Harness Context

Use this skill when the user asks about Open Harness layout, lifecycle, conventions, git workflow, or where files live.

Steps

1. Read CLAUDE.md for the orchestrator contract — what the root-level agent does and does not do.
2. For layout questions, read the per-directory README.md files (e.g. .agro/README.md, .agro/tasks/README.md, .agro/scripts/README.md) plus the nested agent guides .worktrees/AGENTS.md, projects/AGENTS.md, and crons/AGENTS.md, and the How the system fits together section of AGENTS.md. There is no single comprehensive tree — use the filesystem and the directory READMEs.
3. For behavioral norms, now collapsed from an always-on tier into on-demand skills, the canonical homes are:
   • /git (.agro/skills/git/SKILL.md) — issue / branch / commit / PR conventions
   • /t3 (.agro/skills/t3/references/sandbox-processes.md) — tmux sessions for long-running processes
   • references/directory-readme.md — when a directory needs a README
   • /delegate (.agro/skills/delegate/SKILL.md) — wave decomposition and the recursion budget for sub-agents
4. For skill-listing or skill-source-of-truth questions, load references/skill-source-of-truth.md and separate the tracked Open Harness shared skill library (.agro/skills/*/SKILL.md, exposed through .agents/skills for Codex and Pi, .claude/skills for Claude, and Hermes' runtime .hermes/skills/openharness symlink) from the active Hermes runtime catalog (.hermes/skills/ plus bundled/profile skills). Do not present hermes skills list as the repo source of truth without this distinction.
5. For shared-skill path migrations or cross-agent skill wiring, load references/shared-skills-symlink-migration.md; update symlinks, runtime boot setup, docs, CI path filters, cron invocations, exact-path eval probes, and fixed-depth support scripts together before declaring the migration done.
6. For optional sandbox agent/runtime integrations, audit the full runtime/config/auth surface together: compose mounts and env, entrypoint setup/migration, default config seeding, banner sentinels, .gitignore, docs/wiki/changelog, and compose config validation. If Hermes is involved, load references/hermes-state-auth-split.md for the established single-device project-local runtime (config + auth + sessions all under one HERMES_HOME) + shared-skill symlink pattern, and why auth must not be a cross-device symlink.
7. For user-facing optional-runtime warnings, update the user-facing runtime docs/README when the fix is likely to recur. Example: Hermes gateway Slack warnings are covered in references/hermes-gateway-slack.md; document that Hermes' Slack gateway is separate from the Pi Slack extension, hermes-agent[slack] supplies slack-bolt, Open Harness' installed Hermes uses /usr/local/lib/hermes-agent/venv (so add extras with uv pip install --python /usr/local/lib/hermes-agent/venv/bin/python 'hermes-agent[slack]', not uv tool install), cron still runs without a messaging adapter, and .hermes/README.md is the right home for Hermes runtime-home/gateway notes rather than the top-level README. Ownership invariant: runtime paths a human uses inside the container should be writable by the default sandbox user; if an image-level install needs root during Docker build, constrain root-owned build paths to that RUN layer and reset runtime env/ownership afterward.
8. Answer with path:line citations.

Output shape

• Short factual answer first, then file references.
• Lifecycle questions (setup / validate / teardown): cite the relevant section of CLAUDE.md.
• Convention questions: cite the owning skill (.agro/skills/<name>/SKILL.md).