Skip to content

CONVERSATION_PROTOCOL.md

Latest commit

 

History

History
113 lines (83 loc) · 4.79 KB

CONVERSATION_PROTOCOL.md

File metadata and controls

113 lines (83 loc) · 4.79 KB

CONVERSATION_PROTOCOL — Agent-Orchestrated State Machine

Last updated: 2026-04-15

0) Principle / 原则

This protocol defines a strict conversational flow for execution safety and auditability.

Critical rule: Authenticity and factual truth are judged by the user. The agent provides process structure, evidence organization, and controlled execution only.

关键原则：真实性由用户判断；Agent 只负责流程编排与能力执行，不替代用户做事实认定。

1) State Machine

goal -> platform-scope -> auth -> evidence -> risk-confirm(3x) -> export-decision -> execute -> notify -> close

2) Per-State Specification

State A: goal

  • User prompt: “What outcome do you want to achieve in this cleanup session?”
  • Required fields: goal_type, intent_summary
  • Exit condition: Goal is specific and actionable
  • Failure handling: If vague, ask narrowing questions with examples

中文：先明确目标（删除/下架/投诉/提醒），目标不清则继续澄清。

State B: platform-scope

  • User prompt: “Which platforms/targets are in scope? Please list URLs, IDs, or channels.”
  • Required fields: platforms[], targets[]
  • Exit condition: At least one normalized target exists
  • Failure handling: URL parse/dedup; unresolved targets go to manual review queue

中文：平台与范围必须落地到可执行目标（URL/ID）。

State C: auth

  • User prompt: “Please authorize access using the minimum required scope and short-lived credentials.”
  • Required fields: auth_mode, credential_ref|session_proof, consent_ack
  • Exit condition: Valid auth present or explicit none if channel doesn’t require auth
  • Failure handling: Auth expired/invalid => re-auth request, no silent fallback

中文：凭据遵循最小权限和短时效，认证失败必须显式重试。

State D: evidence

  • User prompt: “Provide evidence and explain why each target is in scope.”
  • Required fields: evidence_refs[], user_statement_authenticity_ack
  • Exit condition: Evidence package passes minimum completeness checks
  • Failure handling: Missing evidence => request supplementation; do not auto-fabricate

中文：证据不足就补证，不允许系统自行捏造事实。

State E: risk-confirm (3x)

  • User prompt:
    1. “This may be high-impact. Confirm #1 (YES).”
    2. “Please reconfirm #2 (YES).”
    3. “Final confirmation #3 (YES).”
  • Required fields: confirm1, confirm2, confirm3 (all exact YES)
  • Exit condition: Three valid confirmations received
  • Failure handling: Any mismatch => block execution and return to review

中文：高风险动作必须三次确认，任一失败即中断。

State F: export-decision

  • User prompt: “Before irreversible delete actions, do you want to export records first? (yes/no/ask)”
  • Required fields: export_before_delete, and export_answer when ask
  • Exit condition: Export choice is explicit
  • Failure handling: Missing answer when required => hard block

中文：删除前必须完成导出决策，避免不可逆损失。

State G: execute

  • User prompt: “Ready to execute now under manual mode. Proceed?”
  • Required fields: manual_trigger=true, execution_plan_ref
  • Exit condition: Adapter returns success|partial|failed
  • Failure handling: Capture structured error, provide retry or manual fallback path

中文：仅手动触发执行，禁止后台静默自动跑。

State H: notify

  • User prompt: “How should I notify results? (none/telegram/email/signal)”
  • Required fields: notify_mode
  • Exit condition: Notification branch resolved
  • Failure handling: Unsupported mode => default to none with explicit notice

中文：通知是可选分支，用户可选不通知。

State I: close

  • User prompt: “Session complete. Do you want an audit export and follow-up reminder?”
  • Required fields: close_ack, optional followup_preference
  • Exit condition: Audit summary persisted and session closed
  • Failure handling: If persistence fails, mark as close_pending and retry save

中文：收尾阶段输出审计摘要，确保可复盘。

3) Minimal Session Object

{
  "session_id": "sess_abc123",
  "state": "execute",
  "goal": { "goal_type": "dmca_takedown", "intent_summary": "Remove infringing reposts" },
  "scope": { "platforms": ["x"], "targets": ["https://x.com/.../"] },
  "auth": { "auth_mode": "oauth", "credential_ref": "vault://..." },
  "evidence": { "evidence_refs": ["ev1"], "user_statement_authenticity_ack": true },
  "risk_confirm": { "confirm1": "YES", "confirm2": "YES", "confirm3": "YES" },
  "export": { "export_before_delete": "ask", "export_answer": "yes" },
  "notify": { "notify_mode": "telegram" }
}