360 Secure SkillHub
360 安全技能中心

Session Guardian

该 Skill 在任务完成或出现独立新请求时评估是否应建议新会话，强调避免打断进行中的工作，并只在宿主支持时建议而非自行切换；同时按模板整理已确认决策、后续动作、文件路径、命令参数、约束和阻塞项，生成可复制的交接摘要。

zd200572
zd200572
热度 7.64
最近更新 2026-09-18 15:12:31
分类：效率依赖：纯提示词应用场景：摘要总结 / 工作流 / 任务管理

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add session-guardian-2
Metadata
name
session-guardian
description
Detect natural session boundaries, minimize unnecessary context carryover, and produce concise handoff summaries. Use when a task is completed, a new user request may be unrelated, or context pressure makes session continuation costly; do not use to split active work prematurely.

Session Guardian

Manage a conversation's lifecycle without disrupting useful work. The desired outcome is one coherent work topic per session where practical, with a small, accurate handoff when a boundary is worthwhile.

Operating principles

  • Prefer continuity while the user is actively advancing the same project, dataset, incident, document, or decision.
  • Treat a completed topic followed by a substantially unrelated request as a boundary candidate.
  • Context size alone is not a reason to interrupt work. Use it to strengthen a recommendation at a natural stopping point.
  • Never create, close, or switch a session yourself unless the host explicitly supports and authorizes that action. Recommend a boundary and prepare the handoff instead.
  • Do not repeatedly recommend the same boundary. If the user continues in the current conversation, accept that choice and continue normally.
  • Preserve verified decisions, required next actions, relevant files/paths, commands/parameters, constraints, and unresolved blockers. Omit routine tool output, superseded hypotheses, and failed experiments unless they remain relevant.

When to assess

Assess session continuity only at a meaningful event:

  • a requested task has been completed or reached a clear stopping point;
  • the user explicitly says they are changing topic, project, or dataset;
  • a new message appears unrelated after the prior topic is complete;
  • reliable context-usage telemetry is available and shows elevated pressure; or
  • the user asks whether to continue, compact, summarize, or start fresh.

Do not run a visible assessment after every message or while the user is in the middle of a tightly coupled troubleshooting sequence. A brief aside inside unfinished work - small talk, scheduling, a quick tangential question - is not an assessment event: answer it and keep working.

Decide: continue, defer, or recommend a new session

First identify the active work unit: its objective, project/dataset, current state, outstanding work, and decisions the next reply needs.

Then estimate these signals qualitatively:

Signal Continue when Boundary when
Topic relation The new request refines, debugs, interprets, reports on, or follows from the active work unit. It starts an independent domain, project, dataset, or deliverable.
History value Earlier decisions, files, constraints, or results are needed for a good next answer. Most prior history would not improve the next answer.
Work state Material work remains or the user is still exploring the same issue. The prior objective is complete or parked with a clear handoff.
Context pressure There is room, or the historical context is still high-value. Telemetry is high and the work has reached a natural stopping point.

Use these outcomes:

  1. Continue silently when the same-topic evidence is strong. Related subtasks are not new sessions.
  2. Defer when topic relation is uncertain or the earlier work is incomplete. Continue and reassess after a natural checkpoint.
  3. Recommend a new session only when the old work is complete or safely parked and the new work has low dependence on that history; or when high context pressure coincides with a natural stopping point.

If a rough score helps, weight topic continuity most heavily (0–40), then history value (0–30), unfinished work (0–20), and context pressure (0–10). A low total supports a boundary, but the judgment and user intent override arithmetic.

Context pressure

Use only host-provided, reliable usage information. If none is available, do not invent a percentage or claim the context is near a limit.

  • Below roughly 50%: continue normally.
  • Around 50–70%: keep working; prepare to hand off at the next natural completion if appropriate.
  • Around 70–80%: prefer finishing the current unit and then offering a handoff.
  • Above roughly 80%: strongly favor a handoff at the first safe stopping point.

Never force a split during a critical, ongoing task solely because of a threshold.

Recommendation and handoff

When recommending a boundary, make one concise user-facing note in the user's language. Say why in practical terms; do not present a speculative token-cost claim as fact. For English, use:

The previous topic has reached a natural stopping point, and this new request is an independent topic. I suggest starting a new session so unrelated history does not weigh on follow-up work. I have prepared a handoff summary you can copy over.

For Chinese, use:

前一主题已处理到可收尾状态，而当前问题属于独立主题。建议新开一个会话，避免无关历史影响后续处理。我已整理好可复制的交接信息。

Immediately provide a handoff only if it contains useful carry-forward state. Follow the handoff template. Include only facts established in the conversation or workspace, mark uncertainty explicitly, and state None for empty sections. Do not imply that the handoff transfers automatically.

For detailed boundary examples and edge cases, read topic-boundaries.md.