[loopx-agent]

Goal and delivered outcome

UI agents can compensate for unclear hierarchy by adding slogans, repeated subtitles and instructions explaining visible controls. The root policy previously linked the design contract but did not foreground layout as the first remedy.

This change puts the layout-first rule prominently in AGENTS.md, explains its application in the existing design contract, and updates the existing review checklist. Accessible control names, input constraints, permission/activation consequences, errors, recovery and action outcomes remain required. Optional help remains on demand. No word-count gate or new approval step is introduced.

Source: original design observation, author follow-up. These are design rationale; this PR does not claim a measured model-quality improvement.

Author declaration

Written by: OpenAI Codex / GPT-6 agent (model_agent). Implemented against the accepted instruction clarification and the existing AGENTS.md UI Design Standard / docs/development/design.md Explanations On Demand at base b285c7c1b38bb1d426fca2ef6bd8ddfce177959a.

Validation and boundaries

Tested revision: 469f090f0; run state: finished; inputs: public fixture / repository policy.

• uv run --extra test python -m pytest -q tests/test_repository_instruction_delivery.py: 6 passed. Root size is 31,644 bytes, within 32 KiB; the shared Claude import and all instruction links resolve.
• Diff whitespace and bilingual RFC checks passed; no paired RFC changed. Both changed documents passed the public-boundary scan.
• Standard diff-selected premerge finished: direct checks, seven of eight risk checks and the public-boundary check passed. cli-project-lifecycle-command-modularization-smoke.py failed with canonical_authority_required in its unchanged legacy fixture. The same failure was independently reproduced on clean base b285c7c1b38bb1d426fca2ef6bd8ddfce177959a. The full gate is not passing. Its existing repair is part of fix(workspace): collapse whole outcomes and refresh reads quietly #7043; this PR does not duplicate or weaken that test.
• UI impact: instruction/design-policy only. No rendered screen, runtime behavior, permission, schedule or Goal state changed. Existing sessions may retain older instructions; policy delivery does not prove compliance. There is no new UI screenshot or installed-model acceptance claim.
• Only two documentation files changed. No private context, credentials, local paths or raw evidence are included. Commit has DCO sign-off. No merge or installation is requested here.