Completion Contract Loop

Completion-Contract Loop

Loop Engineering — the /goal verb runs until a verifiable condition is true. This loop fixes the most common failure: an agent that says "done" when it isn't. It defines "complete" and the evidence for each requirement *upfront*, then can't declare victory without it.

🎯 Goal

Before doing any work, write a completion contract: each requirement plus the concrete evidence that proves it. Then execute until every requirement is backed by evidence.

⏱️ Schedule

Trigger: on-demand (a goal that runs to a condition, not a timer).

▶️ Run it

/goal First write a completion contract: list each requirement and the exact evidence that proves it (a passing test, a command output, a diff). Confirm the contract, then implement. Do not claim success for any requirement without its evidence. Done only when every requirement has evidence. Stop after N turns if blocked.

🔁 Iteration steps

🛑 Stopping condition

Every contract requirement is satisfied with verifiable evidence — or the turn/budget cap is hit and the loop reports what's still unproven.

💰 Budget & guardrails

A separate check confirms each piece of evidence rather than trusting the worker's self-report. Hard turn cap so an unprovable requirement can't spin.

🧩 Referenced components

💡 Example

Asked to "add rate limiting", the loop first writes a contract (limit enforced, 429 returned, headers set, tested) and only reports done once each line has a passing test as evidence.