← 返回 Skills 市场

Context Switch Token Optimizer

Name: Context Switch Token Optimizer
Author: nicochow

作者 nicochow · GitHub ↗ · v1.0.1 · MIT-0

cross-platform ✓ 安全检测通过

236

总下载

当前安装

版本数

在 OpenClaw 中安装

/install context-switch-token-optimizer

功能描述

智能对话上下文切换与 Token 优化。分析对话主题连续性以管理上下文、在话题间切换并节省 Token。适用于多任务对话、项目工作流、学习研究等需管理上下文与 Token 使用的场景。

使用说明 (SKILL.md)

Context Switch Token Optimizer

技能描述

通过关键词相似度、弱关联子串与时间衰减判断「连续 / 渐变漂移压缩 / 硬切换」，并可监控 Token、触发压缩与清理。实现见 context_manager.py、token_optimizer.py；详细算法见 docs/TOPIC_AND_CONTINUITY.md。

上下文动作（与代码一致）

动作	条件	行为
`load_memory`	无当前主题且无历史	建主题、可选加载 `memory/topic` 下记忆
`continuous`	与上一轮相似度 ≥ `similarity_threshold`	追加历史，不压缩
`drift_compress`	相似度 ∈ [`continuity_threshold`, `similarity_threshold`)	对话仍连续；压缩与当前弱相关的历史轮（`is_compressed`，保留短摘要）
`switch_context`	相似度 \x3C `continuity_threshold`	硬切换，换记忆、增 `switch_count`

相似度计算：calculate_topic_similarity()（关键词 Jaccard 式重叠 × 时间衰减 + 2 字子串/首字弱关联 + continuity_bonus_for_partial，弱关联封顶 drift_similarity_cap）。

持久化状态格式（`memory/context_switch_state.json`）

与 ContextState / TopicSummary 一致，主题历史项包含：

topic, keywords, timestamp, tokens_used, content_snippet, is_compressed
顶层另有：current_topic, memory_context, last_switch_time, switch_count, total_tokens

配置文件（`config.json`）

所有键名与 load_skill_config() 默认值一致；用户 JSON 按顶层块与默认合并（同一块内键覆盖，未写的子键保留默认）。

`context_switch`

键名	类型	默认	说明
`similarity_threshold`	number	0.7	≥ 此值判为 `continuous`
`continuity_threshold`	number	0.35	\x3C 此值判为硬切换 `switch_context`
`compress_relevance_threshold`	number	0.3	`drift_compress` 时，历史轮与当前相似度低于此则压缩该轮
`continuity_bonus_for_partial`	number	0.3	存在弱关联时的相似度加成
`drift_similarity_cap`	number	0.65	弱关联时相似度上限，便于进入 `drift_compress`
`time_decay_factor`	number	0.95	按小时衰减的底数
`max_topic_history`	int	10	主题历史条数上限
`memory_relevance_threshold`	number	0.5	记忆文件匹配阈值
`log_level`	string	（不设）	可由环境变量 `CONTEXT_SWITCH_LOG_LEVEL` 写入

`token_optimization`

键名	类型	默认	说明
`enabled`	bool	true	为 false 时不自动触发优化（对应环境变量 `TOKEN_OPTIMIZER_ENABLED`）
`token_limit`	int	80000	Token 预算
`compression_threshold`	int	56000	压缩参考阈值
`context_cleanup_threshold`	number	0.8	上下文清理参考
`memory_load_limit`	int	2000	记忆加载参考
`optimization_interval`	int	300	优化间隔（秒，预留）
`max_history_size`	int	100	Token 使用记录条数上限

`memory_search`

键名	类型	默认	说明
`max_search_results`	int	5	最多返回记忆条数
`keyword_limit`	int	5	每轮提取关键词数
`search_depth`	int	2	可由环境变量 `MEMORY_SEARCH_DEPTH` 覆盖（1–3）
`file_types`	string[]	`["*.md"]`	记忆目录 glob

`optimization_settings`

键名	说明
`auto_optimization`	是否自动优化（策略预留）
`health_score_threshold`	健康分阈值
`suggestions_check_interval`	建议检查间隔（秒）

环境变量（与 `apply_environment_overrides()` 一致）

环境变量	映射配置键	取值与行为
`CONTEXT_HISTORY_SIZE`	`context_switch.max_topic_history`	正整数，限制在 1–100
`MEMORY_SEARCH_DEPTH`	`memory_search.search_depth`	正整数，限制在 1–3
`TOKEN_OPTIMIZER_ENABLED`	`token_optimization.enabled`	`true`/`1`/`yes`/`on` 为启用；`false`/`0`/`no`/`off` 为禁用（禁用后使用率 >70% 也不触发 `trigger_optimization`）
`CONTEXT_SWITCH_LOG_LEVEL`	`context_switch.log_level`	`DEBUG` / `INFO` / `WARNING` / `ERROR` / `CRITICAL`；设置根 logger 级别

未设置上述环境变量时，行为完全由 config.json 与代码默认决定。

CLI 与入口

主入口：python3 context-switch-token-optimizer.py [--config path] --process|--status|--reset|--test
子模块：context_manager.py、token_optimizer.py（参数见 --help）

维护: Hermosa
版本: v1.1
最后更新: 2026-03-17

安全使用建议

This skill appears internally consistent and implements the described context-switching and token-optimization behavior without requesting secrets or performing network calls. Before installing/running: 1) Inspect or run it in an isolated directory (it will create/modify memory/context_switch_state.json and other files). 2) If you don't want automatic optimization, set TOKEN_OPTIMIZER_ENABLED=false to disable auto-triggering. 3) Enable debugging (CONTEXT_SWITCH_LOG_LEVEL=DEBUG) if you want to observe similarity calculations and compression decisions. 4) Review the memory/ files after test runs and confirm behavior on representative conversations. If you need the skill to never write to disk, do not run it or run with a sandboxed working directory.

功能分析

Type: OpenClaw Skill Name: context-switch-token-optimizer Version: 1.0.1 The skill bundle implements a legitimate utility for managing LLM conversation context and optimizing token usage through topic similarity analysis and history compression. The code operates locally on state files within a designated 'memory/' directory, lacks network communication capabilities, and contains no evidence of malicious execution (e.g., eval/exec), data exfiltration, or prompt injection attacks. All logic in 'context_manager.py' and 'token_optimizer.py' is consistent with the stated purpose of maintaining conversation continuity while reducing token overhead.

能力评估

✓ Purpose & Capability

Name/description (context switching and token optimization) match the code and SKILL.md. The skill only requests configuration and logging environment overrides relevant to topic management and token optimization; no unrelated credentials, binaries, or external services are required.

ℹ Instruction Scope

SKILL.md and CLI instruct running the included Python scripts which read/write a local state file (memory/context_switch_state.json), load config.json, and may scan local memory files (configured via file_types). The runtime instructions do not ask the agent to read arbitrary OS secrets or network endpoints. Note: the skill will create and modify files in the working directory (memory/), so consider where the agent runs.

ℹ Install Mechanism

There is no install spec (instruction-only from registry perspective), which is low-risk. However the package includes multiple Python source files (no external downloads or package installers). Because code is present, running the skill will write state to disk; no remote fetch/install behavior was observed.

✓ Credentials

Environment variables the skill reads (CONTEXT_HISTORY_SIZE, MEMORY_SEARCH_DEPTH, TOKEN_OPTIMIZER_ENABLED, CONTEXT_SWITCH_LOG_LEVEL) are directly tied to its configuration and behavior. No secrets/keys/tokens/passwords are required.

ℹ Persistence & Privilege

The skill persists its own state under memory/context_switch_state.json and updates it during operations (save_context_state). always:false and normal autonomous invocation are set. This is expected for a context manager, but be aware it will create/modify files in the agent's working directory.

如何使用

确保已安装 OpenClaw（本地或 Docker 部署）
在对话框中输入安装命令：/install context-switch-token-optimizer
安装完成后，直接呼叫该 Skill 的名称或使用 /context-switch-token-optimizer 触发
根据 Skill 的参数说明提供必要输入，即可获得结构化输出

版本历史

v1.0.1

- Added initial implementation files: context management and token optimization modules, plus test and debug scripts. - Provided detailed algorithm and configuration documentation. - Introduced a richer state persistence format including compression status and token usage tracking. - Supported advanced topic similarity calculation (Jaccard overlap, substring/partial bonus, time decay). - Environment variable overrides and merged user config supported. - Initial CLI support for processing, status, reset, and testing operations.

v1.0.0

Initial release of context-switch-token-optimizer: - Introduces intelligent context switching, summarizing dialog themes and managing conversation continuity. - Implements a three-level topic similarity mechanism for continuous, drift-compress, or hard context switching. - Features topic extraction, keyword analysis, and related memory search for efficient context management. - Includes token optimization strategies with progressive memory loading and context cleanup. - Provides monitoring and reporting of token usage and topic switches. - Suitable for workflows requiring efficient context and token management across conversations.

元数据

Slug context-switch-token-optimizer

版本 1.0.1

许可证 MIT-0

累计安装 0

当前安装数 0

历史版本数 2

常见问题