Kickoff

Run one engineering episode from intent to a verified terminal state, with as little human involvement as safety allows. Two layers make one system:

• The episode layer (route, episode state, corrections, delivery states, closeout) decides how much process the work gets and when it is really done.
• The engineering layer (non-negotiables, principles, autonomy, subagents, reply, playbooks) decides how the work is done.

Skills are invoked by name (/kickoff). There are no mode toggles. Tool names below are T3 Code's. In another harness, map each one through references/harness.md before the first call. A skill not exposed to the harness's skill tool is read from its sibling folder (<skills-root>/<name>/SKILL.md).

Scope and authorization

Honor explicit user scope and applicable prior authorization over workflow defaults. Continue authorized preparation and reversible work using established conventions. Preserve engineering safeguards and access controls. If a skill blocks progress, identify the exact instruction and concrete conflict, fix the skill in its own PR when writes are in scope, and continue independent authorized work.

For a read-only or no-change request, return findings in chat without creating or updating plans, state, memory, configuration, or knowledge records unless those writes are explicitly authorized.

Intake

Make three decisions once and record them in the episode state.