LongHorizonOS

KO | EN

gitlite — search

#tensorflow #docker #csharp #html #ai-agent #pytorch #json #database #large-language-models #java

(★ 9)

No description available.

파일 탐색기

ZIP 다운로드

- - bug.md
  - core-semantic.md
  - feature.md
- - ci.yml
  - vpg-stress.yml
- PULL_REQUEST_TEMPLATE.md
- - baseline-fairness-audit.md
  - capability-manifest.json
  - coverage-summary.json
  - coverage.json
  - existing-tags.txt
  - git-commit.txt
  - git-status.txt
  - milestone1-final-report.md
  - mutation-audit.md
  - mypy-resolution.md
  - mypy.txt
  - pip-freeze.txt
  - pytest-cov.txt
  - pytest-quick.txt
  - python-version.txt
  - ruff.txt
  - sigkill-recovery-audit.md
  - sigkill-results.json
  - stub-leakage-audit.md
  - test-report.md
  - vertical_slice_v2_stop_gate.md
- - multi-agent-runtime.json
  - vpg-incremental-history-2026-08-12-frontier-summary-final.json
- .gitkeep
- adaptive-wallclock-runtime-20260816.json
- resource-aware-runtime-20260815.json
- - banner.svg
  - build_brand.py
  - icon-dark.svg
  - icon.svg
  - logo-horizontal-on-dark.svg
  - logo-horizontal-on-light.svg
  - logo-stacked-on-dark.svg
  - logo-stacked-on-light.svg
  - mark-mono-black.svg
  - mark-mono-white.svg
  - mark-on-dark.svg
  - mark-on-light.svg
  - rasterize.js
  - README.md
- - lhos-architecture.svg
  - lhtb-46-rewards.png
  - lhtb-46-tokens.png
- - xhs-01-cover.png
  - xhs-01-cover.svg
  - xhs-02-kpi.png
  - xhs-02-kpi.svg
  - xhs-03-mechanisms.png
  - xhs-03-mechanisms.svg
  - xhs-04-roadmap-cta.png
  - xhs-04-roadmap-cta.svg
- - oracle.json
  - README.md
  - sensenova-pi-ai-long.cordis.patch.yml
  - sensenova-pi-ai.cordis.patch.yml
  - stepfun-3.7-pi-ai.cordis.patch.yml
  - stepfun-pi-ai.cordis.patch.yml
- - lhtb_dsh_langchain_baseline_gate.yaml
  - lhtb_dsh_langchain_lhos_gate.yaml
- - cost_aware.yaml
  - fifo.yaml
- default.yaml
- development.yaml
- lhtb-controlled-queue-20260822.json
- lhtb-controlled-queue-r4-20260822.json
- lhtb-extended-natural-86400-queue-20260823.json
- lhtb-invalid9-rerun-queue-20260823.json
- - CORE-V1-FREEZE.md
  - LONGHORIZONOS-CORE-V1.md
  - README.md
- - ADAPTIVE-RUNTIME.md
  - ASYNC-AGENTOS.md
  - CASE-PROFILING-MANUAL.zh-CN.md
  - HARNESS-ADAPTIVE.md
  - HIDDEN-PROVENANCE.md
  - LHTB-EXTERNAL-CHECKPOINTS.zh-CN.md
  - LHTB-SOFTWARE5-PAIR.md
  - LHTB-STEPFUN-SMOKE.zh-CN.md
  - ONLINE-COMPUTE-CONTROL.md
  - REAL-DSH-DYNAMIC-CODING-PILOT.zh-CN.md
  - REAL-DSH-DYNAMIC-STEPFUN37-RESULT.zh-CN.md
  - REAL-MODEL-END-TO-END.md
  - SEMANTIC-REPAIR.md
  - SWE-BENCH-DOCKER-VALIDATION.zh-CN.md
  - SWE-BENCH-HOST-NATIVE-PILOT.zh-CN.md
  - SWE-BENCH-STEPFUN-RUNBOOK.zh-CN.md
  - UNIFIED-COMPUTE-CONTROL.md
  - VPG-INCREMENTAL-HISTORY.md
- - PROVENANCE-REPAIR.md
  - RECOVERY-REPAIR.md
- - DEEPSEEK-HARNESS-ADAPTER.md
  - HARNESS-OPTIMIZATION.md
- - v0.1.0.md
- - API-SURFACE.md
  - PUBLIC-API.md
- ACTIVE-STOPLOSS-20260831.md
- ADAPTIVE-WALLCLOCK-RUNTIME-BENCHMARK.md
- ARCHITECTURE-PATHS.md
- AUTOMATIC-REBASE.md
- COMPUTE-BUDGET.md
- CONCEPTS.md
- CONFLICT-GRAPH.md
- CTX-MANAGEMENT-FIX-20260830.md
- EVENT-DRIVEN-SUPERVISOR.md
- FRONTIER-POLICY.md
- HARNESS-SESSION-PROTOCOL.md
- IMPLEMENTATION-PROGRESS-2026-08-14.md
- IMPLEMENTATION-PROGRESS-2026-08-15-1300.md
- IMPLEMENTATION-PROGRESS-2026-08-15-1310.md
- IMPLEMENTATION-PROGRESS-2026-08-15-1420.md
- IMPLEMENTATION-PROGRESS-2026-08-15-1450.md
- IMPLEMENTATION-PROGRESS-2026-08-15-1525.md
- IMPLEMENTATION-PROGRESS-2026-08-15-1611.md
- IMPLEMENTATION-PROGRESS-2026-08-15-1810.md
- IMPLEMENTATION-PROGRESS-2026-08-15-1820.md
- IMPLEMENTATION-PROGRESS-2026-08-15-1835.md
- IMPLEMENTATION-PROGRESS-2026-08-15-1851.md
- IMPLEMENTATION-PROGRESS-2026-08-15-1915.md
- IMPLEMENTATION-PROGRESS-2026-08-15-2010.md
- IMPLEMENTATION-PROGRESS-2026-08-15-2017.md
- IMPLEMENTATION-PROGRESS-2026-08-15-2027.md
- IMPLEMENTATION-PROGRESS-2026-08-15-2141.md
- IMPLEMENTATION-PROGRESS-2026-08-15-2229.md
- IMPLEMENTATION-PROGRESS-2026-08-15-2307.md
- IMPLEMENTATION-PROGRESS-2026-08-15-2315.md
- IMPLEMENTATION-PROGRESS-2026-08-15-2355.md
- IMPLEMENTATION-PROGRESS-2026-08-15.md
- IMPLEMENTATION-PROGRESS-2026-08-16-0827.md
- IMPLEMENTATION-PROGRESS-2026-08-16-1010.md
- IMPLEMENTATION-PROGRESS-2026-08-16-1023.md
- IMPLEMENTATION-PROGRESS-2026-08-16-1135.md
- IMPLEMENTATION-PROGRESS-2026-08-16-1223.md
- IMPLEMENTATION-PROGRESS-2026-08-16-1545.md
- IMPLEMENTATION-PROGRESS-2026-08-16-1958.md
- IMPLEMENTATION-PROGRESS-2026-08-16-2028.md
- IMPLEMENTATION-PROGRESS-2026-08-16-2108.md
- IMPLEMENTATION-PROGRESS-2026-08-16-2130.md
- IMPLEMENTATION-PROGRESS-2026-08-16-2209.md
- IMPLEMENTATION-PROGRESS-2026-08-16-2322.md
- IMPLEMENTATION-STATUS.md
- ISSUE-INVENTORY.md
- LHOS_FRAMEWORK_IMPROVEMENT_DESIGN_20260823.md
- LHTB_OFFICIAL_PROTOCOL.md
- LONG-HORIZON-COMPUTE-MANAGEMENT.md
- LONGHORIZONOS-ALGORITHM-OS-HARNESS-EVALUATION-AUDIT.zh-CN.md
- LONGHORIZONOS-OS-READINESS-HARNESS-ADAPTATION.zh-CN.md
- LONGHORIZONOS_REVIEW_AND_ROADMAP_2026-08-11.md
- MEDIATED-WORKSPACE-COMMIT-VALIDATION.md
- MIND-VLA-IMPLEMENTATION-MATRIX.md
- OWNERSHIP-HANDOFF-CONTRACT.md
- PROVENANCE-CONTRACT.md
- QUICKSTART.md
- RELEASE-CHECKLIST.md
- RESOURCE-AWARE-PARALLELISM.md
- RESOURCE-AWARE-RUNTIME-BENCHMARK.md
- RESOURCE-REPLANNING-E2E.md
- RESOURCE-TELEMETRY.md
- ROADMAP.md
- SEMANTIC-INTERRUPT.md
- USER-OPERATOR-MANUAL.md
- USER-OPERATOR-MANUAL.zh-CN.md
- WORKSPACE-WATCH-LOOP.md
- WORKSPACE-WATCHER.md
- - __init__.py
  - context_basic_load.py
  - context_budget_eviction.py
  - context_process_isolation.py
  - context_snapshot_restore.py
  - context_version_pinning.py
  - crash_recovery.py
  - multi_process_artifacts.py
  - optimistic_conflict.py
  - private_workspace.py
  - shared_readonly.py
- - canonical_recovery_repair.py
- - __init__.py
  - capacity.py
  - crash_reassignment.py
  - no_eligible_agent.py
  - parallel_ready.py
  - semantic_operational_separation.py
  - specialized_pipeline.py
- - hello_world.py
  - multi_agent.py
  - observation_repair.py
  - real_coding_task.py
  - repair.py
- - artifact_version_reopens_task.py
  - basic_graph_closure.py
  - deterministic_ready_frontier.py
  - evidence_backed_verification.py
  - optimistic_patch_conflict.py
  - sigkill_recovery_demo.py
- __init__.py
- harness_subprocess_preemption.py
- online_resource_supervisor.py
- resource_replanning_e2e.py
- workspace_watch_loop.py
- analyze_3seed_results.py
- analyze_milestone_2_3_costs.py
- analyze_parse_failures.py
- analyze_v3_p4_p8.py
- b4_content_hash_baseline.py
- benchmark_adaptive_runtime.py
- benchmark_multi_agent_runtime.py
- benchmark_provider_routing.py
- benchmark_vpg_incremental_history.py
- build_report.py
- compare_lhtb_control.py
- compose_dynamic_pair.py
- context_sigkill_recovery.py
- d21_cross_process_double_claim.py
- d21_demo_reverify.py
- d21_multi_task_stress.py
- d21_mutation_audit_v2.py
- d21_random_sm.py
- d21_sigkill_adversarial.py
- d31_concurrency_atomicity.py
- d31_cone_shapes_audit.py
- d31_determinism_sm.py
- d31_evidence_audit.py
- d31_mutation_audit_v2.py
- d31_over_under_invalidation.py
- d31_partial_verified_invalidation.py
- d31_projection_audit.py
- d31_repair_authority.py
- d31_sigkill_audit_v2.py
- d3_demos.py
- d3_heavy_audit.py
- d3_mutation_audit.py
- d3_sigkill_audit.py
- d3_sigkill_worker.py
- debug_e2e.py
- e5_scale_engine_vs_sdk.py
- e6_context_residency_motivation.py
- e6_prototype_sigma_scheduling.py
- export_swe_predictions.py
- export_trace.py
- gen_coverage_summary.py
- gen_d2_artifacts.py
- gen_token_accounting.py
- harbor_job_worker.py
- i5_projection_rebuild_probe.py
- i6_rebuild_semantic_preservation.py
- init_db.py
- lhtb_dsh_harbor_agent.py
- lhtb_dsh_phase_bridge.py
- lhtb_external_checkpoint.py
- lhtb_semantic_policy_v3_candidate.py
- lhtb_snapshot_restore_agent.py
- m11_undeclared_dependency_ratio.py
- prepare_lhtb_budget_sweep.py
- prepare_lhtb_local_pilot.py
- profile_lhtb_trial.py
- profile_real_dsh_dynamic_case.py
- profile_swe_case.py
- profiling_report.py
- recollect_v3_artifacts.py
- reconstruct_n3_timeline.py
- replay_semantic_policy_v2.py
- run_3seed_pairing.py
- run_controlled.py
- run_deepseek_adapter_smoke.py
- run_deepseek_local_provider_smoke.py
- run_dynamic_lhos_only.py
- run_lhtb_background.ps1
- run_lhtb_controlled_queue.py
- run_lhtb_lhos_after_baseline.ps1
- run_lhtb_lhos_batch.py
- run_lhtb_lhos_control.py
- run_lhtb_software5_pair.py
- run_multi_run_stress.py
- run_phase_b_demos.py
- run_stuck_recovery_debug_v3.py
- run_swebench_stepfun_pytest11143.ps1
- run_vertical_slice.py
- sigkill_recovery_test.py
- summarize_dsh_paired_experiment.py
- summarize_lhtb_budget_sweep.py
- summarize_lhtb_prefix.py
- summarize_swe_lite_pilot.py
- update_v3_summary.py
- - __init__.py
  - bootstrap.py
  - config.py
  - subprocess_policy.py
- crash_task.json
- example_task.json
- - __init__.py
  - test_action_fencing_races.py
  - test_action_resource_capabilities.py
  - test_action_state_machine.py
  - test_architecture.py
  - test_audit_action_mutation.py
  - test_audit_blocked_polling.py
  - test_audit_capability_bypass.py
  - test_audit_deadlock.py
  - test_audit_journal_atomicity.py
  - test_audit_journal_rebuild.py
  - test_audit_lease_invariants.py
  - test_audit_process_mutation.py
  - test_audit_sigkill.py
  - test_audit_starvation.py
  - test_capabilities.py
  - test_deadlock_determinism.py
  - test_deadlocks.py
  - test_demos.py
  - test_dispatcher_ownership.py
  - test_isolation.py
  - test_journal.py
  - test_kernel_loop.py
  - test_lease_renew_atomicity.py
  - test_leases.py
  - test_process_lease_lifecycle_race.py
  - test_process_state_machine.py
  - test_process_step_caller_identity.py
  - test_recovery.py
  - test_recovery_policy.py
  - test_resource_admission_safety.py
  - test_signals.py
  - test_sqlite_storage_isolation.py
  - test_sqlite_storage_migrations.py
  - test_strict_effect_contracts.py
  - test_transactional_outbox.py
- - __init__.py
  - test_adaptive_control.py
  - test_adaptive_runtime.py
  - test_adaptive_wallclock_runtime.py
  - test_baseline_vs_lhos.py
  - test_benchmarks.py
  - test_computation_utility.py
  - test_compute_budget.py
  - test_dsh_dynamic_coding.py
  - test_harness_adaptive.py
  - test_hidden_provenance.py
  - test_live_model.py
  - test_parallelism_regimes.py
  - test_preemption_payoff.py
  - test_provider_routing.py
  - test_real_build_workload.py
  - test_resource_aware_runtime.py
  - test_scheduling_regimes.py
  - test_swe_host_native.py
  - test_unified_control.py
- - __init__.py
  - test_adaptive_wallclock_runtime.py
  - test_cli.py
  - test_compute_budget.py
  - test_harness_adaptive.py
  - test_online_compute_provider.py
  - test_resource_aware_runtime.py
  - test_unified_control.py
  - test_vpg_lifecycle_cli.py
- - __init__.py
  - test_demo.py
  - test_online_supervisor.py
  - test_provenance_repair.py
- - __init__.py
  - test_tiny_repository_task.py
- - __init__.py
  - test_attempt_semantics.py
  - test_benchmark_run.py
  - test_checkpoint_restore.py
  - test_crash_injection.py
  - test_event_projection.py
  - test_invalidation_loop.py
  - test_llm_logging.py
  - test_migration_compatibility.py
  - test_milestone_2_2_integration.py
  - test_multi_run_isolation.py
  - test_run_resume.py
  - test_verification_commit.py
  - test_worker_tool_loop.py
- - __init__.py
- - test_harbor_job_worker.py
  - test_lhtb_budget_sweep.py
  - test_lhtb_dsh_harbor_agent.py
  - test_lhtb_external_checkpoint.py
  - test_lhtb_semantic_policy_v3_candidate.py
  - test_run_lhtb_controlled_queue.py
  - test_run_lhtb_software5_pair.py
  - test_summarize_lhtb_prefix.py
- - __init__.py
  - test_access_correction.py
  - test_adaptive_run.py
  - test_architecture.py
  - test_async_measured_span.py
  - test_async_run.py
  - test_automatic_provenance_boundaries.py
  - test_automatic_rebase.py
  - test_child_resource_usage.py
  - test_computation_control.py
  - test_compute_budget.py
  - test_compute_budget_audit_edges.py
  - test_compute_budget_facade.py
  - test_compute_budget_run.py
  - test_compute_calibration.py
  - test_compute_routing.py
  - test_compute_usage.py
  - test_conflict_derivation_equivalence.py
  - test_conflict_graph.py
  - test_conflict_query_index.py
  - test_context_budget.py
  - test_context_delta.py
  - test_context_hot_potato.py
  - test_context_utilization.py
  - test_context_vm_integration.py
  - test_dsh_phase_bridge.py
  - test_effect_gateway_integration.py
  - test_epoch_controller.py
  - test_error_isolation.py
  - test_event_supervisor.py
  - test_execution_closure.py
  - test_frontier_policy.py
  - test_graph_analysis.py
  - test_graph_analysis_incremental.py
  - test_handoff_transaction.py
  - test_harness_control_bridge.py
  - test_harness_dispatcher_fencing.py
  - test_harness_protocol.py
  - test_harness_usage_accounting.py
  - test_host_capacity.py
  - test_kernel_loop.py
  - test_large_goal_atomic_compile.py
  - test_live_context_rebase_e2e.py
  - test_live_context_rebase_fences.py
  - test_model_tier_measured.py
  - test_model_tier_routing.py
  - test_online_epoch.py
  - test_online_epoch_cleanup.py
  - test_online_epoch_execution.py
  - test_online_epoch_harness_handoff.py
  - test_online_epochs.py
  - test_online_resource_supervisor_e2e.py
  - test_outcome_calibration.py
  - test_parallelism_policy.py
  - test_policy_facade.py
  - test_preempt_superseded.py
  - test_provenance_append_scaling.py
  - test_provenance_integration.py
  - test_provider_routing.py
  - test_provider_routing_benchmark.py
  - test_rebase_public_exports.py
  - test_rebase_runtime.py
  - test_resource_aware_run.py
  - test_resource_configuration.py
  - test_resource_policy.py
  - test_resource_replanning_e2e.py
  - test_resource_telemetry.py
  - test_runtime_state.py
  - test_scheduling_epoch_persistence.py
  - test_sdk.py
  - test_semantic_interrupt.py
  - test_semantic_interrupt_persistence.py
  - test_semantic_interrupt_runtime.py
  - test_streaming_dispatch.py
  - test_subprocess_agent.py
  - test_subprocess_harness.py
  - test_undeclared_reads.py
  - test_unified_policy.py
  - test_unified_policy_run.py
  - test_verification_routing.py
  - test_waste_projection.py
  - test_waste_wrong_direction.py
  - test_workspace_watch_loop.py
- - __init__.py
  - test_architecture.py
  - test_benchmark_metrics.py
  - test_capability_manifest.py
  - test_command_verifier_policy.py
  - test_context_compiler.py
  - test_effect_gateway.py
  - test_external_grader.py
  - test_failure_codes.py
  - test_file_verifier_n3_regression.py
  - test_generator.py
  - test_invalidation.py
  - test_milestone_2_2.py
  - test_minimal_lhos_mode.py
  - test_oracle_isolation.py
  - test_patch_validator.py
  - test_provenance.py
  - test_readiness.py
  - test_scheduler.py
  - test_sensenova_parsing.py
  - test_state_machine.py
  - test_tool_name_contract.py
- __init__.py
- conftest.py
.env.example
.gitattributes
.gitignore
CHANGELOG.md
CODE_OF_CONDUCT.md
CONTRIBUTING.md
LICENSE
Makefile
MANIFEST.in
NOTICE
pyproject.toml
README.md
README.zh-CN.md
SECURITY.md
uv.lock

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

### The operating system for long-running agents

**Your harness keeps the agent alive.<br/>LongHorizonOS keeps it honest.**

<br/>

[![Python](https://img.shields.io/badge/Python-3.11%2B-3776AB?style=for-the-badge&logo=python&logoColor=white)](https://www.python.org/)
[![License](https://img.shields.io/badge/License-Apache--2.0-D22128?style=for-the-badge)](https://raw.githubusercontent.com/Yang-Jiashu/LongHorizonOS/main/LICENSE)
[![Status](https://img.shields.io/badge/status-research%20alpha-yellow?style=for-the-badge)](https://raw.githubusercontent.com/Yang-Jiashu/LongHorizonOS/main/docs/releases/v0.1.0.md)

[![GitHub Stars](https://img.shields.io/github/stars/Yang-Jiashu/LongHorizonOS?style=for-the-badge&logo=github&color=gold)](https://github.com/Yang-Jiashu/LongHorizonOS/stargazers)
[![GitHub Forks](https://img.shields.io/github/forks/Yang-Jiashu/LongHorizonOS?style=for-the-badge&logo=github&color=forestgreen)](https://github.com/Yang-Jiashu/LongHorizonOS/network/members)
[![GitHub Issues](https://img.shields.io/github/issues/Yang-Jiashu/LongHorizonOS?style=for-the-badge&logo=github&color=orange)](https://github.com/Yang-Jiashu/LongHorizonOS/issues)
[![Last Commit](https://img.shields.io/github/last-commit/Yang-Jiashu/LongHorizonOS?style=for-the-badge&logo=github&color=blueviolet)](https://github.com/Yang-Jiashu/LongHorizonOS/commits/main)

[![Benchmark](https://img.shields.io/badge/LHTB-46%20task%20paired-4A7DBF?style=for-the-badge)](https://raw.githubusercontent.com/Yang-Jiashu/LongHorizonOS/main/)
[![Top Language](https://img.shields.io/github/languages/top/Yang-Jiashu/LongHorizonOS?style=for-the-badge&color=3776AB)](https://github.com/Yang-Jiashu/LongHorizonOS)
[![Code Size](https://img.shields.io/github/languages/code-size/Yang-Jiashu/LongHorizonOS?style=for-the-badge&color=teal)](https://github.com/Yang-Jiashu/LongHorizonOS)
[![Contributors](https://img.shields.io/github/contributors/Yang-Jiashu/LongHorizonOS?style=for-the-badge&color=crimson)](https://github.com/Yang-Jiashu/LongHorizonOS/graphs/contributors)

<br/>

English | [简体中文](README.zh-CN.md)

</div>

---

## 📑 Table of Contents

- [The failure you already know](#the-failure-you-already-know)
- [The idea: an exoskeleton, not a transplant](#the-idea-an-exoskeleton-not-a-transplant)
- [✨ Key Mechanisms](#-key-mechanisms)
- [🏗️ What ships in this release](#️-what-ships-in-this-release)
- [📊 Evaluation: 46-task LongHorizonBenchmark (LHTB), paired arms](#%EF%B8%8F-evaluation-46-task-longhorizonbenchmark-lhtb-paired-arms)
- [Why not just \[prompt harder / bigger context / native compaction\]?](#why-not-just-prompt-harder--bigger-context--native-compaction)
- [🚀 Quick Start](#-quick-start)
- [🗺️ Roadmap](#️-roadmap)
- [📚 Documentation](#-documentation)
- [🧑‍💻 Development](#-development)
- [Contributing · Security · License](#contributing--security--license)
- [Acknowledgments](#acknowledgments)

---

## The failure you already know

Long-horizon agents don't crash. They do something worse: **they keep going**.

Real traces from our benchmark fleet, before the OS layer existed:

> **`apex-openroad`** — 158 slice-resumes over 3.5 hours. Every resume re-read its entire conversation; by the end the agent was paying **183,000 tokens per model call**. 45.9M tokens burned, reward went 0.292 → 0. Nobody stopped it. Nothing even noticed.

> **`super-mario`** — 120M tokens. The agent never cleared the first level. It was "working" the entire time.

> **`audio-visual`** — a 90-minute session in which **54% of the slices did literally nothing**: zero tool calls, zero progress, while the verifier kept asking "done yet?" and the harness kept answering "keep going."

Every long-running agent stack has these sessions. They are not edge cases; they are the tail that eats the budget. The harness cannot fix them, because the harness *is* the thing that is running. You need a layer outside it.

---

## The idea: an exoskeleton, not a transplant

LongHorizonOS wraps an existing agent harness and treats every run as a **process under an operating system**:

Same model. Same harness. Same task. The only difference is that **someone is watching the meter**.

---

## ✨ Key Mechanisms

OS analogies are intentional. Every gate is fail-closed: ambiguous telemetry means *keep running*.

| | OS Concept | LongHorizonOS Implementation |
|---|---|---|
| ⏱️ | **Time Quantum** | Bounded execution slices — a drifting run is never more than one slice away from intervention |
| 🔄 | **Context Switch** | Restart with compacted semantic handoff — goals, evidence, artifact URIs forward; the 100K-token transcript stays behind |
| 💡 | **Investment, Not Alarm** | Restarts must be *earned*: fires only if handoff carries state and the previous restart demonstrably paid for itself |
| 🧨 | **OOM Killer** | Rate-limited severe valve — cache-per-call ≥ 4× the soft line forces a restart; never churns |
| 🐕 | **Watchdog** | Stop-loss halt — 8M tokens of burn with zero writes, zero tests, and frozen verifier → stop. Any life resets the clock |
| 📑 | **Page Table** | Versioned progress graph (goals / artifacts / evidence / validity). Mid-run changes invalidate only the affected subgraph. Verified work is never redone. |

</div>

The reward red line is structural, not aspirational.

---

## 🏗️ What ships in this release

The artifact core is a crash-consistent, content-addressed store wrapped in a capability model:

| Capability | What it means |
|---|---|
| 📝 **Process / Action / Journal** | Every mutation is a journaled, atomic event; a process may act only through explicitly spawned actions |
| 🔐 **Capability / Lease / Signal** | Resource access is capability-gated, ownership is lease-based, control flows through typed signals |
| 💥 **Crash Recovery** | SIGKILL-resilient with exactly-once semantics; replaying the journal reconstructs the exact pre-crash state |
| 🗃️ **Versioned Artifact FS** | Artifacts are immutable, content-addressed, and versioned; reads & writes go through an atomic write protocol |
| 🧱 **Namespace Isolation** | Each process namespace is isolated; cross-namespace access requires an explicit capability |
| ✅ **Version-checked Commits** | Optimistic concurrency via expected versions — stale writes are refused instead of silently lost |
| 🔗 **Canonical URI Security** | Artifact URIs are canonicalized and path traversal is rejected at the boundary |

</div>

### Not yet implemented

- **Distributed multi-agent cluster** — the control plane is single-host today.
- **General belief revision** — the progress graph is versioned but does not yet retract prior conclusions in general.

---

## 📊 Evaluation: 46-task LongHorizonBenchmark (LHTB), paired arms

Each of the 46 **LHTB** (**LongHorizonBenchmark**) tasks (real software-engineering and data-science workloads, machine-checked verifiers) runs **twice** — identical model (`stepfun/step-3.7-flash`), task binary, budget, and verifier:

- **fresh** — the harness alone
- **lhos** — the same run under LongHorizonOS

### ── Core KPIs ──

# <span style="font-size:2.4em;color:#2ea043">37 / 42</span>

### reward parity-or-better

<br/>

# <span style="font-size:2.4em;color:#4A7DBF">22 / 27 (81%)</span>

### where either arm scored

<br/>

# <span style="font-size:2.4em;color:#8957e5">~2.4×</span>

### fewer input tokens

<br/>

# <span style="font-size:2.4em;color:#e6850a">0</span>

### harness changes required

<br/>

</div>

**What you are looking at:** left — every task's verifier reward, both arms; red marks the five dropoffs we disclose below; grey marks the three tasks with no valid lhos result. The denominator, spelled out: 46 tasks − 3 without a valid lhos result − 1 invalid pairing (matpower, see †) = 42; fifteen of the 42 are 0–0 pairs (both arms scored zero), which count as parity but carry no signal — the 22/27 rate excludes them. Right — input tokens per task on log axes; below the diagonal means the OS layer spent less. `commit0` is the visible counterexample above the line — we disclose it rather than crop it.

<details>
<summary><b>Full 46-task paired results table</b> (single-seed rewards; ±0.01 tolerance)</summary>

| # | task | fresh | lhos | Δ | note |
|---|---|---:|---:|---:|---|
| 1 | grammar-fuzz | 0.939 | 0.828 | -0.111 | **dropoff** |
| 2 | poc-exploit | 0.892 | 0.892 | +0.000 | parity |
| 3 | spot | 0.855 | 0.909 | +0.054 | improved |
| 4 | spice | 0.606 | 0.636 | +0.030 | improved |
| 5 | foldseek | 0.333 | 0.333 | +0.000 | parity |
| 6 | alp | 0.300 | 0.200 | -0.100 | **dropoff** |
| 7 | apex-openroad | 0.292 | 0.000 | -0.292 | **dropoff** |
| 8 | great-expectations | 0.273 | 0.273 | -0.000 | parity |
| 9 | satellite | 0.200 | 0.200 | +0.000 | parity |
| 10 | su2 | 0.200 | 0.200 | +0.000 | parity |
| 11 | unison | 0.167 | 0.167 | -0.000 | parity |
| 12 | apexmgmt | 0.160 | 0.118 | -0.042 | **dropoff** |
| 13 | modflow6 | 0.092 | 0.000 | -0.092 | **dropoff** |
| 14 | climate | 0.084 | 0.086 | +0.002 | parity |
| 15 | audio-visual | 0.068 | 0.066 | -0.002 | parity |
| 16 | materials | 0.067 | 0.064 | -0.004 | parity |
| 17 | rush_hour | 0.040 | 0.040 | +0.000 | parity |
| 18 | scientific-figure | 0.040 | 0.040 | +0.000 | parity |
| 19 | robotics-slam | 0.029 | 0.029 | +0.000 | parity |
| 20 | microscopy | 0.017 | 0.121 | +0.104 | improved |
| 21 | apex-ib244 | 0.014 | 0.144 | +0.130 | improved |
| 22 | document-table | 0.012 | 0.012 | +0.000 | parity |
| 23 | nrel | 0.005 | 0.000 | -0.005 | parity |
| 24 | epa-swmm | 0.002 | 0.031 | +0.029 | improved |
| 25 | opensees | 0.001 | 0.000 | -0.001 | parity |
| 26 | matpower | 0.000 | 0.583 | +0.583 | invalid pairing † |
| 27 | generals-bot | 0.000 | 0.430 | +0.430 | improved |
| 28 | sudoku | 0.000 | 0.027 | +0.027 | improved |
| 29 | chess-mate | 0.000 | 0.000 | +0.000 | parity |
| 30 | nbody | 0.000 | 0.000 | +0.000 | parity |
| 31 | langchain | 0.000 | 0.000 | +0.000 | parity |
| 32 | 2048 | 0.000 | 0.000 | +0.000 | parity |
| 33 | apex-investment | 0.000 | 0.000 | +0.000 | parity |
| 34 | unknown-config | 0.000 | 0.000 | +0.000 | parity |
| 35 | tabular | 0.000 | 0.000 | +0.000 | parity |
| 36 | gdal | 0.000 | 0.000 | +0.000 | parity |
| 37 | snake_maze | 0.000 | 0.000 | +0.000 | parity |
| 38 | super-mario | 0.000 | 0.000 | +0.000 | parity |
| 39 | dicom | 0.000 | 0.000 | +0.000 | parity |
| 40 | epidemic | 0.000 | 0.000 | +0.000 | parity |
| 41 | commit0 | 0.000 | 0.000 | +0.000 | parity |
| 42 | sokoban | 0.000 | 0.000 | +0.000 | parity |
| 43 | duckdb | 0.000 | 0.000 | +0.000 | parity |
| 44 | apex-law433 | 0.452 | — | — | no valid lhos result |
| 45 | vector-db | 0.300 | — | — | no valid lhos result |
| 46 | riscv | 1.000 | — | — | no valid lhos result |

† fresh arm timed out in this pairing; that task's historical fresh value is 1.0 — treat the row as invalid, not as a gain.

</details>

**The honest part, because it matters:** 5 of the 42 valid tasks score *lower* under the OS layer. We know why, per case: restart-sensitive recovery on some, a context-bloat limitation on two long sessions. 3 tasks produced no valid result (provider censorship, budget hard-stop). The current policy generation (v2: investment-gated restarts, severe valve, stop-loss halt) targets exactly these; the research log tracks whether they move. If they don't move, this section will keep saying so.

---

## Why not just [prompt harder / bigger context / native compaction]?

- **Prompting** asks the agent to police itself. The agent is the thing that is drifting.
- **Bigger context** makes the 183K-tokens-per-call problem *more expensive*, not less.
- **Native compaction** happens inside the session, blind to budget, verifier signal, and whether the last three restarts paid off. The OS sees all of it — it is the only layer that never gets compacted.

---

## 🚀 Quick Start

```bash
python -m pip install -e ".[dev]"

# No API key needed. Build a goal, crash mid-run, recover, change an
# upstream file — and watch only the affected subgraph re-verify:
lhos demo recovery-repair --json
```

---

## 🗺️ Roadmap

| Phase | Status | What's inside |
|---|---|---|
| **Now** | 🟢 Shipped | Research alpha · single-host · DeepSeek Harness integration · 46-task paired **LongHorizonBenchmark (LHTB)** evaluation |
| **Next** | 🟡 In focus | Harness-agnostic capability matrix · partial in-session compaction · richer policy telemetry |
| **The Point** | 🔭 Horizon | A training substrate — continuation decisions logged, replayable, and policy-addressable. Schedule policies (and eventually learned models) trained & benchmarked *inside* the OS against real long-horizon workloads |

</div>

---

## 📚 Documentation

| | |
|---|---|
| **Using it** | [User & Operator Manual](docs/USER-OPERATOR-MANUAL.md) · [Quick Start](docs/QUICKSTART.md) · [Python API](docs/sdk/PUBLIC-API.md) · [Harness Protocol](docs/HARNESS-SESSION-PROTOCOL.md) |
| **Understanding it** | [Concepts](docs/CONCEPTS.md) · [Architecture](docs/architecture/LONGHORIZONOS-CORE-V1.md) · [Compute Management](docs/LONG-HORIZON-COMPUTE-MANAGEMENT.md) |
| **Status** | [Release Notes](docs/releases/v0.1.0.md) · [Implementation Status](docs/IMPLEMENTATION-STATUS.md) · [Roadmap](docs/ROADMAP.md) · [Changelog](CHANGELOG.md) |

---

## 🧑‍💻 Development

```bash
python -m pytest -q -m "not slow"
python -m ruff check src tests && python -m ruff format --check src tests examples scripts
python -m mypy src/lhos
```

---

## Contributing · Security · License

[CONTRIBUTING.md](CONTRIBUTING.md) · [SECURITY.md](SECURITY.md) · [Apache-2.0](LICENSE)

---

## Acknowledgments

**Benchmark & Harness foundation.** The 46-task evaluation in this release uses **LHTB (LongHorizonBenchmark)**, the long-horizon agent benchmark suite. All 46 agent sessions were executed on top of **DeepSeek Harness** (see `LHTBDeepSeekHarnessAgent` in `scripts/lhtb_dsh_harbor_agent.py`); the paired-experiment bridge and per-slice telemetry gate assume DeepSeek Harness as the baseline harness. LongHorizonOS is evaluated *against* this harness and benchmark, not on top of any other runtime.

**Model & inference.** All experiments were enabled by the [StepFun Builder Program](https://platform.stepfun.com/builder-program), which provided the model API and inference quota (model `stepfun/step-3.7-flash`, credential via the `STEPFUN_API_KEY` environment variable).

---

**[⬆ Back to Top](#longhorizonos)**

</div>

LongHorizonOS

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Download README