Harness

English · 한국어

A version-controlled personal workstation setup for coding agents: shared skills, T3 Code integration, guardrails, and guarded automation.

Harness configures an existing T3 Code workstation whose agents run as threads on Claude Code, Codex CLI, and Antigravity CLI (agy). It is a personal configuration and extension showcase, not a standalone agent, a hosted service, or a replacement for those tools. Herdr stays only as a plain terminal multiplexer for terminal work a person watches. Retired Herdr agent-runtime, Oh My Pi, and benchmark sources are frozen in archived/; nothing installs or runs them.

Installation · Documentation · Usage · Reuse policy · MIT License

What it adds

Capability	What you get
Shared skills	Repository-owned instructions linked into Claude Code, Codex, and AGY, with retained notices for adopted material.
UI design workflows	Four focused skills for interface design/review, image-to-web implementation, UI/brand reference-image generation, and optional design-system documentation; no mandatory image or document pipeline. See usage.
T3 Code path	t3-orchestrate, review-loop, and council delegate through T3's orchestration tools; t3-handoff continues the work in a new thread from a handoff document; T3 human-in-the-loop rules route questions through the thread; codex/review.ts runs an independent review in Codex's read-only sandbox outside T3 (inside T3 the Codex reviewer is a delegate_task child); t3/scheduled/guard.mjs adds timeouts and run limits to scheduled tasks.
Watched terminal work	t3-terminal opens Herdr workspaces, tabs, and panes for dev servers, interactive prompts, REPLs, ssh, and TUIs that a person must watch or step into. It is not a substitute for background processes.
Interactive visuals	t3-show builds explorable views inside the thread with html_render: codebase maps, algorithm steppers, assumption-driven simulations such as GTM metrics, and data exploration. Each page is grounded in code read or numbers computed in the shell, self-checks against them, and is driven through its interactions in html_preview before it is shown.
MR lifecycle and work modes	GitLab MR and GitHub PR skills adapted from tstack (file-mr, babysit-mr, leaving-mr-comments, takeover, modernize) and the user-selected full-send/full-stop modes. A merge-guard hook in the hstack plugin for Claude Code and AGY, and codex/harness.rules for Codex, block merges the mode does not allow; full-send merges only in repositories you list in an allowlist. See AGENT-CONFIG.
Personal operations	Cost audits over Claude Code and Codex sessions, and guarded host synchronization. These are not hosted services.

Recommended environment

The maintained showcase target is Linux or WSL2 with mise. mise.toml pins Node.js 24 and rulesync; every script runs on that Node, and Python 3 runs the cost-audit analyzer and skill collectors. Install and authenticate T3 Code, Claude Code, Codex CLI, and AGY separately. Native Windows and macOS are not maintained installation targets. Future agent CLI, Node, or rulesync releases must be reviewed rather than assumed compatible.

git clone https://github.com/huketo/harness.git
cd harness
mise trust && mise install
mise run plan

Review the installation guide, the source, and the proposed workstation changes before proceeding.

mise run install

Installation changes your workstation: it links shared skills and the Codex execpolicy file, backs up an eligible existing file before replacing it with a link, merges credential-read deny rules into the Claude Code and AGY settings files through rulesync without dropping any existing setting, and installs the hstack plugin into Claude Code (from this repository's marketplace, with auto-update) and AGY when they are on PATH. A foreign or unexpected entry remains a conflict and is not removed; mise run plan reports the plan without mutating anything. Running install again on an unchanged checkout changes nothing. Start new agent sessions afterwards so they load the current skills. AGY settings and T3 scheduled tasks are not restored by the installer.

On a shared workstation, node scripts/install.ts --skills-only --apply links only skills and leaves tool configuration unchanged. Keep the clone at a stable path; mise run update fast-forwards a clean main clone and reinstalls. Review updates first because installed links are live. See installation and update boundaries.

To take only the merge guard and read-only reviewer into Claude Code, without the skills or settings changes, install the plugin from this repository's marketplace in Claude Code: /plugin install hstack --marketplace huketo/harness.

Everyday use

Agents run as T3 Code threads. In a thread, t3-orchestrate splits independent work across worktree children, review-loop and council run read-only reviews by other model families, and the MR skills (file-mr, babysit-mr, takeover, modernize) carry GitLab MRs and GitHub PRs. Questions and approvals go through the thread under the human-in-the-loop rules.

Send terminal work a person must watch or step into through t3-terminal; attach with herdr in your own terminal to watch it or type. Run short reads, builds, tests, and background processes with normal tools.

Check the boundary of the Codex reviewer used outside T3 without calling a model:

node codex/review.ts --cwd <worktree> --inspect-tools

See detailed usage and state boundaries (Korean). Dated measurements live in FACTS and pricing evidence in PRICING; they are observations, not performance guarantees.

Configuration and safety

• One owner per asset: hand-maintained assets are symlinked; tool-managed state stays with its owning tool. Distributed snapshots are review examples, not host backup or restore instructions.
• Credentials stay local: keep API keys, OAuth state, private calendar URLs, messaging tokens, session logs, and report inputs outside version control. .gitignore is not a secret scanner and does not erase history.
• Review external effects: T3 schedule activation, messages, plugin updates, remote writes, merges, and Git publication require deliberate owner action. The merge guard is a guardrail against accidental merges, not an authorization system. No distributed snapshot should be activated unchanged.
• No sandbox claim: permission policies and file-tool deny rules do not isolate arbitrary shell access. Treat permissive workstation settings as personal risk choices, not security defaults.
• Live links matter: editing or updating this checkout changes installed assets immediately. There is no automated uninstaller; manual rollback boundaries explain what link removal does not undo.

Find your way around

Path	Purpose
skills/	Personal skills and adopted skills with retained notices.
plugin/	The hstack plugin for Claude Code and AGY: merge guard, read-only reviewer agent, AGY rules, builder, and plugin installer. The repository root is its marketplace.
scripts/ and .rulesync/	The installer, updater, and rulesync wrapper behind mise run, and the credential deny rules it applies.
codex/	Read-only Codex reviewer and the execpolicy rules that make Codex ask before merging.
t3/	Guard for T3 Code scheduled tasks and the guarded host sync; the distributed job list is empty.
agy/	An example AGY settings snapshot.
third-party/	External provenance, lock snapshots, and patches.
docs/	Installation, usage, ownership, evidence, and maintenance guidance.
archived/	Frozen retired sources (Herdr agent runtime, OMP, benchmark runner); nothing installs, runs, or tests them.
var/	Ignored local reports and generated state; only .gitkeep is tracked.

License, reuse, and maintenance

Harness is released under the MIT License. Third-party and adopted components retain their own copyright notices and license terms; adopted-skills.json records their provenance.

This repository is a personal showcase maintained only for its owner. Bug reports, feature requests, pull requests, other contributions, support requests, and vulnerability reports are not accepted. You may fork, modify, and redistribute the code under the MIT License, but no maintainer service or response is offered. See CONTRIBUTING.md for the complete policy.

Coding agents enter through AGENTS.md. The README's language switch, capability overview, quick start, and deeper-documentation layout take structural inspiration from oh-my-hermes; its branding, artwork, implementation, and product claims are not copied.