agent-harness-engineering

# Agent Harness Engineering Bundled commands use `$PLUGIN_ROOT` (`$env:PLUGIN_ROOT` in PowerShell; same path suffix) for the plugin root. Set it once: use the host's plugin-root variable when defined (Claude Code: `PLUGIN_ROOT="$CLAUDE_PLUGIN_ROOT"`), otherwise the absolute path of this skill folder's `../..`. Build a vendor-neutral control plane around an LLM. The harness, not generated text, owns state transitions, authority, resource limits, and recovery. Read `$PLUGIN_ROOT/references/agent-harness-contracts.md` for the artifact and runtime contracts. Read `$PLUGIN_ROOT/references/agent-harness-patterns.md` before selecting control, persistence, or delegation mechanisms. Use `$PLUGIN_ROOT/references/agent-harness-landscape.md` only for build-versus-adopt or provider/runtime comparisons. Choose the evidence surface before producing durable harness paperwork. A bounded diagnosis or fix may use an inline checklist plus direct tests. Create a versioned design artifact only for a new harness, a material control-plane revision, or a durable release or handoff claim with a named downstream consumer. ## Non-Negotiable Boundaries - Model output is an untrusted proposal, never authority to mutate state or invoke a side effect. - A transcript is evidence, not a checkpoint. Persist explicit typed state, pending work, policy decisions, budgets, and replay position. - Prompt or skill text is not a sandbox. Enforce permissions, approvals, isolation, and side-effect policy outside t...