Skip to content

ai-assisted.md

Latest commit

 

History

History
42 lines (30 loc) · 4.34 KB

ai-assisted.md

File metadata and controls

42 lines (30 loc) · 4.34 KB

AI-Assisted Development

Canton's documentation is centralized and Daml has thin training-data coverage, so working effectively with AI on Canton is worth treating as its own skill.

General-purpose coding assistants

  • Claude Code - agentic CLI and IDE integration with large context windows
  • Cursor - AI-first IDE with a rules system for encoding Daml conventions
  • GitHub Copilot - ubiquitous inline completions
  • Windsurf - agentic IDE
  • Aider - terminal-native pair programmer
  • Continue - open-source IDE assistant with swappable models

Which assistant is best suited for Daml

No model is Daml-native. The relevant signal is how well a model handles Haskell-family syntax, since Daml derives from Haskell. In practice, frontier models from Anthropic and OpenAI tend to handle Daml better than code-specialized models whose training skews toward JavaScript, Python and Solidity (DeepSeek Coder, Codestral). Large context windows matter more than code specialization because current Canton docs must be fed in-context.

Making an LLM better at Daml

Ordered from lowest to highest effort.

  • Rules file: encode Daml idioms in CLAUDE.md, .cursorrules, or AGENTS.md. Pin the SDK version, list template and choice naming conventions, and instruct the agent to derive observers from business rules rather than copy them from other templates.
  • Few-shot examples: paste canonical templates (Iou, Proposal-Accept, Delegation) into the prompt, and at least one Daml Script that must fail with a missing-authorization error.
  • RAG over official sources: index docs.daml.com, docs.digitalasset.com, forum.canton.network, and the digital-asset/daml source.
  • MCP servers: point an MCP-capable agent at the Daml SDK tree, Daml Finance source, and the Canton forum. For project-local context, a lightweight MCP over the repo's daml/ tree and daml.yaml scopes the agent to your own code.
  • Agent skills: package domain knowledge as reusable skills or agent tools: a Daml code review skill that checks authorization and privacy invariants, a test scaffold skill that generates Daml Scripts per template choice, a Canton deployment skill that produces participant or validator config from a template.
  • Custom scripts: a pre-commit hook that runs daml build and daml test and feeds failures back to the agent; an authorization linter that flags mismatched signatory/observer/controller sets; a DAR diff tool that warns on breaking template-evolution changes.
  • Fine-tuning: LoRA or full fine-tune on open-weights models (Llama, Qwen, DeepSeek) using your internal Daml corpus plus the public sources. Only worth it at scale.

AI for Daml testing

  • Prompt the model to enumerate signatory and observer edge cases before writing any contract code.
  • Ask explicitly for a Daml Script that should fail with a missing-authorization error, and verify the expected failure.
  • Use AI to review contracts for missing signatories, observers, controllers, and privacy leaks.

AI for general Canton usage

  • Query ledger state via the JSON Ledger API with an AI agent as an operational assistant. Use the prebuilt @canton-network/core-ledger-client or generate a typed client from the OpenAPI spec.
  • Scaffold dApp frontends against the @canton-network/dapp-sdk with AI-generated wiring.
  • For synchronizer operations, AI is useful for log and trace triage but unreliable for config generation; cross-reference the operate docs.
  • Expose Canton tooling to an agent via MCP: the canton-mcp-server wraps daml and canton CLIs, and the Lighthouse explorer MCP gives read access to CNS, governance, validators, and contracts.
  • Pin the target Canton SDK version in the repo's rules file; the agent will otherwise mix 3.3/3.4/3.5 doc paths and API shapes.