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. Load the mapped delivery workflow, current STATUS, ROADMAP and any applicable checkpoint before planning. Check dates and executable commands against current evidence; a stale profile is a finding, not a fact to silently preserve. Hooks provide orientation, not proof that these reads happened.
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

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

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.