Writing a Custom Solver

This guide covers everything you need to build, package, test, and submit a custom IntentSolver to Minotaur Subnet 112.

Overview

A solver submission is a git repository containing:

my-solver/
├── Dockerfile        # FROM ghcr.io/subnet112/solver-base:v1
├── solver.py         # class MySolver(IntentSolver): ... ; SOLVER_CLASS = MySolver
├── requirements.txt  # Additional pip dependencies (optional)
└── README.md         # Description of your solver's approach

The validator clones this repo, runs it through a three-stage screening pipeline, benchmarks it against active App Intents, and adopts it if it delivers strictly more than the current champion under the relative reference-bar rule (per-order: zero regressions/drops and at least one strict win or blind-spot cover).

Dockerfile Requirements

Your Dockerfile must meet these requirements:

Base image: Must use FROM ghcr.io/subnet112/solver-base:v1
No CMD or ENTRYPOINT: The harness manages the entry point. Including either directive causes screening failure.
Repo size: Total repository must be under 100MB (excluding .git).
No suspicious binaries: Binary files (.so, .dll, .exe, .bin, etc.) over 10MB outside of models/ directories are rejected.

Minimal Dockerfile

FROM ghcr.io/subnet112/solver-base:v1

COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt

COPY . /app
WORKDIR /app

Container Execution Environment

When the validator runs your solver, the container is launched with strict isolation:

Constraint	Value
Network	`--network=none` (no internet access)
Filesystem	`--read-only` (with `/tmp` tmpfs)
Memory	`--memory=4g`
CPU	`--cpus=2.0`

Your solver must work entirely offline. Any data your solver needs (routing tables, model weights, pool lists) must be bundled in the Docker image.

solver.py Structure

Your solver module must:

Subclass IntentSolver from minotaur_subnet.sdk.intent_solver
Implement initialize(), generate_plan(), and metadata()
Export SOLVER_CLASS at module level

Complete Example

import time
from typing import Any

from minotaur_subnet.sdk.intent_solver import IntentSolver, MarketSnapshot, SolverMetadata
from minotaur_subnet.shared.types import (
    AppIntentDefinition,
    ExecutionPlan,
    Interaction,
    IntentState,
)
from minotaur_subnet.v3.contexts import SwapIntentContext


class AdvancedSwapSolver(IntentSolver):
    """Solver with RPC-based pool discovery and multi-hop routing."""

    def __init__(self):
        self.rpc_urls: dict[int, str] = {}
        self.chain_ids: list[int] = []
        self.routing_table: dict = {}

    def initialize(self, config: dict[str, Any]) -> None:
        self.chain_ids = config.get("chain_ids", [1])
        self.rpc_urls = config.get("rpc_urls", {})
        # Build routing tables, load models, etc.
        self.routing_table = self._build_routing_table()

    def generate_plan(
        self,
        intent: AppIntentDefinition,
        state: IntentState,
        snapshot: MarketSnapshot | None = None,
    ) -> ExecutionPlan:
        chain_id = state.chain_id or 1
        if isinstance(state.typed_context, SwapIntentContext):
            input_token = state.typed_context.input_token
            output_token = state.typed_context.output_token
            input_amount = state.typed_context.input_amount
        else:
            input_token = state.raw_params.get("input_token", "")
            output_token = state.raw_params.get("output_token", "")
            input_amount = int(state.raw_params.get("input_amount", "0"))

        # Query pool states via RPC if available, else use snapshot
        if self.rpc_urls.get(chain_id):
            pool_states = self._query_pools_rpc(chain_id, input_token, output_token)
        elif snapshot and snapshot.pool_states:
            pool_states = snapshot.pool_states
        else:
            pool_states = {}

        # Find best route and build interactions
        route = self._find_route(pool_states, input_token, output_token, input_amount)
        interactions = self._build_interactions(route, chain_id)

        return ExecutionPlan(
            intent_id=intent.app_id,
            interactions=interactions,
            deadline=int(time.time()) + 300,
            nonce=state.nonce,
            metadata={
                "route": "custom_multi_hop",
                "hops": len(route),
                "chain_id": chain_id,
            },
        )

    def check_trigger(
        self,
        intent: AppIntentDefinition,
        state: IntentState,
        snapshot: MarketSnapshot | None = None,
    ) -> bool:
        # For perpetual intents: check if market conditions warrant execution
        return False

    def metadata(self) -> SolverMetadata:
        return SolverMetadata(
            name="advanced-swap-solver",
            version="1.0.0",
            author="5Grwva...",
            description="Multi-hop DEX aggregation with cross-protocol routing",
            supported_chains=[1, 8453],
            supported_intent_types=["swap"],
        )

    def serialize_state(self) -> bytes:
        # Persist learned routing data for next epoch
        import json
        return json.dumps(self.routing_table).encode()

    def restore_state(self, data: bytes) -> None:
        import json
        self.routing_table = json.loads(data.decode())

    # --- Private methods ---

    def _build_routing_table(self) -> dict:
        return {}

    def _query_pools_rpc(self, chain_id, token_in, token_out) -> dict:
        return {}

    def _find_route(self, pool_states, token_in, token_out, amount) -> list:
        return []

    def _build_interactions(self, route, chain_id) -> list[Interaction]:
        return [
            Interaction(
                target="0xC02aaA39b223FE8D0A0e5C4F27eAD9083C756Cc2",
                value="0",
                call_data="0xd0e30db0",
                chain_id=chain_id,
            ),
        ]


# Required: tells the harness which class to instantiate
SOLVER_CLASS = AdvancedSwapSolver

Prefer state.typed_context when the validator provides it. The raw state.raw_params dict remains available for untyped access, and runtime control metadata such as the intent function lives in state.control.

Three-Stage Screening Pipeline

Every submission goes through progressive screening before benchmarking. Screening stops at the first failure.

Stage 1: Static Checks (~10 seconds)

Check	Requirement
Required files	`Dockerfile`, `solver.py`, `README.md` must exist
Base image	Dockerfile must contain `FROM ghcr.io/subnet112/solver-base`
No entrypoint	Dockerfile must not contain `CMD` or `ENTRYPOINT`
Repo size	Total size must be under 100MB
No suspicious binaries	No `.so`, `.dll`, `.exe`, etc. over 10MB outside `models/`

Stage 2: Build Check (~2 minutes)

Step	What happens
Docker build	`docker build --network=none --memory=4g -t <tag> <repo>`
Import check	`from solver import SOLVER_CLASS` -- verifies the module loads
Subclass check	Verifies `SOLVER_CLASS` is a subclass of `IntentSolver`
Init check	Calls `SOLVER_CLASS().initialize({"chain_ids": [1]})`
Metadata check	Calls `metadata()` and verifies `name` and `version` are non-empty

Stage 3: Smoke Test (~5 minutes)

Step	What happens
Synthetic intents	3 synthetic intents are generated and passed to `generate_plan()`
Plan validation	Each plan is checked for structural correctness (see below)
Trigger check	`check_trigger()` is called for auto-triggered intents; must return `bool`
Per-plan timeout	Each `generate_plan()` call must complete within 30 seconds

Plan Validation Rules

A valid ExecutionPlan must satisfy:

intent_id matches the intent's app_id
interactions list is non-empty
deadline is after the snapshot timestamp
Each interaction's target is a 42-character hex address starting with 0x
Each interaction's call_data starts with 0x

Benchmarking

After passing all three screening stages, the solver is benchmarked against active App Intents on the network.

Scoring

Plans are scored by each app's JS scoring function (score(plan, state, context)). The JS score ranges from 0.0 to 1.0. Plans are also simulated on an Anvil fork to capture on-chain scores. Both scores must exceed the app's threshold.

Champion/Challenger Model

The currently active solver is the champion — the relative-scoring baseline (no score of its own).
A new submission is a challenger.
Each order is compared champion-vs-challenger at the same pin → win / regression / matched (within a ±0.1% / 10 bps tie band), plus blind_spot_cover (challenger serves an order the champion can't → win) and dropped (the reverse → regression).
The challenger dethrones only with zero regressions/drops and at least one strict win or blind-spot cover. Matching everywhere is rejected; any regression makes the verdict behind.
Once adopted, the challenger becomes the new champion and processes real orders.

Auto-Triggered Intents

For perpetual (auto-triggered) intents, solvers are also evaluated on trigger accuracy. The composite score is:

composite = 0.4 * trigger_accuracy + 0.6 * plan_quality

Where trigger_accuracy measures how well check_trigger() predicts when execution is warranted.

Using BaselineSwapSolver as Reference

The BaselineSwapSolver at minotaur_subnet/sdk/solvers/baseline_solver.py is the default champion. Study it to understand:

RPC-first architecture: Queries Uniswap V3 pool states via RPC, falls back to snapshot.
Factory-based pool discovery: Uses the Uniswap V3 Factory contract to find pools for any token pair across all fee tiers (100, 500, 3000, 10000).
Multi-hop routing: Finds optimal routes through intermediary tokens (WETH, USDC) when direct pools have poor liquidity.
Cross-chain support: Generates multi-leg plans (source swap + bridge + destination action) when dest_chain_id differs from the source chain.
Pool state caching: Caches pool states with a 12-second TTL (one Ethereum block).
Price derivation: Derives USD prices from pool sqrtPriceX96 values.

Strategies to Beat the Baseline

More pool discovery: Scan factory events for all deployed pools, not just known addresses.
Cross-DEX aggregation: Route through multiple DEXes (SushiSwap, Curve, Balancer) for better prices.
Split routing: Split large orders across multiple pools to reduce price impact.
MEV protection: Use Flashbots-style techniques to protect orders from sandwich attacks.
ML-based parameter tuning: Use serialize_state() / restore_state() to learn optimal slippage tolerances and routing preferences across epochs.
Gas optimization: Minimize the number of interactions and calldata size.

Using the RoutingSolver with Strategies

If your solver needs to handle multiple apps, use the RoutingSolver with per-app Strategy instances:

from minotaur_subnet.sdk.routing_solver import RoutingSolver
from minotaur_subnet.sdk.strategy import Strategy
from minotaur_subnet.shared.types import ExecutionPlan, Interaction

class SwapStrategy(Strategy):
    APP_ID = "swap-app-001"
    INTENT_FUNCTIONS = ["execute"]

    def generate_plan(self, intent, state, snapshot=None):
        # Swap-specific logic
        ...

class VaultStrategy(Strategy):
    APP_ID = "vault-app-002"
    INTENT_FUNCTIONS = ["buyDip", "withdraw"]

    def generate_plan(self, intent, state, snapshot=None):
        intent_function = (
            getattr(state.typed_context, "intent_function", "")
            or state.control.get("_intent_function", "")
        )
        if intent_function == "buyDip":
            return self._buy_dip_plan(intent, state, snapshot)
        else:
            return self._withdraw_plan(intent, state, snapshot)

    def check_trigger(self, intent, state, snapshot=None):
        # Check price conditions for auto-triggered buyDip
        return True

    # ... private methods ...


# Wire it up
solver = RoutingSolver()
solver.register_strategy(SwapStrategy())
solver.register_strategy(VaultStrategy())

SOLVER_CLASS = RoutingSolver

The RoutingSolver generates a minimal fallback plan for any intent that does not match a registered strategy.

Testing Before Submission

1. Local smoke submission

curl -X POST http://localhost:8080/v1/submissions/source \
  -H "Content-Type: application/json" \
  -d '{"solver_source":"<python source>", "hotkey":"local-miner", "epoch":0, "solver_name":"local-smoke"}'

2. Static Check Only

Run just Stage 1 on your repo directory to verify file structure before pushing:

from minotaur_subnet.harness.screening import run_stage_1
result = run_stage_1("/path/to/my-solver")
print(result.passed, result.details)

3. Full Screening Locally

If you have Docker available, run the full screening pipeline:

import asyncio
from minotaur_subnet.harness.screening import ScreeningPipeline

async def test():
    pipeline = ScreeningPipeline()
    result = await pipeline.run_all("/path/to/my-solver", commit_hash="abc123")
    print(result.to_dict())

asyncio.run(test())

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Writing a Custom Solver

Overview

Dockerfile Requirements

Minimal Dockerfile

Container Execution Environment

solver.py Structure

Complete Example

Three-Stage Screening Pipeline

Stage 1: Static Checks (~10 seconds)

Stage 2: Build Check (~2 minutes)

Stage 3: Smoke Test (~5 minutes)

Plan Validation Rules

Benchmarking

Scoring

Champion/Challenger Model

Auto-Triggered Intents

Using BaselineSwapSolver as Reference

Strategies to Beat the Baseline

Using the RoutingSolver with Strategies

Testing Before Submission

1. Local smoke submission

2. Static Check Only

3. Full Screening Locally

Submission Checklist

Uh oh!

FilesExpand file tree

custom-solver.md

Latest commit

History

custom-solver.md

File metadata and controls

Writing a Custom Solver

Overview

Dockerfile Requirements

Minimal Dockerfile

Container Execution Environment

solver.py Structure

Complete Example

Three-Stage Screening Pipeline

Stage 1: Static Checks (~10 seconds)

Stage 2: Build Check (~2 minutes)

Stage 3: Smoke Test (~5 minutes)

Plan Validation Rules

Benchmarking

Scoring

Champion/Challenger Model

Auto-Triggered Intents

Using BaselineSwapSolver as Reference

Strategies to Beat the Baseline

Using the RoutingSolver with Strategies

Testing Before Submission

1. Local smoke submission

2. Static Check Only

3. Full Screening Locally

Submission Checklist