Module micro_mode 

Micro Evolution Mode — For 0.8B-3B Parameter Models

Small models struggle with:

• Large context windows (>4K tokens)
• Complex multi-file reasoning
• Precise JSON array formatting
• Multi-step search-and-replace

This module provides simplified mutation strategies that trade sophistication for reliability on resource-constrained hardware.

Constants

MICRO_MAX_CONTEXT_CHARS

Maximum context size for micro mode (in characters) ~4K tokens ≈ 16K chars for code

MICRO_MAX_EDITS

Maximum edits per hypothesis in micro mode Small models do better with single, focused changes

MICRO_MAX_FILES

Maximum files to include in micro mode context

Functions

build_micro_context

Build micro-mode context with aggressive truncation

build_micro_system_prompt

Simplified system prompt for micro models

is_micro_model

Detect if we should use micro mode based on model name. Matches small parameter counts as word-bounded tokens to avoid false positives (e.g. “32b” matching “2b”, “13b” matching “3b”).

select_micro_targets

Filter and limit mutation targets for micro mode Returns the smallest files first, limited by MICRO_MAX_FILES

validate_micro_hypothesis

Validate that a hypothesis is “micro-safe”