goal-prompt-drafting

# Goal prompt drafting ## Contract | Field | Bound contract | |---|---| | Trigger | User asks to draft, rewrite, compress, or prepare a copy-ready goal-mode objective or /goal command for a long-running agent. | | Authority | Read-only: no file, VCS, credential, paid, published, deployed, or remote mutation. | | Side effect | Chat output only; applies a deterministic one-line normalizer and returns one normalized /goal command plus only necessary missing-information warnings. | | Done | The objective states a measurable end state, grounded proof, invariants against easy-outs, and a stop or blocked clause; it is one line within the 4,000-character harness limit, and any ungrounded contract element is listed as missing. | ## Inputs - The user's draft, request, or conversation describing the intended work (required). - Repository context the agent may read to ground elements: files, the exact test command, logs, or a plan (optional; looked up rather than guessed). - The harness character limit is fixed at 4,000; `/goal` is taken as one line. ## Procedure 1. Read the user's draft and any referenced repository context. Never invent an element: look things up in the request, conversation, or repository; an element that cannot be grounded is left out and flagged (see Failure and recovery). Done when: the grounding set is listed and every element is sourced from the request, conversation, or repository, with ungroundable elements flagged for the Missing list. 2. Draft the condi...