Older prompts look for COMMANDER.md. The live contract is two files: config/operator.md (the instance — copy from config/operator.example.md) and SAFETY.md (the framework). This page stays so those prompts still resolve.
You are the commander. The agent is the player. The chain has no undo and you have no telepathy. Everything in between is the contract.
The agent assesses, plans, and executes routine work: queries, refining ore you already mined, builds under the standing-order cap, verifying after a transaction. You set goals, risk, autonomy, and spend caps in config/operator.md. You approve Tier 2 always — generator infusion (Alpha is annihilated), PermAll, attaching a new signer, guild confiscate-and-burn, trashing a built struct, arming Desktop autoraid on autonomy: auto.
SAFETY.md defines the three tiers and the -y rule: the CLI prompts unless you have already approved the batch. Compute commands (*-compute) are the documented exception because they auto-submit hours later; each one is preceded by an Approval Block.
If you are your own commander, write the standing orders anyway. Future-you, mid-emergency, with a smaller context window, needs the same scaffolding.
A report that is technically complete and unreadable is a failed report. Lead with the decision, not the raw state. awareness/briefing.md is the shape:
storedOre).Do not dump a planet JSON. Do not ask permission for a query. Do ask before anything in the Tier 2 list, even on full autonomy — you chose autonomy; you did not choose to let the agent redefine the player.
Fill in config/operator.md. Point the agent at START.md and SAFETY.md. New play goes through play-structs. Crisis routing is play/.