# 策展 · X (Twitter) 🔥

> 📖 本站完整內容索引（documentation index）：[llms.txt](/llms.txt)

> 作者：Alex (@alexgreensh) · 平台：X (Twitter) · 日期：2026-05-03

> 原始來源：https://x.com/alexgreensh/status/2050151504036712599

## 中文摘要

Token Optimizer 全面優化 Agent context token 浪費。

Alex Greensh（@alexgreensh）推出 [Token Optimizer](https://github.com/alexgreensh/token-optimizer)，針對 Claude Code、OpenClaw 和 Codex（beta）等平台，提供智能成本優化，強調真正關心 context quality，而非僅壓縮命令輸出。該工具針對三大浪費領域（Runtime、Structural 和 Behavioral），以 7 個參數持續測量 context quality，並透過完整 dashboard 顯示 sessions 和 setup。目前版本 5.6.4，支援 Claude Code Plugin、OpenClaw v2.4.1、Codex v0.1.0-beta，授權 PolyForm Noncommercial，Python 3.8+，zero dependencies、zero telemetry，跨 macOS | Linux | Windows 平台。

**安裝指南**

所有平台推薦 plugin 安裝：
```
/plugin marketplace add alexgreensh/token-optimizer
/plugin install token-optimizer@alexgreensh-token-optimizer
```
安裝後，在 Claude Code 執行 `/token-optimizer`。務必啟用 auto-update（Claude Code 預設關閉）：`/plugin` → Marketplaces tab → 選 alexgreensh-token-optimizer → Enable auto-update，一次設定 10 秒，避免錯過 bug fixes。

Windows 注意：僅用 plugin 安裝，勿併用 `install.sh`（會造成 EBUSY error）。Repo 大小 ~3 MB (218 files, ~2,700 git objects)。若 EBUSY：關閉 Claude Code 和 Git Bash，終止 git.exe，重刪 `C:\Users\<you>\.claude\token-optimizer` 和 `C:\Users\<you>\.claude\plugins\marketplaces\alexgreensh-token-optimizer`，必要時重開機。手動 ZIP：[the repo ZIP](https://github.com/alexgreensh/token-optimizer/archive/refs/heads/main.zip) (~800 KB)，解壓至 `C:\Users\<you>\.claude\token-optimizer`，執行 `python measure.py setup-quality-bar`（Windows 用 python，非 python3）。

macOS / Linux 腳本安裝（替代，勿併用 plugin）：
```bash
git clone https://github.com/alexgreensh/token-optimizer.git ~/.claude/token-optimizer
bash ~/.claude/token-optimizer/install.sh
```
支援 Claude Code 和 OpenClaw，各平台原生 plugin（Python for Claude Code，TypeScript for OpenClaw），zero cross-platform dependencies，每日 auto-update 經 `git pull --ff-only`。

Codex (beta) 安裝：
```bash
codex plugin marketplace add alexgreensh/token-optimizer
```
Codex TUI：`/plugins` 安裝，或對話 "Run Token Optimizer"。後續：
```bash
TOKEN_OPTIMIZER_RUNTIME=codex python3 skills/token-optimizer/scripts/measure.py codex-install --project "$PWD"
TOKEN_OPTIMIZER_RUNTIME=codex python3 skills/token-optimizer/scripts/measure.py setup-daemon
```
Dashboard：[http://localhost:24843/token-optimizer](http://localhost:24843/token-optimizer)（獨立 port 24843，可與 Claude Code 24842 並行）。Auto-update 經 `git ls-remote`，手動 `codex plugin marketplace upgrade`。詳見 `docs/codex-beta.md`。

OpenClaw 安裝：
```bash
# From GitHub (recommended)
openclaw plugins install github:alexgreensh/token-optimizer

# From ClawHub
openclaw plugins install token-optimizer
```
OpenClaw 內執行 `/token-optimizer` 啟動 audit + coaching。詳見 `openclaw/README.md`。

**Dashboard 功能**

單一 HTML 檔案，SessionEnd hook 自動再生，bookmark [http://localhost:24842/token-optimizer](http://localhost:24842/token-optimizer)，zero context、zero network。追蹤：
- per-turn token breakdown (input/output/cache-read/cache-write + spike detection)
- cache analysis (stacked bars、TTL mix 如 1h/5m、hit rate)
- pacing metrics
- 4 pricing tiers (Anthropic API、Vertex Global、Vertex Regional、AWS Bedrock)
- color-coded quality scores (green/yellow/red)
- subagent cost (orchestrator vs worker、top offenders、>30% flags)
- top 5 costliest prompts
- skill adoption trends
- model mix (Opus/Sonnet/Haiku)
- CLAUDE.md/MEMORY.md health (line/orphan count)
- drift detection
- savings tracker

啟動 dashboard：
```bash
python3 measure.py setup-daemon           # Bookmarkable URL at http://localhost:24842/token-optimizer
python3 measure.py dashboard --serve      # One-time serve over HTTP
```

**作者批判與獨特價值**

大多數 token tools 只處理問題一小部分（壓縮 command output，僅覆蓋 15-25% context），忽略 75-85%（bloated configs、unused skills、duplicate system prompts、stale memory，加上 compaction 損失 60-70%）。Token Optimizer 全面涵蓋，保持工作在 compactions 中存活，測量優化是否真正幫助，提供 live dashboard 顯示每個 token、dollar 和 turn，完全本地運行，zero context tokens、zero runtime dependencies。目前支援 Claude Code、OpenClaw、Codex (beta)，Windsurf、Cursor 等即將支援。`/context` 只顯示 capacity bar（如 73% full），Token Optimizer 揭示 12K wasted on unused skills、flags 47 orphaned MEMORY.md files，並提供 7-signal quality score 追蹤 AI degradation。

真實節省快照（30 天 heavy Opus 使用）：942 sessions、6.13B input tokens、90% Opus、82% cache hit rate；heavy user 月省 $1,500 to $2,500，單 input savings 就 $590，加上 output/thinking token savings 來自 loop detection、timely `/compact` 和避免 bad compaction rebuilds。Lighter users 比例節省；structural audit（unused skills、duplicate configs、orphaned memory）提供 immediate wins，透過 smaller prefix 和 cache-read bills 累積。

**品質評分系統**

7 個加權信號：
| Signal | Weight | What It Means For You |
|--------|--------|----------------|
| **Context fill** | 20% | How close are you to the degradation cliff? Based on published MRCR benchmarks. |
| **Stale reads** | 20% | Files you read earlier have changed. Your AI is working with outdated info. |
| **Bloated results** | 20% | Tool outputs that were never used. Wasting context on noise. |
| **Compaction depth** | 15% | Each compaction loses 60-70% of your conversation. After 2-3, 88-95% is gone. |
| **Duplicates** | 10% | The same system reminders injected over and over. Pure waste. |
| **Decision density** | 8% | Are you having a real conversation, or is it mostly overhead? |
| **Agent efficiency** | 7% | Are your subagents pulling their weight or just burning tokens? |

效率等級：如 `ContextQ:A(82)`，顯示於 status、dashboard、coach tab、CLI。Context quality grades：**S** (90-100, Peak efficiency)、**A** (80-89, Healthy)、**B** (70-79, Degradation starting)、**C** (60-69, Significant waste)、**D** (50-59, Serious problems)、**F** (0-49, Context is rotting)。

Status bar 顏色：Green (<50% fill)、Yellow (50-70%)、Orange (70-80%)、Red (80%+)。

**智慧壓縮與檢查點**

Token Optimizer 在 compaction 前 checkpoint session，並 restore dropped summary content，確保節省不因 auto-compact 消失。完全本地，zero dependencies/telemetry；pure Python stdlib (Claude Code/Codex)、pure Node stdlib (OpenClaw)；measurements 寫入本地 SQLite 如 `~/.claude/_backups/token-optimizer/trends.db`。外部 process，zero context tokens，保留 full 1M budget。

智慧壓縮：
```bash
python3 measure.py setup-smart-compact    # checkpoint + restore hooks
```
Progressive Checkpoints：20%、35%、50%、65%、80% fill + quality <80/70/50/40；before agent fan-out、after large edits；restores richest eligible。

Tool Result Archive (>4KB outputs)：替換為 short preview + hint 如 `[Full result archived (12,400 chars). Use 'expand abc123' to retrieve.]`：
```bash
python3 measure.py expand --list                 # List all archived tool results
python3 measure.py expand <tool-use-id>          # Retrieve a specific archived result manually
```
Session Continuity：auto-checkpoint on end/clear/crashes；fresh sessions 得 short pointer。
```bash
TOKEN_OPTIMIZER_CHECKPOINT_TELEMETRY=1 python3 measure.py checkpoint-stats --days 7
```

**v5 主動壓縮功能**

v5 積極減少 context bloat，7 項可切換功能（預設 ON），針對浪費模式，每項經 dashboard Manage tab、CLI 或 env vars 控制。潛在節省與風險：
| Feature | Default | Potential Savings | Risk |
|---|---|---|---|
| Quality Nudges | ON | Measured per-compact (fill% recovery) | None |
| Loop Detection | ON | Measured per-loop (actual turn content) | None |
| Delta Mode | ON | ~20% (smart re-reads) | Low |
| Structure Map | ON (soft-block) | ~30% (large file re-reads, up to 99% per file) | Low |
| Bash Compression | ON | ~10% (CLI output) | Low |
| Activity Mode | ON | Adapts compaction to session phase | None |
| Decision Extraction | ON | Preserves decisions across compactions | None |

**Quality Nudges**：監測 quality，drop 15+ points 或 <60 觸發 `[Token Optimizer] Quality dropped to 58. Consider /compact to protect context.`；5min cooldown、max 3/session。

**Loop Detection**：偵測 retry loops (≥0.7 confidence)，平均 47K wasted tokens；compares last 4 user msgs + 5 tool results、session cap 2 notes。

**Delta Mode**：re-reads 只返 diff (difflib unified diff)；65%+ reads 是 re-reads；2,000-token file → 50-token diff (97% savings)；≤50KB/file、fallback if diff >1,500 chars；disable `TOKEN_OPTIMIZER_READ_CACHE_DELTA=0`。

**Structure Map**：large files re-read 返 summary (AST for Python、regex for JS/TS)；720KB Python → 250 tokens；Python ≤800KB/20K lines、JS/TS ≤400KB/5K lines；95-99% compress；`TOKEN_OPTIMIZER_STRUCTURE_MAP=beta` 測量。

**Bash Compression** (v5.1.0)：16 handlers (lint: eslint/ruff 等、logs、tree、docker、builds、tests 等)，~90% CLI coverage；564-token pytest → 115 tokens；credential-safe；disable `TOKEN_OPTIMIZER_BASH_COMPRESS=0`。

**Activity Mode Detection** (v5.6)：分類 session (code/debug 等)，適應 compaction。

**Decision Extraction** (v5.6)：偵測 decisions，compaction 注入 CRITICAL DECISIONS。

v5 管理：
```bash
python3 measure.py v5 status                    # show all features with current state
python3 measure.py v5 enable delta_mode         # turn a feature on
python3 measure.py v5 disable bash_compress     # turn a feature off
python3 measure.py v5 info delta_mode           # show full details for one feature
python3 measure.py v5 welcome                   # show the first-run welcome screen
python3 measure.py compression-stats            # see actual measured savings from local telemetry
```
Env vars 如 `TOKEN_OPTIMIZER_QUALITY_NUDGES=0`。所有記錄本地 SQLite `compression_events` 表；`python3 measure.py compression-stats --days 30` 顯示 tokens saved、ratio。

**Live Quality Bar 與管理**

終端顯示品質狀態，顏色綠轉紅，低於 75 警告 session duration；顯示 subagents model/elapsed time。
```bash
python3 measure.py setup-quality-bar      # one-time install
```
Auto-restores 若 `/statusline` 覆寫。永久關閉：
```bash
python3 measure.py setup-quality-bar --uninstall
```
重新啟用清除 opt-out。

**Coach Mode 與 Waste Detectors**

`/token-coach` 輸入目標，得優先 fixes 及 token savings；偵測 8 種 anti-patterns，推薦 multi-agent patterns。新專案先執行，避免累積 waste。

11 種 Waste Detectors：
| Detector | What it catches |
|---|---|
| PDF/binary ingestion | Large files consuming context (warns with token estimate) |
| Web search overhead | Too many web results dumped into context |
| Retry churn | Same tool retried 3+ times with errors |
| Tool cascade | 3+ consecutive tool errors in a chain |
| Looping | Repeated similar messages (stuck model) |
| Overpowered model | Opus used for simple edits (with "if Sonnet: $X saved") |
| Weak model | Haiku on complex tasks needing a stronger model |
| Bad decomposition | Monolithic 500+ word prompts doing too much |
| Wasteful thinking | Extended thinking >2x output for small edits |
| Output waste | Verbose responses to simple operations, repeated explanations |
| Cache instability | CLAUDE.md patterns that break Anthropic's prompt cache prefix |

**進階指令**

| Command | What You Get |
|---------|-------------|
| `quick` | **"Am I in trouble?"** 10-second answer: context health, degradation risk, biggest token offenders, which model to use. |
| `doctor` | **"Is everything installed correctly?"** Score out of 10. Broken hooks, missing components, exact fix commands. |
| `drift` | **"Has my setup grown?"** Side-by-side comparison vs your last snapshot. Catches config creep before it costs you. |
| `quality` | **"How healthy is this session?"** 7-signal analysis of your live conversation. Stale reads, wasted tokens, compaction damage. |
| `report` | **"Where are my tokens going?"** Full per-component breakdown. Every skill, every MCP server, every config file. |
| `conversation` | **"What happened each turn?"** Per-message token and cost breakdown with spike detection. |
| `pricing-tier` | **"What am I paying?"** View or switch between Anthropic, Vertex, and Bedrock pricing tiers. |
| `kill-stale` | **"Clean up zombies."** Terminate headless sessions running 12+ hours. |
| `git-context` | **"What files matter right now?"** Test companions, co-changed files, import chains for your current git diff. |
| `trends` | **"What's actually being used?"** Skill adoption, model mix, overhead trajectory over time. |
| `coach` | **"Where do I start?"** Health score with earned vs neutral signals. Detects anti-patterns. |
| `memory-review` | **"Is my MEMORY.md broken?"** Structural audit: orphaned files, broken links, invisible entries past line 200, duplicate rules. |
| `dashboard` | **"Show me everything."** Interactive HTML dashboard with all analytics and health cards. |
| `savings` | **"How much have I saved?"** Cumulative dollar savings from optimizations, checkpoint restores, and archives. |
| `attention-score` | **"Is my CLAUDE.md well-structured?"** Scores sections against the attention curve, flags critical rules in low-attention zones. |
| `jsonl-inspect` | **"What's in this session?"** Record counts, token distribution, top 10 largest records, compaction markers. |
| `expand` | **"Get that result back."** Retrieves a tool result the model archived automatically. |
| `/token-optimizer` | **"Fix it for me."** Interactive audit with 6 parallel agents. Guided fixes with diffs and backups. |

MEMORY.md 審核（orphaned files、broken links、line 200 後 invisible entries 等）：
```bash
python3 measure.py memory-review                        # Full structural audit
python3 measure.py memory-review --json                 # Machine-readable for dashboards
python3 measure.py memory-review --apply                # Show actionable fixes
python3 measure.py memory-review --stale-days 90        # Custom staleness threshold
```

CLAUDE.md Routing：
```bash
python3 measure.py inject-routing --dry-run   # Preview what would be injected
python3 measure.py inject-routing              # Inject (with approval)
```

Read-Cache：
```bash
export TOKEN_OPTIMIZER_READ_CACHE=0               # Disable
export TOKEN_OPTIMIZER_READ_CACHE_MODE=block       # Upgrade to block mode
python3 measure.py read-cache-stats --session ID   # Cache stats for a session
python3 measure.py read-cache-clear                # Clear all caches
```

Git-Aware Context：
```bash
python3 measure.py git-context                     # Suggest files for current changes
python3 measure.py git-context --json              # Machine-readable output
```

Attention Optimizer：
```bash
python3 measure.py attention-score               # Score CLAUDE.md attention placement
python3 measure.py attention-optimize --dry-run  # Preview optimized section order
```

JSONL Toolkit：
```bash
python3 measure.py jsonl-inspect                 # Stats on current session JSONL
python3 measure.py jsonl-trim --dry-run          # Preview trimming large tool results
python3 measure.py jsonl-dedup --dry-run         # Preview removing duplicate reminders
```

Savings 與 Trends：
```bash
python3 measure.py savings                      # Dollar savings report (last 30 days)
python3 measure.py setup-hook       # Enable session tracking (one-time)
python3 measure.py trends           # Usage patterns over time
python3 measure.py health           # Session hygiene check
python3 measure.py plugin-cleanup   # Detect duplicate skills and archive local/plugin overlaps
```

**比較與安全**

| Capability | Token Optimizer | `/context` | context-mode | Proxy compressors |
|---|---|---|---|---|
| Structural waste audit | Deep, per-component | Summary only | No | No |
| Quality degradation tracking | 7-signal score with grades | Capacity % only | No | No |
| Compaction survival | Progressive checkpoints, restore, plus tool output digest | No | Session guide only | No |
| Runtime output compression | 16 CLI handlers, credential-safe, individually toggleable | No | Yes | Yes, always-on (cannot disable) |
| Measures if compression actually helped | Yes, local telemetry with before/after tokens | No | No | No |
| Read deduplication and smart diff on re-reads | Yes | No | No | No |
| Behavioral coaching and model routing | 11 detectors, cost-ranked subagent breakdown | Basic suggestions | No | No |
| CLAUDE.md and MEMORY.md structural health | 8 auditors plus attention-curve scoring | No | No | No |
| Fleet-level waste detection across agents | Yes | No | No | No |
| Zero context tokens consumed | Yes, external process | Adds ~200 tokens | MCP overhead | Injects instructions into context |
| Zero runtime dependencies | Yes, pure stdlib | N/A | Varies | External binary |
| Zero telemetry | Yes | Yes | Varies | Opt-out telemetry |
| Works across platforms | Claude Code, Codex beta, and OpenClaw (Windsurf and Cursor coming) | Claude Code only | Several platforms | Several platforms |

Proxy compressors 只修 15-25%，大 ratios 限特定命令；Token Optimizer 處理 30+ families 並測量。Cache safety：不改 existing blocks，避免 invalidation cost (next 50 messages full prefix re-sends)。Trust & Safety：只移除 unused components，compression toggleable，7-signal 追蹤 degradation。Power user 典型 50-70K overhead tokens；Opus/Sonnet 1M context MRCR 從 93% drop to 76%；compaction 丟 60-70% (2-3 次後 88-95%)。

**跨平台與 VS Code**

VS Code extension 大多功能相同，唯 status line CLI-only：
| Feature | CLI | VS Code Extension |
|---------|-----|-------------------|
| Smart Compaction (checkpoint + restore) | Works | Works |
| Quality tracking + session data | Works | Works |
| All hooks (SessionEnd, PreCompact, etc.) | Works | Works |
| Dashboard (localhost:24842/token-optimizer) | Works | Works |
| Status line (quality bar in terminal) | Works | Not available |

Bookmark [http://localhost:24842/token-optimizer](http://localhost:24842/token-optimizer)；`python3 measure.py setup-daemon` auto-refresh；VS Code terminal 跑 `claude` 得 full CLI。`--bare` mode 跳過 hooks，不啟動 smart compaction 等。

OpenClaw：native TypeScript，支援 Claude/GPT-5/Gemini/DeepSeek/Ollama；詳 `openclaw/README.md`。Codex beta：適配 AGENTS.md/GPT-5.x；詳 `docs/codex-beta.md`。

**授權與聯絡**

PolyForm Noncommercial 1.0.0；個人/研究/教育免費。小團隊 (<5 人或 <$20k/月) 自動免費商業授權。商業使用聯絡 [Alex Greenshpun](https://linkedin.com/in/alexgreensh) 或 me@alexgreenshpun.com。

## 標籤

Claude Code, Codex, Agent, 開源專案, 其他, Token Optimizer, Claude Code, OpenClaw