leviathantalon-swarm 1.0.0

LeviathanTalon: hierarchical Kabbalistic multi-agent swarm orchestrator. 2,832 concurrent agents at full scale.

LeviathanTalon Swarm

Hierarchical Kabbalistic multi-agent orchestrator. 1 master + 7 sefirotic sub-orchestrators + 700 LLM subagents + 3 worker pools of 708 each = 2,832 agents at full scale.

Built for HaChazal (Chaz Leland Hamm) | Metanoia Unlimited LLC | Anthropic CPN.

---

Architecture

                    HaChazal (Da'at) — Master Orchestrator
                                  |
        +-------+--------+--------+--------+--------+--------+
        |       |        |        |        |        |        |
    Zadkiel  Gavriel  Raphael   Uriel   Michael  Malika  Katan HaShem
    Chesed  Gevurah Tiphareth Netzach    Hod    Yesod     Malkuth
   (LeadGen)(Security)(Creative)(Crypto)(Comms) (Hive)  (Manifest)
        |       |        |        |        |        |        |
      [up to 100 LLM subagents per sub-orchestrator = 700 total]
                                  |
        +=====================================================+
        |     WORKER POOLS (drawn from by any sub-orch)       |
        +=====================================================+
        |  ZeroClaw  (708)  - deterministic, no LLM, FREE     |
        |  PicoClaw  (708)  - haiku LLM only                  |
        |  NanoClaw  (708)  - haiku LLM + tool use            |
        +=====================================================+

TOTAL: 1 + 7 + 700 + 2,124 = 2,832 concurrent agents

Top triad (Keter / Chokmah / Binah) is accessed as source, not commanded.

Feedback loop: Malkuth → Yesod → Da'at closes the self-learning circuit via the neural bus.

---

Quickstart

pip install -e ".[llm]"          # install with Anthropic SDK
export ANTHROPIC_API_KEY=sk-...   # optional, enables LLM tiers

# Launch the swarm at any scale 0..2832
lvtn start --config config.yaml

# Run a quick demo
python examples/demo_run.py

# Run the test suite (smoke + live scaling)
pytest -v

---

Scaling

The public scale knob is 0..2832. Three pools, 708 each:

Tier	LLM?	Cost	Use case
ZeroClaw	No	Free	Deterministic ops (hash, parse, transform, math, regex)
PicoClaw	haiku	Cheap	Single-shot text generation
NanoClaw	haiku + tools	Cheap-ish	Tool-using subtasks (search, fetch, etc.)

LLM tier (708 sub-orch subagents) uses haiku by default.

# config.yaml
scale:
  zeroclaw_pool: 708
  picoclaw_pool: 708
  nanoclaw_pool: 708
  max_llm_subagents_per_sub: 100
mesh:
  backend: in_memory   # or 'redis'
  redis_url: redis://localhost:6379

---

Three Laws (Sage Bet Discipline)

This package is the immutable v1 baseline stored under .lpls/baseline/.

1. FREEZE — working state is frozen. Do not edit in place.
2. LAYER — extensions ship as patches in .lpls/patches/NNNN_name/.
3. VERIFY — each layer is tested before the next stacks.

See HONEST_STATUS.md for the current truth of every feature.

---

Honest Status Files (Sage Gimel)

• HONEST_STATUS.md — every feature labeled WORKING / PARTIAL / MOCKED / BLOCKED / NOT STARTED
• BUILD_PROGRESS.md — append-only session log
• BLOCKERS.md — open issues
• TEST_REPORT.md — evidence-based test results
• SECURITY_NOTES.md — threat model + mitigations
• COMPLETION_CHECK.md — 15-item No-Scaffolding Contract

---

License

Proprietary — Metanoia Unlimited LLC. Built under the Preservation Doctrine.