Token Saver

> Universal token audit & optimization framework for OpenClaw agents.

> Based on real-world practice (2026-05-04).

Core Principles

1. Tier your model usage — Simple tasks use cheap models; complex reasoning

uses expensive ones. Don't mix the two.

1. Prompts say what, not why — Background rationale and philosophy are

noise to an agent. Strip them.

1. Batch > Serial — One call for 10 results costs marginally more than

three calls for 3+3+4 results. Combine.

1. Context = Cost — Every file loaded at session start, every tool schema

registered, every past message injected — all have a token price.

1. Idle = Zero burn — Nighttime, weekends, and idle periods should run

nothing. Configure active hours.

Output

After each full execution, write a report (token-audit-report-YYYY-MM-DD.md)

containing: before/after comparison table, estimated weekly savings per change,

items deferred and why, recommended next step.

Quick Start

Not every audit needs the full Phase 1-5 treatment. Use these shortcuts

based on your goal:

🚀 Express Audit (15 min)

Trigger: "快速 token 审计"

1. Run Phase 1A (enumerate cron tasks) + Phase 1D (model tier map)
2. Skip directly to Phase 3 category table and pick the lowest-hanging fruit
3. Apply Safe techniques only, no user confirmation needed
4. Skip Phase 4 and Phase 5 — just log the changes

🎯 Quick Wins (<5 min)

Trigger: "快速省 token"

Go straight to these high-impact, zero-risk techniques:

1. A3 (constrain output — add conciseness instruction)
2. B1 (right-size each task — cheapest viable model)
3. G1 (fixed prefix first — static prefix + dynamic suffix)
4. H1/H2 (default to working path, fail once and switch)

Apply directly without audit preamble.

🔄 When to run full audit

Phase 1: DISCOVER — Map the Full Token Landscape

1A Enumerate All Automated Tasks

Read your cron/scheduled task configuration (e.g. ~/.openclaw/cron/jobs.json).

For each task record:

• name
• model (or "default" if unset)
• message / prompt length in chars
• schedule frequency (daily / weekly / other)
• delivery.mode (announce / none)
• sessionTarget (isolated / main)

1B Analyze Agent Configuration

Inspect your gateway config (e.g. openclaw.json):

• agents.defaults.heartbeat.* — interval, active hours, isolated session,

light context flag

• agents.defaults.compaction.mode — message retention aggressiveness
• agents.list[].tools.profile — full, coding, or custom
• agents.list[].model — per-agent model override

1C Measure Context Load

List every file that is injected at session start (typically files in the

workspace root directory). Measure each in chars and estimate token cost

(~3 chars per token for CJK-heavy text, ~4 for English-heavy).

If LCM (Lossless Context Management) is active, note the number and average

size of compacted summary blocks injected per turn.

If tool schemas are accessible, estimate total schema chars:

(count of registered tools × average schema size in chars).

1D Map Models to Tiers

Categorize all available models into three tiers based on capability and cost:

• 🏆 Premium (strong reasoning, high cost): e.g. deepseek-v4-pro, gpt-5.x
• 🟡 Standard (balanced): e.g. deepseek-v4-flash, minimax-m2.7
• 🟢 Economy (lightweight): e.g. minimax-m2.7-highspeed, ollama local

Map each task from 1A to its current model tier.

> ⚠️ Checkpoint: Before moving to Phase 2, present your Phase 1 findings

> (task inventory, file sizes, model tier map) to the user.

> Confirm that the inventory is complete and the measurements are correct.

> This prevents optimizing the wrong things.

Phase 2: PRIORITIZE — Build Your Decision Matrix

Score each finding from Phase 1 along three independent dimensions:

How to Score

Compute a relative priority for each finding by inverting Risk and Effort:

Priority = ImpactWeight × (1 / RiskWeight) × (1 / EffortWeight)

Where each dimension maps to a simple numeric weight:

• Impact: High=3, Med=2, Low=1
• Risk: Safe=1, Moderate=2, High=3
• Effort: Easy=1, Med=2, Hard=3

Focus on items scoring ≥ 1.5 first. Skip items < 1.0 unless they are

trivially easy (effort=1) and safe (risk=1).

Common High-Impact Patterns

These patterns tend to score high across most deployments:

> ⚠️ Checkpoint: Show your top-3 priority items to the user.

> Confirm direction before starting optimization.

> If the highest-score items seem wrong, revisit Phase 1 measurements.

Phase 3: OPTIMIZE — Apply Categorical Techniques

> ⚠️ User confirmation gate: Techniques marked Moderate or High risk

> involve config changes, profile switches, or task merging. Before applying them,

> present the proposed change using this template and get explicit approval:

>

> ```

> ## Proposed Change

> Technique: [category/technique name]

> Target: [file/config path]

> Before: [current state, chars/tokens if measurable]

> After: [proposed state, estimated savings]

> Risk: [Moderate/High]

> Rollback: [how to undo]

> ```

>

> Techniques marked Safe can be applied directly.

Each category below contains a set of techniques. Apply them in priority

order from Phase 2 — start with the highest-score items first, regardless

of which category they fall into.

Failure Recovery

If a technique causes a problem:

• Config change: Restore the backed-up config file and reload.
• Cron merge broken: Restore the old separate cron job from version control

or re-create it from the original prompt.

• Profile switch issue: Revert to "full" profile, report the missing tool.
• Prompt compression over-aggressive: Restore from the diff backup (keep

pre-optimization prompt versions in a prompts/backup/ directory).

Category Selection Guide

Match your Phase 2 findings to the best starting category:

Category Decision Tree

If you're not sure which category to start with, follow this tree from top

to bottom — the first match tells you your likely best starting category:

1. Is the main session slow or expensive?
   → Check B (tiering) and J (routing)
   → Also check D (too many tools loaded?)

2. Are cron jobs consuming more than expected?
   → Check A (prompts too wordy?), then B (wrong model?)
   → If F (same-tier jobs not batched?)

3. Is context getting cut off mid-task?
   → Check C (bootstrap too large?) → I (progressive disclosure?)
   → Then J3 (incremental delivery?)

4. Are agent outputs too verbose?
   → Check E (output discipline) → H (behavioral discipline)

5. Is the same heavy prompt repeated across tasks?
   → Check G (provider-side caching: fixed prefix first?)

6. Are you seeing the same errors repeatedly?
   → Check H2 (fail once, switch) → H4 (fix root cause)

7. Default (no obvious symptom):
   Run Phase 1 from scratch → Phase 2 will tell you where to go

> Pro tip: Start with G (Provider-Side Caching) if you use DeepSeek.

> Cache pricing is 0.83% of uncached — fixing prefix structure alone

> can cut token costs by 90%+.

A. Prompt Simplicity

| A1 Strip preamble | Remove background/rationale paragraphs from task prompts. Keep only: trigger, action, output format.

Before: "你是系统监控助手。每天检查服务器状态：CPU使用率>80%告警、内存>90%告警、磁盘>85%告警、SSL证书<30天告警。每个告警按严重程度分别处理：严重→立即通知值班、一般→发运维邮件、提示→记录日志。"

After: "系统监控。检查：CPU(>80%) Mem(>90%) Disk(>85%) SSL(<30d)。告警：严重→立即、一般→邮件、提示→日志。" (360→110 chars, -69%) | Safe |

B. Model Tiering

C. Context Slimming

| C4 Trim to one-liner | Convert verbose descriptions to single-line summaries.

Before: "This project's coding conventions were established after three code reviews revealed inconsistent patterns: use 2-space indent for HTML/CSS, 4-space for Python, tabs for Go. Prefix private methods with underscore. No Hungarian notation. Import order: stdlib, third-party, local."

After: "Coding conventions (see CONTRIBUTING.md) — 6 rules, numbered."

Actionable instructions stay; background context goes. | Safe |

D. Tool Profile Optimization

E. Output Discipline

F. Session Lifecycle

G. Provider-Side Caching

> Impact is 10× any other category. DeepSeek V4 Pro cached price is 0.83% of

> uncached. Cache hit rates of 91-96% are achievable with proper prompt structure.

| G1 Fixed prefix first | Design all prompts as [static prefix] + [dynamic suffix]. Static prefix includes system instructions, bootstrap summary, and tool schemas. Dynamic suffix includes runtime instruction. This maximizes KV cache hits on the provider side.

Wrong: "Analyze this code for memory leaks...你是代码审查助手，审查规则如下：..."

Right: "你是代码审查助手，审查规则如下：...现在分析这段代码的内存泄漏：..." | Safe |

H. Behavioral Discipline

> These are zero-config, zero-cost techniques. The savings come from how you use

> the system, not how it's configured.

| H1 Default to working path | Use known-working tools before alternatives. Don't retry tools known to be broken in the current deployment — each retry is a wasted tool call + error response.

Bad: web_search (broken) → error → web_search again → error → baidu-search → works

Good: baidu-search → works (first attempt) | Safe |

I. Context Engineering (2026-05-12 新增)

> Context Engineering 是 2026 年从 Prompt Engineering 演进出的上层方法论。

> 核心原则：渐进式披露 (Progressive Disclosure) — 仅在任务需要时加载特定模块。

> 北京大学论文《Meta Context Engineering via Agentic Skill Evolution》实测：

> Token 消耗降低 60%，任务成功率提升 45%。

| I1 Progressive disclosure | 按任务类型分级加载系统能力描述。不需要的技能说明不加载。与 token-saver 的 C1-C4 互补：C 系列负责文件级裁剪，I1 负责能力级裁剪。

Wrong: Agent 启动时加载全部技能描述

Right: Agent 启动时按任务类型加载对应技能（搜索类任务只加载搜索相关技能描述） | Safe |

| I2 Semantic context scoring | 用语义重要性评分代替固定滑动窗口。计算历史交互与当前 query 的语义相似度，只保留 top-3 + 最近 1 轮。

推荐实现：Sentence-BERT 计算余弦相似度，按得分排序裁剪到 token budget。

参考开源项目：Agent Skills for Context Engineering | Safe |

| I3 Fixed prefix pattern | 设计所有 prompt 为「static prefix + dynamic suffix」。静态前缀包括系统指令、bootstrap 摘要、工具 schema 描述。动态后缀包括本次运行指令。最大化 provider 端 KV cache 命中。

Wrong: 每次运行时先写任务指令再补系统描述

Right: 系统描述在前，任务指令在后 | Safe |

| I4 Tool call circuit breaking | 在调用高成本外部工具（SQL 查询、图像生成、文件写入）前增加预检代理。校验参数合法性 + 资源预估。超限或非法直接拒绝，不触发实际调用。

适用场景：web_fetch 超长 URL、exec 高风险命令、文件写入大文件。 | Safe |

| I5 Cost-aware context pruning | 按 taskType 和 budget tokens 动态决定上下文精度。简单任务用精简 prompt，复杂任务才加载完整能力描述。

实现：在 cron payload 中嵌入 task_type 字段，根据该字段调度上下文模板。 | Safe |

J. Intelligent Model Routing (2026-05-14 新增)

> 基于 OpenSquilla（开源 Token 优化引擎）方案。

> Core insight: 路由决策本身不消耗 Token — 在本地判定任务复杂度后决定走哪个模型。

> 与 B 系列（Model Tiering）的区别：B 说"每个任务应该固定分配一个 tier"，

> J 说"同一个任务的每一次调用动态判断走哪个 tier"。

| J1 Dynamic inbound grading | 对每次入站请求做轻量复杂度判定（prompt length + expected reasoning depth + tool count），自动路由到对应 tier 的模型。

判定规则示例：

• 🟢 Economy: 简单信息查询（token 消耗 < 1K，无推理要求）→ sensenova-6.7-flash-lite / minimax-m2.7
• 🟡 Standard: 一般分析任务（1K-5K，轻度推理）→ deepseek-v4-flash
• 🏆 Premium: 深度推理（>5K，复杂推理/多工具）→ deepseek-v4-pro

实现：在 cron/子代理 payload 中嵌入 model 字段，或使用中间路由代理做分类。

Before: 所有每日 cron 统一用 deepseek-v4-pro，包括简单版本检查

After: 版本检查→sensenova-6.7-flash-lite，领域探针→deepseek-v4-flash，

深度分析→deepseek-v4-pro。按任务特征分级，非一刀切。 | Safe |

| J2 Local routing decision | 路由判定在本地完成（不需要发 request 到远程判断路由）。

方案：在任务 payload 中预置 routing_hint 字段（economy/standard/premium），

由编排层基于任务特征（prompt length、任务类型、预期工具调用数）决定。

Bad: 每次调用先发一个轻量请求问"这个走哪个模型？"（浪费 Token）

Good: 任务描述本身包含路由线索，编排层本地匹配 | Safe |

| J3 Incremental context delivery | OpenSquilla 核心方案之一：先发送最小必要上下文，不足时增量补充。

与 C 系列（Context Slimming）区别：C 是文件级裁剪（减少启动加载量），

J3 是调用级递进（先发主干，模型需要时再发细节）。

实现：prompt 中先给出高度压缩的摘要版本，后面用"如需详情请见"标记展开点。

结合 OpenClaw 的 lossless-claw 机制：LCM 摘要本身已经是一个渐进式实现。 | Safe |

| J4 Cache-aware routing | 统计各 provider 的 KV cache 命中率，优先路由到命中率高的 provider。

当前已知：DeepSeek V4 Pro cache price = 0.83% of uncached。

如果命中了缓存的 prompt prefix（含 bootstrap + 技能描述），成本降低 99%+。

实现：在路由时优先选择上次使用同一模型+同一 prefix 结构的 provider。 | Moderate |

| J5 Fallback chain | 当分配的路由模型不可用（API 超时/限流/报错），自动降级到备用模型。

与 H2（Fail once, switch）的区别：H2 是行为纪律（失败了就换方法），

J5 是配置级的自动 failover 链。

示例 Fallback Chain：

1️⃣ deepseek-v4-pro → 2️⃣ deepseek-v4-flash → 3️⃣ sensenova-6.7-flash-lite

实现：在 cron/子代理 payload 中使用 fallbacks 字段配置降级顺序。

Before: cron job 用 minimax-m2.7，卡死后无降级 → 任务永远失败

After: 同一 cron 配置 fallbacks: ["deepseek-v4-flash", "deepseek-v4-pro"] →

minimax 卡死后自动切到 deepseek-v4-flash 执行 | Safe |

> 权衡：智能路由的收益上限取决于任务复杂度分布。如果 80% 的任务是简单查询，

> 智能路由的 Token 节省可达 60-80%。如果大部分任务已经是标准/经济 tier，

> 额外收益有限。建议做一次任务复杂度分布扫描后再选择启用哪些 J 技术。

Phase 4: VALIDATE — Confirm Results

4A Prompt Length Delta

Before/after comparison of all modified prompts and files. Include total

chars and estimated tokens saved.

4B Config Integrity

After editing JSON configuration files, validate:

python3 -c "import json; json.load(open('<config-path>')); print('OK')"

4C Functional Test

• Verify cron tasks still start correctly (check cron action=runs or next

scheduled trigger)

• Verify heartbeat runs in configured active window
• Read through compressed cron prompts to ensure key instructions survive

4D Generate Report

Write token-audit-report-YYYY-MM-DD.md summarizing:

• Changes made and per-change token savings
• Total estimated weekly token reduction
• Items deferred and why
• Recommended next optimization

Log each optimization cycle in results.tsv (see skill directory for

format reference). This creates an audit trail for the quarterly deep audit (5B).

Phase 5: MONITOR — Guard Against Regrowth

5A Periodic Token Watch (Optional)

Optionally create a weekly cron (cheapest available model) that checks

prompt lengths haven't crept back:

{
  "name": "token-watch-weekly",
  "schedule": { "kind": "cron", "expr": "0 10 * * 1", "tz": "Asia/Shanghai" },
  "payload": {
    "kind": "agentTurn",
    "model": "<cheapest-model>",
    "message": "Check all cron prompt lengths. Flag any that grew >20% since last baseline.",
    "timeoutSeconds": 120
  },
  "sessionTarget": "isolated",
  "delivery": { "mode": "none" }
}

5B Quarterly Deep Audit

Run the full Phase 1-4 cycle every quarter using the cheapest available

model. Compare results against previous reports to spot regrowth trends.

The quarterly audit MUST include:

• Compare each cron's prompt length against last audit baseline
• Check if unused categories crept back (complacency regrowth)
• Verify all J routing hints still reflect actual task complexity
• Re-run the Priority Matrix from Phase 2 to catch new high-impact items

5C Real-Time Drift Alerts (Optional)

When token consumption suddenly spikes, it's usually one of three causes.

Know which one:

If you have access to provider dashboards:

• Monitor cache hit rate — DeepSeek dashboard shows prefix cache stats
• If hit rate drops below 80% -> inspect recent prompt changes for prefix variation
• Track per-model token consumption weekly — a 2× week-over-week jump is a red flag

Safety Boundaries

Configs That Need Gateway Restart

Some configuration paths require a gateway restart to take effect:

• agents.defaults.heartbeat.* (edit config file + restart)
• agents.list[].tools.profile
• gateway., auth.
• plugins.* — certain sub-fields

What NOT to Compress

These core mechanisms must be preserved even in an aggressive token budget:

• Error detection logic (consecutive errors, failure alerts)
• Essential signal handling (high-priority alerts → auto-escalation)
• Drift detection for recurring tasks

External References

• OpenClaw Cron Jobs: https://docs.openclaw.ai/automation/cron.md
• OpenClaw Standing Orders: https://docs.openclaw.ai/automation/standing-orders.md
• OpenClaw Gateway Config: openclaw gateway config CLI
• OpenClaw Agent Profiles: agents.list[].tools.profile in openclaw.json
• Test Prompts: test-prompts.json in skill directory
• Results Log: results.tsv in skill directory

Appendix: Local Deployment Configuration

This section is populated by the first execution of the Token Saver in a

specific deployment. Replace the example values below with real ones.

Configuration Paths

Baseline Measurements (example: Wave 2026-05-04)

Per-session token savings from bootstrap compression: ~6,720 tokens.

Benchmark: Compression by File Type

Task-to-Model Map

Deferred Items

Deployment-Specific Constraints

• Network: GFW blocks chatgpt.com, api.openai.com. All OpenAI/Codex models unavailable.
• Models available: deepseek-v4-pro (premium), deepseek-v4-flash (standard),

minimax-m2.7 (economy).

• File paths: Standard OpenClaw paths under ~/.openclaw/.
• Git: Workspace is a git repository; all changes version-controlled.