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
INLINECODE0	Full dashboard — files, models, context usage %
INLINECODE1

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: CODEBLOCK0

📁 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. 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 INLINECODE13
• Unknown models get safe defaults (200K context) + warning
• Easy to add new models to INLINECODE14

📝 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

CODEBLOCK1

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