LWLT-AI/agent设计规范/skills/ltjt-business-operation/references/safety-and-results.md

Safety Gates and Result Semantics

Before any write

Require all applicable gates:

1. The operation matches the canonical schema and exactly one action.
2. Required business fields and existing identifiers are present.
3. Product/customer/route/OP/sales/parent mappings are unique or explicitly resolved by deterministic runtime logic.
4. submit_mode remains dry_run until local authorization enables execution.
5. Duplicate checks, task/profile lock, session readiness, and sender authorization pass.
6. The route has a verified form/API contract and post-operation query.
7. A live browser preflight and submit-intercept evidence exists for a new write route.

The Agent or Skill must not set allowRealSubmit, credentials, whitelist entries, authorization codes, browser selectors, or cookies.

Result meanings

Result	Meaning	Next action
dry_run	Valid plan; no ERP write	Review plan or authorize a separate execution.
operation_dry_run_completed	Read-only browser/context probe completed	Use the evidence to decide the next technical step.
blocked	A required rule, capability, safety, or authorization gate failed	Fix the named blocker; do not claim success.
execution_uncertain	Transport/page result does not prove durable state	Re-query registry/ERP before retrying.
completed	Durable identifier/artifact and post-operation verification exist	Return a safe business summary.
post_save_recovery_required	Save is verified; export/delivery failed	Route only to export recovery with neverResave=true.

Customer-facing boundary

Return only business status, safe identifiers, artifact type, next step, and necessary blocker. Never return selectors, stack traces, credentials, cookies, authorization codes, internal audit paths, or traveler PII.