v2.0.0

Added

• Workflow 0 orchestration contract in templates/guide.md and platform
  boot docs: every substantive prompt runs platform detect → prompt-optimize
  → HyperRecall bundle → skills ensure → orchestrate, unless the user prompt
  itself opts out (no-optimize, skip prompt optimize,
  MEMORY_HIVE_NO_OPTIMIZE, or <!-- mh:no-optimize -->).
• New CLI verbs powered by memory_hive_orchestrate.py:
  prompt-optimize, platform detect, orchestrate, skills match|ensure|build,
  bench suite.
• HyperRecall recall expand to resolve HiveCodes / path:lines.
• Multi-root skill indexing across Cursor, Claude, Codex, Hermes, and
  ~/.agents/skills.
• SessionStart hooks prefer recall bundle --cache over legacy mega-concat.
• Reproducible memory-hive bench suite with dual baselines (naive full-boot
  vs HyperRecall/v0.3.2-style) and JSONL history under
  hive/.hivecode/bench/history.jsonl.

Measured (this machine, 2026-07-09)

Baseline	Tokens into agent turn	Notes
Naive full-boot corpus	139,614	index+registry+knowledge+distilled+silo heads
v0.3.2-style HyperRecall bundle	1,190	recall bundle max 1200
v2.0.0 optimize+bundle	1,005	prompt-optimize + budgeted bundle

• 99.28% token reduction vs naive (≥70% faster agent-turn claim; methodology: token reduction dominates model latency).
• 138.9× efficiency vs naive (naive_tokens / v2_tokens, ≥2.2× / “120% more efficient”).
• Improved vs HyperRecall-only: v2 bundle 969 ≤ v032 1190 tokens with orchestration metadata.
• Warm HyperRecall wall clock ~46ms vs naive boot+grep ~83ms (local I/O; not the primary claim).

Changed

• Installer/platform docs emphasize Grok/Cursor model preference for worker
  lanes; IDE-selected model remains the planner.

---

Commits since v1.9.0

• feat: Memory Hive v2.0.0 — orchestrate, optimize, route, prove (548115e)
• Add AGENTS.md with Cursor Cloud dev environment instructions (#31) (f44daae)