Skip to content

Latest commit

 

History

1 Commit

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

HALT — Hierarchical Automaton for Long-horizon Tasks

Version 1.0.2 · Status: Stable

A process specification for running long-horizon tasks with AI agents — interruptible at any point, persistently memorized, arbitrarily nestable.

A task is a directed graph of nodes driven by journaled transactional state machines; progress exists only where write→verify→commit has completed; any agent — new or old, human or machine — resumes by asking "where is my last provably stable position?" and recomputing what may legally run next.

What's here

Path Contents
00-index.md Entry point & document map
0111 The specification: core model, lifecycle FSMs, SOP interface, persistence layout, recovery, nesting, parallel dispatch, roles, human gates, memory store, conformance
12-validation-report.md Field-run validation: incident register, per-clause verdicts, amendments, fixture map
glossary.md Authoritative definitions
provenance.md Source mapping (papers / field practice) + Evidence Matrix
runtime/ v1.1 minimal reference runtime (single-node, stdlib-only Python) + 8 fixture tests — all passing
zh/ Chinese edition of the core document

Highlights

  • Field-validated: every mechanism carries an evidence class (paper / practice / design) and a field-test flag; the four v1.0.2 amendments (seq tail-derivation, path-checksum immutability, artifact-presence verification, audit recomputability) each trace to a real incident they now prevent.
  • Falsifiable: incidents from the validation run are replayable as automated fixtures against the reference runtime.
  • Executor-agnostic: the same spec has been executed by a human-coordinated LLM agent, by plain Python, and packaged as an agent skill.

Quick start

cd runtime
python test_halt_rt.py     # 8 passed
from halt_rt import HaltRuntime

rt = HaltRuntime("./my-task-ws", task_id="my-task")   # fresh workspace
# ... crash happens ...
rt = HaltRuntime.resume("./my-task-ws")               # rebuild from ledger

Reading paths

  • Just the idea: 00-index.md01-core-model.md
  • Implement it: 04-persistence-layout.mdruntime/README.md
  • Trust it: 12-validation-report.mdprovenance.md (Evidence Matrix)
  • 中文读者: zh/01-core-model.md

Status

Single-node scope is stable. Multi-node parallel dispatch, process-kill benchmark suite (HALT-Bench), and cross-task memory store are specified but not yet runtime-implemented.

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages