId: string): Promise<RunRecord | null>;
  getRuns(spellId: string, limit?: number): Promise<RunRecord[]>;
  saveLedger(spellId: string, runId: string, entries: LedgerEntry[]): Promise<void>;
  loadLedger(spellId: string, runId: string): Promise<LedgerEntry[] | null>;
  listSpells(): Promise<string[]>;
  upsertRunTrack(track: RunTrackRecord): Promise<void>;
  getRunTracks(runId: string): Promise<RunTrackRecord[]>;
  upsertRunHandoff(handoff: RunHandoffRecord): Promise<void>;
  getRunHandoffs(runId: string): Promise<RunHandoffRecord[]>;
  upsertRunStepResult(step: RunStepResultRecord): Promise<void>;
  getRunStepResults(runId: string): Promise<RunStepResultRecord[]>;
}

SQLite schema

Database lives at .grimoire/grimoire.db (configurable via --state-dir). Core tables: spell_state (persistent state), runs (execution history), ledger (event logs per run). Cross-chain tables: run_tracks, run_handoffs, run_step_results. Migrations are additive and versioned via PRAGMA user_version. Old runs are pruned beyond maxRuns (default 100).

CLI flags

• --state-dir <dir> - custom directory for grimoire.db (on simulate, cast, history, log)
• --no-state - disable persistence entirely (on simulate, cast)
• --advisory-trace-verbose - stream detailed advisory trace (prompt/schema, tool payloads, model text/thinking deltas) in non-JSON simulate/cast runs
• --destination-spell <spell> - enable cross-chain orchestration by specifying destination spell
• --destination-chain <id> - destination chain ID for cross-chain mode
• --handoff-timeout-sec <seconds> - required handoff timeout in cross-chain mode
• --poll-interval-sec <seconds> - handoff polling interval (default 30)
• --watch - keep process alive to continue after handoff settlement
• --morpho-market-id <actionRef>=<marketId> / --morpho-market-map <path> - explicit Morpho market identity mapping for cross-chain runs

Action Constraints (Slippage)

Action constraints are resolved at runtime and attached to Action.constraints for adapter use.

• max_slippage, min_output, max_input, deadline
• max_price_impact, min_liquidity, require_quote, require_simulation, max_gas

For swaps, always set both max_slippage and min_output to prevent unexpected losses.

Bridge Actions

bridge actions compile from venue.bridge(asset, amount, to_chain) and are supported by the Across adapter. The IR requires a literal chain ID for to_chain.

CLI

• grimoire format <paths...> [--write|--check] [--diff] [--json] - canonicalize .spell source or check formatting
• grimoire compile <spell> - compile a .spell file to IR
• grimoire compile-all [dir] - compile all .spell files in a directory
• grimoire validate <spell> [--strict] [--json] - validate a .spell file
• grimoire simulate <spell> [--rpc-url <url>|<chainId>=<url>] [--destination-spell <spell>] - simulate execution (dry run), with state persistence
• grimoire cast <spell> [--rpc-url <url>|<chainId>=<url>] [--destination-spell <spell>] - execute a spell onchain, with state persistence
• grimoire setup [--chain <id>] [--rpc-url <url>] [--adapter <name>] [--keystore <path>] [--password-env <name>] [--key-env <name>] [--import-key] [--no-save-password-env] [--no-doctor] [--non-interactive] [--json] - guided local execute onboarding (wallet + RPC + readiness checks, with optional .grimoire/setup.env writing/autoload)
• grimoire resume <runId> [--watch] - resume a waiting cross-chain orchestration run
• grimoire history [spell] - view execution history (all spells or runs for one spell)
• grimoire log <spell> <runId> - view ledger events for a specific run
• grimoire venues - list available venue adapters
• grimoire venue doctor [--chain <id>] [--adapter <name>] [--rpc-url <url>] [--json] - venue env/chain/RPC diagnostics
• grimoire init - initialize a new .grimoire directory
• Per-venue CLIs in @grimoirelabs/venues:
  • grimoire-aave
  • grimoire-uniswap
  • grimoire-morpho-blue
  • grimoire-across
  • grimoire-hyperliquid
  • grimoire-pendle
  • grimoire-polymarket

Documentation and Skills Maintenance

Docs live in docs/ and follow the Diataxis structure.

Skills live in skills/ and provide LLM-consumable context:

• skills/grimoire/ - Core CLI commands
• skills/grimoire-aave/ - Aave V3 venue CLI + amount format
• skills/grimoire-uniswap/ - Uniswap V3/V4 venue CLI
• skills/grimoire-morpho-blue/ - Morpho Blue venue CLI + default markets
• skills/grimoire-across/ - Across Protocol bridge quotes, routes, and deposit status
• skills/grimoire-hyperliquid/ - Hyperliquid venue CLI
• skills/grimoire-pendle/ - Pendle Hosted SDK metadata workflows
• skills/grimoire-polymarket/ - Polymarket CLOB custom action workflows

Keep docs and skills in sync with code changes:

Change	Update
New/changed CLI command or flag	docs/reference/cli.md, skills/grimoire/SKILL.md, this file
New/changed DSL syntax or feature	docs/reference/grimoire-dsl-spec.md, docs/reference/spell-syntax.md
New/changed venue adapter	docs/reference/venues.md, matching skills/grimoire-<venue>/SKILL.md
New venue adapter added	Create skills/grimoire-<venue>/SKILL.md, update docs/reference/venues.md
Test runner changes	docs/how-to/run-tests.md
Wallet/keystore changes	docs/reference/cli.md (setup + wallet sections), docs/how-to/use-wallet-commands-end-to-end.md
Bridge/amount thresholds	docs/how-to/bridge-with-across.md, docs/reference/venues.md
State persistence changes	This file (State Persistence section)
New example spells	README.md (Examples section), docs/reference/grimoire-dsl-spec.md

Runtime Notes (Bun-first, Node-supported)

• Use Bun for local development and tests.
• Packages in packages/core, packages/venues, and packages/cli must remain Node-compatible.
• Prefer standard fs APIs over Bun.file in shared packages.
• If using bun:sqlite, provide a Node fallback (better-sqlite3).

Testing Conventions

import { describe, test, expect } from "bun:test";

describe("Feature", () => {
  test("does something", () => {
    expect(result).toBe(expected);
  });
});

Test files are named *.test.ts and colocated with source files.

Onchain Test Suite

# Simulate only (no wallet needed):
./scripts/run-onchain-tests.sh

# Dry-run (builds txs, estimates gas, does NOT send):
./scripts/run-onchain-tests.sh --dry-run

# Live execution on Base:
CHAIN=8453 ./scripts/run-onchain-tests.sh --execute

# Resume from a specific phase after failure:
./scripts/run-onchain-tests.sh --execute --start-phase 4

# Recovery mode - return stranded funds to Base:
./scripts/run-onchain-tests.sh --recover

Phases: 0 (compile) -> 1 (simulate pure) -> 2 (simulate features) -> 3 (cast features) -> 4 (cast venues) -> 5 (multi-chain). Phase 5 writes checkpoints to .grimoire/test-suite/checkpoint and skips completed sub-steps on resume. Phase 5 is skipped if earlier phases have failures.

Expression Parser

The expression parser (packages/core/src/compiler/expression-parser.ts) handles:

• Arithmetic: +, -, *, /, %
• Comparison: ==, !=, <, >, <=, >=
• Logical: and, or, not (case-insensitive)
• Ternary: condition ? then : else
• Function calls: max(a, b), min(x, y), abs(n), sum(arr), avg(arr), to_number(n), to_bigint(n)
• Query functions: price(base, quote, source?), balance(asset, address?)
• Type conversions: to_number(bigint_val) → number, to_bigint(number_val) → bigint
• Property access: obj.prop, arr[0]

Usage Example

import { compile, execute, SqliteStateStore, createRunRecord } from "@grimoirelabs/core";
import { adapters, createAlchemyQueryProvider } from "@grimoirelabs/venues";

// Compile a spell
const result = compile(spellSource);
if (!result.success) {
  console.error(result.errors);
  process.exit(1);
}

// Load persisted state
const store = new SqliteStateStore();
const persistentState = (await store.load(result.ir.id)) ?? {};

// Execute the spell
const execResult = await execute({
  spell: result.ir,
  vault: "0x...",
  chain: 1,
  params: { amount: 1000 },
  persistentState,
  adapters,
  queryProvider: createAlchemyQueryProvider({
    provider,
    chainId: 1,
    vault: "0x...",
    rpcUrl: "https://eth-mainnet.g.alchemy.com/v2/YOUR_KEY",
  }),
});

// Persist results
await store.save(result.ir.id, execResult.finalState);
await store.addRun(result.ir.id, createRunRecord(execResult));
await store.saveLedger(result.ir.id, execResult.runId, execResult.ledgerEvents);
store.close();