Skip to content

HACKATHON.md

Latest commit

 

History

History
82 lines (66 loc) · 4.23 KB

HACKATHON.md

File metadata and controls

82 lines (66 loc) · 4.23 KB
Raw

Hackathon submission — ThreadKeeper

Event: BGI Sprint I Track: Improvements to OmegaClaw Team: ThreadKeeper Repository: this repo — a fork of asi-alliance/OmegaClaw-Core, branched from a recent upstream main (June 2026) and extended.

Submission summary

ThreadKeeper is a configurable hybrid OmegaClaw architecture for open, decentralized, budget-aware agent orchestration. It decouples reasoning quality from reasoning frequency:

  • a cheap, persistent local control loop holds the thread (goal tracking, memory, continuity, and the escalation decision);
  • a local worker loop iterates cheaply;
  • cloud specialists are invoked only for hard subproblems, via a governed delegate skill;
  • an optional adjudicator breaks ties / gates high-stakes actions.

A configurable token budget drives the escalation decision, and every LLM call and escalation decision is logged — an ISO/IEC 42001-friendly audit trail. The result is lower cost, better persistence, and no single-model dependency.

It is framed as an extension to OmegaClaw-Core, not a correction: OmegaClaw's persistent MeTTa loop is the ideal foundation for the "thread keeper," and a deployment that ignores ThreadKeeper's additions runs exactly as before.

What this submission adds on top of OmegaClaw-Core

Addition Files Purpose
Subagent dispatch (delegate skill) src/subagent.py, src/skills.metta The "cloud specialist" node — bounded, governed delegation to a right-sized model.
Cost-awareness seam src/threadkeeper_budget.py Per-loop token accounting; supplies live facts to the policy and executes its verdict.
Escalation policy in MeTTa src/escalation.metta, tests/test_escalation_metta_parity.py The routing decision as Atomspace rules (tk-escalate), evaluated through PeTTa — symbolic, auditable, agent-rewritable. Python fallback proven equivalent by a parity test.
Configuration surface threadkeeper.config.yaml The four-node mesh + budget threshold declared in one place.
Architecture docs + diagram docs/architecture.md, docs/architecture.png The concept, the node responsibilities, the escalation logic.
Subagent reference + tutorial docs/reference-skills-subagent.md, docs/tutorial-09-subagents.md How to use the delegate skill end to end.
Persona scaffolding memory/personas-subagent/ Example specialist persona config (env-var keys only — no secrets).

Where to start reading

  1. README.md — elevator pitch, the problem, the solution, the quickstart.
  2. docs/architecture.md — the four-node mesh and escalation-trigger logic.
  3. threadkeeper.config.yaml — the configurable surface (models are examples; all roles swappable).
  4. src/threadkeeper_budget.py — the cost-awareness / escalation seam.
  5. src/subagent.py — the delegation primitive.

Constraints honored

  • No secrets in the repo. All credentials are referenced by the name of an environment variable; no key material is committed. The persona-config directory gitignores real configs and keeps only the .example.
  • Extension, not critique. ThreadKeeper changes none of OmegaClaw's existing behavior and is offered in the spirit of the OmegaClaw / SingularityNET vision.
  • Model-agnostic. No claim that any model is "better." The architecture works regardless of which models fill each role.

Status & notes

  • The base is a recent asi-alliance/OmegaClaw-Core main (branched June 2026); ThreadKeeper's additions are layered on top.
  • The subagent dispatch primitive and the budget module are working code (import-checked and functionally exercised), not vaporware; the escalation policy is a deliberate v1 designed as a swappable seam.
  • The ~4.7× input-token figure cited in the README came from earlier local testing of subagent dispatch against a 30B local model on tool-heavy workloads. It is illustrative — real numbers depend entirely on the models and workload — not a benchmark claim.

Submission editable until 28 Jun 2026.