Changelog¶

All notable changes to FixtureForge are documented here.

[2.1.0] — 2026-04-13¶

pytest plugin — forge_fixture() declares model fixtures in one line from conftest.py
forge fixture — auto-available in every test with zero config; CI-safe by default
forge_swarm_fixture() — parallel multi-model fixture with shared AI cache
seed= parameter — fully deterministic output across runs (Forge(use_ai=False, seed=42))
verbose= parameter — shows exactly which field was generated by AI / Faker / structural router
pytest plugin auto-registered via pytest11 entry point (no conftest.py config needed)

Faker.seed() replaced with faker.seed_instance() — two Forge(seed=42) instances no longer interfere with each other's RNG state
Global random module replaced with random.Random(seed) per instance — full RNG isolation

README rewritten from scratch — full architecture docs, comparison table, all v2.0 features

ForgeMemory — FORGE.md pointer index with on-demand topic files
DataSwarms — parallel multi-model generation with shared AI cache (~90% cheaper per additional model)
Permission Gates — safe / sensitive / dangerous data classification with Mailbox Pattern
ForgeDream — 4-phase coverage consolidation (Orient -> Gather -> Consolidate -> Prune)
Three-Layer Compression Pipeline — Micro / Auto / Full session compression
ForgeDataset — automatic disk-spill for large results (>50K chars)
Feature Flags — FORGE_SWARMS, FORGE_DREAM, FORGE_KAIROS, etc.
SYSTEM_PROMPT_DYNAMIC_BOUNDARY — static/dynamic prompt split for ~40% API cost reduction
Multi-provider support: Anthropic Claude, OpenAI, Google Gemini, Groq, Ollama