Skip to content

AGENTS.md

Latest commit

 

History

History
189 lines (131 loc) · 9.87 KB

AGENTS.md

File metadata and controls

189 lines (131 loc) · 9.87 KB
Raw

AGENTS.md

This file provides guidance to AI Agents when working with code in this repository.

Project Overview

Sovereign SDK is a Rust toolkit for building rollups with real-time soft-confirmations, high performance (30k+ UOPS), and pluggable data availability (Celestia, Bitcoin) and zkVM (Risc0, SP1) adapters.

Build Commands

make lint                   # Run all linters (fmt, clippy, zepter, dylint)
make lint-fix               # Auto-fix linting issues
make test                   # Run tests with nextest
make check-features         # Verify all feature combinations compile
make mini-ci                # Full pre-submission checks
make install-dev-tools      # Install all development dependencies

Instead of running make build use make lint or cargo check --all-features for faster feedback loops.

Running Tests

cargo nextest run <test_name>
cargo nextest run -p <package_name> <test_name>
  • Prefer running tests with -p to reduce rebuild times.
  • Always use cargo nextest over regular cargo test. Cargo nextest runs each test in a separate process which gives isolation of environment variables.
  • This project has strict testing practices. It means that tests on base branch(dev) are always passing. Any consistent failure in the branch is certainly introduced in the branch.

Debugging

  • Logs: add sov_test_utils::initialize_logging() and run test like this. RUST_LOG=debug cargo nextest run -p <pkg> <test>. If there's too much information, adjust RUST_LOG to only emit logs from relevant modules.
  • For DB issues: Check sov-db traces with RUST_LOG=sov_db=trace
  • Feel free to add println! or logs if it helps.

Environment Variables

  • PROPTEST_CASES=50 - Faster local proptest runs (CI uses more)
  • SKIP_GUEST_BUILD=1 - Skip risc0 guest builds during development
  • SP1_SKIP_PROGRAM_BUILD=1 - Skip SP1 program builds during development
  • SOV_TEST_SKIP_DOCKER=1 - Skip docker-based tests, useful for local development without docker

Always use SKIP_GUEST_BUILD=1 unless performing specific ZK related changes, because it massively speeds up compilation time.

Architecture

Key Design Principles

  • Modularity: Pluggable DA(Data Availability) layers, storage backends, and ZK systems
  • Type Safety: Strong typing with Rust's type system, because it allows checking correctness at compile time.
  • Determinism: All operations deterministic for ZK proving. No HashMap/HashSet iteration order, no system time, no random. Breaks ZK proofs.
  • Gas Metering: Automatic tracking of compute and storage costs
  • State Isolation: Each module's state is namespaced. Modules cannot directly access other modules' state; use hooks or explicit dependencies / methods

Four Major Layers

  1. Rollup Interface (crates/rollup-interface/) - Core traits defining rollup behavior
  2. Module System (crates/module-system/) - Framework for building modules
  3. Adapters (crates/adapters/) - DA and zkVM integrations
  4. Full Node (crates/full-node/) - Node infrastructure

Delegate to subagents & AGENTS.md in those directories for more details.

Bindings

  • ./typescript/: TypeScript bindings for serialization, transaction building, and RPC client
  • ./python/: Python bindings for simple transaction building & serialization in python

Code Quality Requirements

Avoid over-engineering solutions. Prioritise clarity and maintainability. Don't prematurely add code or features that aren't needed.

Philosophy

This codebase will outlive you. Avoid shortcuts that create long-term debt.

  • Validate the plan and downstream impact; small mistakes can propagate.
  • Take time to understand the question before changing code.
  • Keep docs/comments in sync with behavior.
  • Improve logging around changed areas if it helps future debugging.

Non-Determinism

  • Always avoid non-deterministic code paths in modules and core logic, because non-determinism breaks consensus code and ZK proofs for the light clients.
  • For tests, avoid using sleep, prefer deterministic detection of expected changes.

Safe Arithmetic

  • Use checked/saturating operations for balances, gas calculations, and memory-bound operations
  • No float arithmetic (lint denial)

Feature Gates

  • native feature: Code not executed in zk proofs
  • Most crates use default-features = false for zkVM compatibility

Linting

  • Clippy with custom configuration
  • Zepter for feature-gate consistency
  • Dylint for Sovereign-specific lints
  • Nightly rustfmt (use rustfmt.nightly.toml)

Communication

  • Always favour clear and concise over verbose. Add more detail when asked
  • Always state if you are guessing or making an assumption
  • Always link to relevant files, lines, or documentation when referencing code or answering questions
    • Prefer style [filename:line_number] for inline references
    • Prefer style [filename] for general references

Code Review

  • Base branch for this project is dev.
  • Focus on business logic, correctness and code quality which has the highest business impact. Ignore untracked files.
  • Ignore uncommitted changes in Cargo.toml, if it removes default-members. This is expected pattern with cargo switcheroo

Scanning for state breakage

A state break is when the same inputs produce a different state root — an unintentional hard fork that desyncs existing chains. It is distinct from a chain-hash break (a change to the runtime schema's identity), which is expected when a message format changes and is absorbed via CHAIN_HASH_OVERRIDES. A change can — and should — be chain-hash-breaking without being state-breaking. Scan any change touching state layout, execution paths, or serialized types for these:

  • Eager genesis initialization. Writing a new #[state] value at genesis adds a key to the tree, changing the genesis (and empty-batch) state root for every chain. Only write when non-default, and make a missing value read back as the default (get(...).unwrap_or_default()).
  • New metered reads/writes in execution paths. State accesses are gas-metered, and gas feeds the fee deducted (and written) from the payer. A single extra (or one less!) StateValue::get on a transaction or hook path changes gas → fees → balances → state root if it's ever exercised. Be careful on the entire execution path: both changes to existing CallMessage handlers in a module, as well as any state access changes across the STF transaction processing pipeline.
  • Serialization layout of replayed or stored types. Transactions are replayed from DA during resync, and state items are hashed into the state root; so any type either reachable from CallMessage / RuntimeCall, or stored in the rollup state, must keep a stable, byte-perfect serialized layout. For both binary codecs used here (borsh for messages, BCS for state), only appending enum variants is safe (preserves existing discriminants); you must NOT add, remove, reorder, or retype struct or tuple fields, nor insert/reorder enum variants. Genesis-config types decoded from JSON using serde_json are exempt when the new field is #[serde(default)], because serde_json's format is self-describing.
  • State item ordering. #[state] field prefixes are positional within a module: appending a new field is safe, but inserting or reordering shifts the storage keys of all following fields. Module discriminants are pinned in constants.toml.

These rules exist to protect deployed state and historical messages. The release boundary is the nightly branch (unless the user says otherwise): commits already in dev but not yet in nightly are unreleased and safe to change in state-breaking ways. In exceptional cases the user may also confirm that some type or commit is known to not have been used in production yet. Test-only modules are also safe to change. When in doubt about whether something is deployed, ask the user; by default assume anything in a nightly branch is released.

Agent working agreement

Default operating mode: Minimal Patch Mode

Unless the user explicitly asks for a refactor, redesign, or broader cleanup, operate in Minimal Patch Mode.

Rules:

  • Make the smallest diff that satisfies the request.
  • Preserve the requested approach, constraints, and test strategy.
  • Prefer editing existing code over introducing new abstractions.
  • Do not fix adjacent issues in the same PR.
  • Put unrelated ideas under "Follow-ups", not in the diff.

Plan Lock

An approved plan is binding.

Rules:

  • Implement the approved plan, not a "better" plan.
  • Do not silently replace requested techniques, test strategies, or constraints.
  • Do not widen scope because a broader change seems cleaner.
  • Do not drop explicit user requirements from earlier discussion just because they were omitted from a later summary.
  • If the plan omitted a prior explicit requirement, repair the plan before coding.

If you believe the approved plan is flawed, stop and report:

  1. the exact conflict or new evidence
  2. the smallest change to the plan that would fix it
  3. zero implementation changes beyond investigation

Any deviation from the approved plan must be called out explicitly under: Deviations from approved plan: ... Silent deviations are not allowed.

Scope budget

Default budget unless explicitly approved otherwise:

  • at most 4 files changed
  • at most 200 added non-test lines
  • no new dependencies
  • no public API changes
  • no file moves or broad renames
  • no opportunistic refactors
  • no formatting-only churn
  • no test rewrites unless explicitly requested

If the task cannot be completed within this budget, stop and propose the smallest split.

Review behavior

Reviewer direction wins over agent preference.

Rules:

  • When asked to reduce scope, reduce scope.
  • Do not argue for the broader patch after the reviewer asks for a smaller PR.
  • You may note one concise technical risk or tradeoff, then comply.
  • Remove incidental changes instead of defending them.