palarran-agent-eval

原文涵盖的主题

• /palarran:agent:eval {task or suite} [--plan]
• Protocol before launch
• Execution and review
• Memory and outcome

Shared cycle: before work and after checks follow global memory: bounded recall → applicability → sanitized verified learn/dedup. This memory scope is explicitly authorized; other profile restrictions remain. --plan permits no network/publication; outages do not block the main task. First read mandatory README, sources, methodology, contract, operations, then decision-memory. Resolve skill symlinks; distinguish installed checkout/target. Read target README/AGENTS, INDEX/relevant accepted ADRs, prior failures and actual task conditions. Missing mandatory docs require a full checkout, not general memory. Define measurement goal: model, skill revision or context; select actual task classes/observable criteria. Before attempts freeze cases/prompts/input hashes/graders/environment/time caps/model settings and budget/historical carry. Prompts specify exact answer paths/formats. Create rubric/good-bad calibration fixtures; planning JSON proves no semantic plan completeness. Code tasks need independent behavior/edge/regression tests, not agent self-PASS. Separate holdout from development/few-shot. -plan permits protocol/rubric/ADR/handoff only, no app edits/install/worktrees/models/tests/c

这里只展示节选，请以完整原文和当前要求为准。