v0.8.0 — Sprint 1-8 complete (219 tests, 20 papers, ContentOptimizer)

Highlights

• All 8 sprints delivered of the 16-week plan (~/.windsurf/plans/are-os-v3-implementation-25a40c.md).
• 219 / 219 tests pass (~6.6 s), runs fully offline (Mock LLM provider + offline arxiv fixtures).
• 20 papers implemented 1-to-1 in sf3/papers/ with smoke registry.
• Post-plan enhancement: factory-level ContentOptimizer + sf3 skill opt --deep CLI.
• Real-world demo: a 192-line Windsurf workflow (figma-to-flutter.md) was ingested, analyzed, optimized (validation_score 0.82 → 0.91), and split into 4 atomic SkillNodes in 12.7 ms total — see analysis_runs/optimization_report.md.

Continuous integration

• GitHub Actions matrix: py3.11 + py3.12 on ubuntu-latest.
• ruff check sf3 tests clean (zero warnings).
• pytest -q must stay green; pyproject.toml is the single source of dev-tooling config.

Modules shipped

• sf3/schema/ — Pydantic v2 SkillNode (P0 / P1 / P2 fields)
• sf3/llm/ — 7-provider abstraction with auto-fallback (DeepSeek / Ollama / Mock / OpenAI / Anthropic / GLM / Qwen)
• sf3/storage/ — LocalFS + fnOS + SMB + NASRouter (3-tier degrade) + SQLiteIndex + ShardManager
• sf3/harvester/ — arxiv + relevance filter + paper-to-skill
• sf3/papers/ — 20 papers (Phase-0 + CAF + RVF + Collision + ELL)
• sf3/factory/ — generate / lifecycle / ContentOptimizer
• sf3/{caf,rvf,collision,attention,lifelong}/ — ARE engine
• sf3/orchestrator/ — DAG + 14 metrics + Go/No-Go
• sf3/cli/ — typer app with 8 subcommands

Sprint roadmap status

Sprint	Key deliverable	Status
1	schema + CLI skeleton	✅
2	paper harvester + LLM fallback + arxiv	✅
3	async neuro-symbolic generator + lifecycle	✅
4	CAF 4-layer async pipeline + 7 papers	✅
5	RVF 3-type extractor + retry/degrade	✅
6	sharded collision engine + Auditor + DEAN	✅
7	ShardManager dynamic rebalance + legacy migrator	✅
8	orchestrator + 14 metrics + Go/No-Go cron	✅

What's next (Sprint 9 candidates)

• chain collision class: a 5th collision type beyond merge/split/rewrite/reject — currently chain dependencies are mis-classified as reject.
• causality_score ceiling lift: validate_causality currently caps at 0.91 (base=0.85 + nudge≤1.0). Needs the 0.6 / 0.4 weight tuned, or a third signal added.
• real-LLM smoke matrix: opt-in CI lane using a tiny budget on DeepSeek/Ollama to catch prompt regressions.

Acknowledgements

This release is the result of pair-programming between the maintainer and Cascade (an AI coding assistant). All commits in git log carry the Co-Authored-By: Cascade footer for full attribution.