Recursive Quality

Recursive Quality

Contract Richness

Wave 0 contracts set the ceiling. Thin contracts -> thin implementations.

Completeness

• Enumerate all operations, not just common ones
• Enums for closed sets (prevents string-typing drift)
• Error types alongside success types (Record | ParseError)
• Custom exception hierarchy -- base + subtypes per failure domain

Flexibility

• Batch interfaces (list[Spec]) over single-item
• Injectable I/O (write callables, output streams) for testability
• Union return types for expected failures; exceptions for unexpected

Modularity

• Protocols/interfaces for DI -- not concrete dependencies
• Factories/registries for runtime selection
• Structured return types (RunResult) over bare values

Typing boundaries

• Untyped/dynamic values (Any, object) are acceptable at input boundaries -- parsed records, CLI args, external data where the shape is genuinely unknown
• All internal result types (aggregation results, pipeline outputs, structured returns) must be fully typed -- the shape is known at design time
• Contracts file contains: protocols, dataclasses, enums, type aliases, constants. No callable stubs.

Critical Rules

These are never judgment calls. Implementers must not produce code that violates them. Reviewers must flag violations as Critical.

• Typed returns. No untyped/dynamic returns (Any, object, interface{}) on public functions or protocol methods. Use the specific type or union.
• Structured returns. No unstructured collection returns where contracts define or should define a structured type.
• No absolute paths. No hardcoded absolute paths in source or tests. Use the language's path library with relative paths.
• Logging, not stderr. Library code uses the language's logging framework with named loggers. Never write directly to stderr. Entry points configure level/format.
• Typed protocol signatures. Protocol/interface method parameters and returns must be fully typed. Untyped signatures cascade to every downstream module.

Quality Patterns

• Graceful degradation -- yield errors for bad input and continue. Don't crash the pipeline.
• System boundary validation -- validate external input at the boundary. Don't trust parsed data.
• Logging -- use the language's framework (logging, pino, log). Library code uses named loggers, never writes stderr. Entry points configure level/format.
• Error handling -- union types for expected failures, exceptions for unexpected. Messages include field name, bad value, what was expected.
• Exception hierarchy -- base project exception + specific subtypes. Part of Wave 0 contracts.
• Cross-platform correctness -- use the language's path library (pathlib, path.join, std::path) for all file operations. Specify encoding explicitly on file I/O (e.g. encoding="utf-8"). No string path concatenation. No hardcoded absolute paths. No platform-specific assumptions.
• Runtime versioning -- expose version programmatically. CLI supports --version.
• Test coverage -- configure with minimum threshold. Coverage config in project config file.
• Content-based detection -- auto-detect formats from content, not just file extensions.

Anti-Patterns

• God orchestrator -- orchestrator micromanaging every file. It manages waves and contracts, not implementation.
• Shared file optimism -- assuming parallel agents won't conflict. Verify scope overlap before dispatching.
• Context dumping -- passing full plan to every child. Each gets contract + scope only.
• Test-after at depth -- skipping TDD at leaf level. Every level writes tests first.
• Unbounded retry -- cap at 2, then escalate.
• Artificial splitting -- files created just to delegate. Split for distinct algorithms only.
• Over-specified plans -- prescribing file structure, test cases, sub-components. Agents discover through TDD.
• Hardcoded absolute paths -- use relative paths from __file__ or project root.