Using the Harness

This is a team operating harness: one shared contract for humans and AI agents working in the same repository — same rules, same risk tiers, same evidence, same definition of done. It exists to make a mixed human–AI team coherent, not to make one agent faster.

A skill supplies a repeatable procedure. It never supplies permission or accountability.

First actions, every session

1. Ground in the contract. If this project has adopted the operating model, read its project profile before anything else (docs/operating-model/PROJECT-OPERATING-PROFILE.md), then the manual it binds. Profile facts beat inference.
2. Classify risk before acting. Map the requested work to R0–R3. Round ambiguity upward. Confirm authority at that tier before any consequential, destructive, external, or credential-touching action.
3. Route before doing. Check the routing table below. If a skill applies to the work, use it. If you are unsure which applies, use delivery-orchestrator.

The shared plugin is provider-neutral. Host adapters implement invocation; the project profile selects authorised reviewers, subscriptions and budgets. Independent Council review is a capability, not a fixed roster. See the review contract.

Installation and project adoption

A user-level installation makes the skills available across that user’s projects; a project installation shares a selected baseline with collaborators. Neither creates project authority. Use the host’s selected skill location, including bundled references and scripts, without assuming a copied .agents/skills/ directory.

For a new adoption, route to operating-model-bootstrap from the consumer project. Ground its project-owned vision, workflow and profile: ticket system, owners, model selection, executable checks, review, release and observation. Existing records and explicit operator authority take precedence. Load detailed skills when needed; an unadopted project needs orientation and an adoption path, not the full delivery process.

Core invariants — not negotiable

• Classify risk before confirming authority. Approval is scoped to actual risk.
• Infer intent; never infer permission. Useful intent is not a grant of authority.
• Receipts, not polish. Produce an artefact or a pass/fail verdict, not a narrative.
• Review binds to the exact candidate. Any change after review invalidates it.
• seed is not active. A seed profile supports design and R0/R1 work only.
• Never invent a project fact. Unknowns get an owner and a resolving trigger.
• Insufficient evidence is a stop, not a prompt to improvise. Tag the roster role that owns the decision; silence never converts to permission.

Skill routing

Work in front of you	Skill
New repo, or operating model install/repair	operating-model-bootstrap
Multi-track epic, or "which skill applies?"	delivery-orchestrator
An agreed goal to drive to a working, merged outcome while the operator is away	get-it-done
The machine is slow, tests or browser skills left processes behind	system-performance-guardrail
Is everything live, any errors this week, the weekly SRE checkup of a GCP project	cloud-checkup
Is the page fast enough, Core Web Vitals, why is this route slow	frontend-performance-audit
Vision, requirements, user journeys and product acceptance	product-owner
Material strategic, cost/risk or operating ownership decision	cxo-oversight
Acceptance coverage and executable user-path checks	test-strategy
Significant technical decision or ADR	the-architect
New feature, epic, or change — plan before code	spec-first-delivery
High-stakes decision, or before claiming done	adversarial-gate
Agent output feeding a decision, store, or another agent	domain-validator
Any PR or agent-generated code before commit	pr-reviewer
Anything about to ship to a real environment	release-readiness
CI triggers, Actions billing, issues, labels, branch protection, or release-tag workflow	github-manager
Release process governance: SemVer, ADR, changelog, release authority	release-manager
Enterprise policy, compliance, governance, or approved-vendor alignment	governance-guardrail
Status, standup, or situation report	sitrep
Plugin-directory, marketplace, or curated-list submission	plugin-submission
LLM/cloud cost estimate or architecture with LLM calls	cost-guardrail
Google ADK multi-agent design	adk-expert
Any cloud-vendor workload (GCP, AWS, Azure, Alibaba Cloud)	cloud-expert (the-architect routes)
Adding or reviewing an MCP server	mcp-server-scaffold

Companion plugins

Agent-craft capabilities may be provided by installed companion plugins instead of these skills — see INTEGRATIONS.md for the evaluated map (simplicity discipline, TDD methodology, frontend design review, and similar craft lanes).

• Route craft work to an installed companion when the project profile names one as the provider. Do not re-implement a capability an installed companion already provides.
• Companions supply capability, never authority. This harness's risk tiers, evidence, and review gates still apply to their output.
• No companion installed? The baseline skills above are the fallback. Record the equivalent and its owner in the project profile; do not silently skip the capability.

Harness tool mapping

Skills name actions, not tools. The per-harness translation lives outside the canonical skill bodies:

• Antigravity (Gemini) — read references/antigravity-tools.md (in this skill's directory) before acting; it maps subagent dispatch, task tracking, and file/search actions to Antigravity's native tools.
• Kimi Code — the plugin manifest's skillInstructions carry the mapping.
• Claude Code / Codex — the skills' action vocabulary resolves to the native tools of the same names.

Anti-rationalization table

Excuse the agent makes	Counter
"The skills are just docs — I can skip them"	The skills are the harness. Skipping them is running without the contract.
"This task is too small for a skill"	Gates are proportional to risk — R0/R1 is cheap. Skipped small gates are where big failures rehearse.
"The companion plugin handles quality"	Companions own craft. Authority, evidence, and completion semantics stay here.
"I'll edit the skill body to fit my tools"	Porting adds a thin adapter or tool mapping. Canonical skill bodies never change per harness.
"I remember the contract from earlier"	Sessions compact and clear. Re-ground in the tracked profile; memory is not the contract.
"The user seems in a hurry"	Speed is easy. Safe speed is engineered. Hurry raises risk tier, it does not lower it.

When unsure

Route through delivery-orchestrator. Record every unresolved fact with an owner and the trigger that resolves it. Stop before any R2/R3 mutation or external action that lacks explicit authority and runnable controls.