Noricum Dev

Noricum Development Conventions

Architecture

• Noricum is an agent orchestrator, not a compiler
• C2Rust is "step zero" (subprocess), not reinvented — but often skippable (P4: --skip-c2rust)
• LLM agents are central from v0
• Semantic Code Map (noricum-ir) tracks metadata, not compiler IR
• rig-rs 0.31 for LLM integration (Rust-native, rustls)

Crate Dependencies (layered)

noricum-cli -> noricum-core -> noricum-agents -> noricum-tools -> noricum-ir
                            -> noricum-validation -> noricum-tools -> noricum-ir
noricum-mcp -> noricum-core

Coding Standards

• Edition 2024, thiserror for lib errors, anyhow for CLI
• tracing for logging, never println! in library code
• No unwrap() in library code
• Async with tokio
• Tests in #[cfg(test)] mod tests within each file
• C compiler: -std=gnu11 (not c11, needed for POSIX like strdup), -lm for math

State Machine

Pending -> Extracted -> Characterized -> C2RustDone -> Analyzed -> Refined -> Validated
                                                                    |
                                                            Repairing (max iterations)
                                                                    |
                                                            FallbackUnsafe

Model Router (Claude 4.6 era)

• Easy: claude-haiku-4-5 (fast, cheap)
• Medium: claude-sonnet-4-6 (also used for analysis)
• Hard: claude-opus-4-6 (translation, complex repair)
• Fallback: Ollama qwen2.5-coder:32b when no API key

Pipeline Improvements (P0-P5, learned from miniz + genann migrations)

• P0: Quality floor — repair rejected if unsafe count exceeds translation baseline
• P1: Best-version tracking — keeps highest-score version, uses it for fallback instead of c2rust
• P2: Per-function C2Rust context — extracts only matching c2rust functions per chunk
• P3: Incremental per-module — split_into_modules() groups C functions by prefix
• P4: Skip C2Rust — --skip-c2rust flag; LLM often translates better without c2rust noise
• P5: Idiomatic improvement hints — when code compiles + diff passes but score 800 LOC (MEDIUMFILELOC)
• Reduced repair iterations: >1000 LOC
• Very large file handling: >2000 LOC
• Structural summary: >800 LOC
• Quality gate: re-translate if >5 unsafe blocks
• Stall detection: 2 consecutive unchanged error counts → re-translate at temp 0.7
• Chunk targets: 400 LOC (medium), 500 LOC (very large)

Pipeline Improvements (P30-P33, learned from miniz_zip.c 9 runs)

• P30: Hybrid Repair Engine — 3-phase: rule engine (free) → surgical per-function (cheap) → legacy whole-file (expensive)
• P31: Assembly Cleanup — fence stripping, syntax error parsing, use import merging
• P32: Brace-Balance Validation — detect/fix truncated module outputs before assembly
• P32b: Smart Truncate + Re-translate — truncate at last balanced brace, re-translate truncated modules
• P33: Type Contract — types-first modular migration:
• generate_type_contract() in type_contract.rs — single LLM call before module translation
• ModuleSplit struct returns shared_context from split_into_modules()
• resolve_header_types() reads #include'd .h files for complete type definitions
• Assembly seeds P27 dedup from contract type names
• Key learning: shared_context only has .c file content; .h headers must be resolved separately

Key Patterns

• Functions migrate independently, ordered by dependency graph (topological sort)
• Every migration must pass differential testing (byte-exact stdout match)
• Translation cache: .noricum-cache/ keyed by SHA-256 of C source
• RAG patterns: successful migrations auto-added to PatternStore for future context
• Structural chunking: data model (structs/constructors) in chunk 0, logic functions in later chunks
• For multi-file C projects: type definitions live in .h headers, not .c files — must resolve includes
• Type contract prompt must enforce idiomatic Rust (no raw pointers, no C-style aliases, complete structs)

---

Powered by Noricum — autonomous C/C++ to Rust migration agent

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: JuanMarchetto
• Source: JuanMarchetto/agent-skills
• License: MIT
• Homepage: https://github.com/JuanMarchetto/agent-skills

Install and usage instructions live in the source repository linked above.