LongHorizon-Harness — TON14 fork

This is my actively maintained fork of AMAP-ML/LongHorizon-Harness, a role-orchestrated harness (Manager → Executor → Auditor) that turns agent CLIs into long-horizon computer-use systems. Upstream went quiet after 2026-08-20 (v0.1.7), so the fork carries the project forward on its own line.

The complete original project documentation — install walkthroughs, the loop engineering story, benchmarks, configuration reference — is preserved verbatim from the v0.1.7 tag: see docs/upstream/README.md (English) and docs/upstream/README.zh-CN.md (中文). Everything below is specific to this fork.

What the fork adds

• ZCode agent backend (lhht run --agent zcode): drives the headless runtime bundled with the ZCode desktop app on GLM models (glm-5.3 default, glm-5.3-flash), with role-scoped permission modes (plan for the manager and auditors, yolo for executors). The reasoning effort (low/high/max) rides in an isolated per-run copy of ZCode's session database plus a .zcode/config.json provider declaration written into the workspace — the operator's ~/.zcode is never touched.
• Reasoning effort for every backend: on top of upstream's reasoning_effort chain (Codex, Claude Code, OpenCode), the DeepSeek Harness backend forwards the level through an llm-deepseek patch layer, and ZCode through its session store. Per-role, verbatim, no cross-backend mapping.
• Bilingual control tokens: the manager/auditor parsers accept route lines and control markers in both English and Russian.
• A modern toolchain floor: Python ≥ 3.14 (tomli fallbacks removed), Node.js ≥ 22 with CI provisioning 24; make check is the pre-push gate.
• Honest CI: tests.yml runs the Python suite on ubuntu/windows × 3.14 plus a web job (frontend core suite + typecheck) on every push and pull request; release.yml builds and verifies distributions without publishing (the PyPI name belongs to upstream).

Installing

The PyPI package lh-harness is upstream's release and does not contain the work above — install from this repository:

pip install lhht            # or: uv tool install lhht

Naming note: the fork owns its identity end to end — the PyPI distribution, the console command, the import module (import lhht), the LHHT_* environment variables, and the .lhht/ state directory are all lhht. The original lh-harness installs alongside it cleanly, so both can drive the same project folder for comparison.

Building from source (what the development loop uses):

git clone https://github.com/TON14/LongHorizon-Harness.git
cd LongHorizon-Harness
npm run build --prefix frontend/web   # the Web workbench bundle
uv tool install --force .   # installs the `lhht` command

Developing

make help        # all targets
make check       # both suites + typecheck, everything expected green
make dev-api     # control API on 127.0.0.1:8799
make dev-web     # Vite dev server on :5173, proxying /api

The tool installs as lhht (import module and the .lhht/ state directory keep their historical names). The harness is also used to develop itself: a repo-root .lhht/config.toml defines the roles (manager/executor on glm-5.3-flash, auditor on glm-5.3, effort max), and tasks run with lhht run --task @task.md. Agents edit the working tree; the operator reviews, runs make check, and commits.

Credits

The foundation is the AMAP-ML team's work — every contributor is listed on the contributors page. The fork's line is maintained by TON14.