Token Saver

Reduce OpenClaw AI costs with model-aware optimization. Features dynamic compaction presets based on your model's context window, intelligent file compression, and robust model detection with fallback. Supports Claude, GPT-4, Gemini, DeepSeek, and more.

Security Scan

The skill mostly does what it says (scan and compress workspace .md files and compute model-aware compaction), but it also reads/writes agent/home files and contains prompt-like text that encourages auto-execution and persistent modifications — these behaviors merit caution before installing.

Purpose & Capability

Overall the code matches the stated purpose: it discovers .md workspace files, has a model registry, computes dynamic compaction presets, and performs local compression. However the skill explicitly reads the user's home dir and OpenClaw config (~/.openclaw/openclaw.json) and writes to agent files such as AGENTS.md and creates a .token-saver-config.json in the workspace. Reading/writing these files is consistent with an optimizer but is broader access than a purely read-only 'analyzer' — worth noting before install.

Instruction Scope

SKILL.md claims 'No external calls — All analysis runs locally', which matches the code (no outgoing network calls seen). But runtime instructions and code intentionally read many workspace and agent files (SOUL.md, AGENTS.md, MEMORY.md, etc.) and provide a 'Persistent Mode' that writes guidance into AGENTS.md. More concerning: compressor code contains replacement patterns and prebuilt compressions with phrases such as 'Auto-execute, no permission needed' and 'Don't ask permission. Just do it.' and a 'SYSTEM:' snippet flagged as prompt-injection — these artifacts could nudge an agent to bypass permission boundaries if combined with autonomous invocation.

Install Mechanism

There is no remote install spec (instruction-only), which limits supply-chain risk. However the package includes runnable JavaScript files (analyzer.js, compressor.js, optimizer.js). Installing/placing these files on disk and running them will execute the included logic locally. No external downloads or archive extraction were found.

Credentials

The skill declares no required env vars or credentials, but the code reads optional environment variables (SKILL_MODEL, OPENCLAW_MODEL, DEFAULT_MODEL) and the user's home directory. While environment model variables are appropriate for model detection, reading the home directory and agent config files grants access to user-scoped configuration and session files — a level of access that is sensitive and should be justified. The skill does not request API keys or secrets, which reduces credential risk.

Persistence & Privilege

The skill writes backups (.backup) and can alter AGENTS.md and write a workspace .token-saver-config.json to persist compaction settings. Those are legitimate for 'persistent mode', but any skill that modifies agent configuration / AGENTS.md is higher-privilege because it can change agent behavior across sessions. Combined with prompt-like strings that encourage 'auto-execute', persistence capability increases the blast radius if misused.

Scan Findings in Context

[path-traversal] expected: Detected path.resolve(dir, '..', '..') usage in scripts — the skill intentionally traverses to workspace root to discover .md files. This is expected for a workspace-wide optimizer but means the code will access files outside the skill folder.

[homedir-access] expected: Code reads process.env.USERPROFILE / HOME and tries ~/.openclaw/openclaw.json to detect model settings. This matches the declared detection chain but means the skill can read files in the user's home directory.

[memory-write] expected: Scanner flagged writing to agent memory/config files (AGENTS.md, MEMORY.md). The skill documents and implements persistent mode that modifies AGENTS.md and creates backups; this is consistent with its stated behavior but is a high-sensitivity action.

[prompt-injection-system] unexpected: A 'SYSTEM:' style snippet and compressor replacement patterns (e.g., 'Don't ask permission. Just do it.' → 'Auto-execute.') were detected. These appear to be prompt-injection or instruction-like strings embedded in compression output and are not necessary for file-compression logic — they raise a red flag because they could cause an agent to behave autonomously or ignore permission prompts.

[memory-file-access] expected: The SKILL.md and scripts explicitly state the skill reads/writes core workspace files (SOUL.md, AGENTS.md, MEMORY.md). This is expected but sensitive — the behavior is documented but still noteworthy.

What to consider before installing

This skill appears to implement the advertised token-optimization features, but it also reads files in your home directory, writes backups and agent configuration (AGENTS.md / .token-saver-config.json), and includes embedded prompt-like strings that urge 'auto-execute'. Before installing: - Review the compressor.js and analyzer.js files yourself (search for 'Auto-execute', 'SYSTEM:', 'AGENTS.md' and the replacement rules). Those phrases are suspicious and could enable bypassing prompts if the agent is configured to follow file content as instructions. - If you want to try it, run it in a safe test workspace (a copy of your real workspace) or inside a restricted container so it cannot access your real ~/.openclaw or other home files. - Back up AGENTS.md, MEMORY.md, SOUL.md, and any other important .md files before running /optimize tokens or enabling 'Persistent Mode'. Use the provided /optimize revert to restore backups, but verify backups yourself first. - If you install, consider disabling automatic/daemonized invocation and avoid enabling persistent mode until you vet the code. Remove or sanitize any compression rules that inject 'Auto-execute' or similar directives. - Because the publisher/source is unknown and there's no homepage, prefer local inspection and containment; ask the author for provenance and a signed release if you need to use it in production.

Token Saver v3

💡 Did you know? Every API call sends your workspace files (SOUL.md, USER.md, MEMORY.md, AGENTS.md, etc.) along with your message. These files count toward your context window, slowing responses and costing real money on every message.

Token Saver v3 is model-aware — it knows your model's context window and adapts recommendations accordingly. Using Gemini's 1M context? Presets scale up. On GPT-4o's 128K? Presets adjust down.

What's New in v3

Feature	v2	v3
Compaction presets	Fixed (80K/120K/160K)	Dynamic (% of model's context)
Model detection	Fragile, env-only	Robust fallback chain
Context windows	Not tracked	Full registry (9 models)
Model info	Hardcoded pricing	JSON registry, easy updates
Already-optimized	Re-compressed	Smart bypass

Commands

Command	What it does
/optimize	Full dashboard — files, models, context usage %
/optimize tokens	Compress workspace files (auto-backup)
/optimize compaction	Chat compaction control (model-aware)
/optimize compaction balanced	Apply balanced preset (60% of context)
/optimize compaction 120	Custom threshold (compact at 120K)
/optimize models	Detailed model audit with registry
/optimize revert	Restore backups, disable persistent mode

Features

📊 Model-Aware Dashboard

Shows current model, context window, and usage percentage:

🤖 Model: Claude Opus 4.5 (200K context)
   Detected: openclaw.json

📊 Context Usage: [████████░░░░░░░░░░░░] 42% (84K/200K)

📁 Workspace File Compression

Scans all .md files, shows token count and potential savings. Smart bypass skips already-optimized files.

File-aware compression:

• SOUL.md — Light compression, keeps personality language
• AGENTS.md — Medium compression, dense instructions
• USER.md / MEMORY.md — Heavy compression, key:value format
• PROJECTS.md — No compression (user structure preserved)

💬 Dynamic Compaction Presets

Presets adapt to your model's context window:

Preset	% of Context	Claude 200K	GPT-4o 128K	Gemini 1M
Aggressive	40%	80K	51K	400K
Balanced	60%	120K	77K	600K
Conservative	80%	160K	102K	800K
Off	95%	190K	122K	950K

🤖 Model Registry

24+ models with context windows, pricing, and aliases:

• Claude: Opus 4.6 (1M), Opus 4.5, Sonnet 4.5, Sonnet 4, Haiku 4.5, Haiku 3.5 (200K)
• OpenAI: GPT-5.2, GPT-5.1, GPT-5-mini, GPT-5-nano (256K), GPT-4.1, GPT-4o (128K), o1, o3, o4-mini
• Gemini: 3 Pro (2M), 2.5 Pro, 2.0 Flash (1M)
• Others: DeepSeek V3 (64K), Kimi K2.5 (128K), Llama 3.3 70B, Mistral Large

🔍 Robust Model Detection

Detection priority:

1. Runtime injection (--model=...)
2. Environment variables (SKILL_MODEL, OPENCLAW_MODEL)
3. Config file (~/.openclaw/openclaw.json)
4. File inference (TOOLS.md, MEMORY.md mentions)
5. Fallback: Claude Sonnet 4 (safe default)

Unknown model handling:

• Strict version matching — opus-6.5 won't fuzzy-match to opus-4.5
• Unknown models get safe defaults (200K context) + warning
• Easy to add new models to scripts/models.json

📝 Persistent Mode

Adds writing guidance to AGENTS.md for continued token efficiency:

File	Writing Style
SOUL.md	Evocative, personality-shaping
AGENTS.md	Dense instructions, symbols OK
USER.md	Key:value facts
MEMORY.md	Ultra-dense data

Safety

• Auto-backup — All modified files get .backup extension
• Integrity > Size — Never sacrifices meaning for smaller tokens
• Smart bypass — Skips already-optimized files
• Revert anytime — /optimize revert restores everything
• No external calls — All analysis runs locally

Installation

clawhub install token-saver --registry "https://www.clawhub.ai"

Version History

• 3.0.0 — Model registry, dynamic presets, robust detection, smart bypass
• 2.0.1 — Chat compaction, file-aware compression, persistent mode
• 1.0.0 — Initial release