palarran-agent-improve

原文涵盖的主题

• /palarran:agent:improve {failed run + feedback} [--plan]
• Diagnosis and regression
• Change and paired verification
• Output

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 README, sources, methodology, contract, operations and decision-memory. Target differs from installed checkout. Read target AGENTS/INDEX/accepted ADRs, actual failing skill/mandatory domain guides. Never expand scope from failure logs/old approvals. ML corrections additionally require ML pilot/actual ML contract. Pin exact run/skill/harness/context/inputs, feedback and expected outcomes. Separate userreport, observed trace/tool evidence, causal hypotheses/reproduced outcomes. Missing original traces/tools/authority means honest NOTRUN, never simulated historical replay. Only approved sanitized observable events/outputs, concise rationale/hashes; request no hidden CoT and upload no secrets/customer logs to Git/few-shot. Calibrate oracles against good/bad/boundary outputs. Freeze regression/prior relevant controls/denominator/limits before candidates. Behavior checks are not instruction-presenc

这里只展示节选，请以完整原文和当前要求为准。