Dashboard & Settings
Context & Prompt Budget
Live input-token estimator across system prompts, Memory/Wiki snapshots, Skills, Tools, and MCP schemas.
Dashboard → Context (/dashboard/context) gives you a live, byte-accurate breakdown of what is consuming your model's input context window before you even send a message.
How to Use the Context Inspector
- Knowledge Level Slider (
Low·Medium·High):- Controls how much background context and tool metadata is injected by default. Prefer High for complex multi-step research, Medium for everyday tasks, and Low for fast, lightweight chats.
- Specialist Preview Selector:
- Pick
None (defaults)to see the token footprint of an un-scoped session, or select any Specialist (e.g., Otium, Soren, Fast Computational Scientist) to see how scoping tools, skills, and MCP servers to that Specialist reduces the input token footprint.
- Pick
- Tokenizer Switch (
tiktokenvs.Provider):- Compare local
tiktoken(o200k_base) counts against provider-reported token counts.
- Compare local
Context Bucket Breakdown
The horizontal distribution bar and table break down your session-start context into six buckets:
| Bucket | What it contains | How to optimize |
|---|---|---|
System prompt (USER + BASE) | Sandwalk base instructions, Deep Agents base prompt, and your frozen Memory, Wiki TOC & Ontology snapshot (USER.md + MEMORY.md + PRIORS.md) | Bounded automatically (~3.5k chars max) so it stays around ~1–1.5k tokens and keeps KV prefix caches warm. |
| Deep Agents middleware prompts | Lightweight planning and runtime coordination rules | Fixed minimal overhead (~25 tokens). |
| Domain / runtime tool schemas | JSON schemas for enabled Core and Extra Sandwalk tools | Keep extra tools disabled globally in Dashboard → Tools and enable them per Specialist. |
| MCP tool schemas | Tool definitions exported by active MCP servers in ~/.sandwalk/mcp.json | Large MCP catalogs (300+ tools) can consume 100k+ tokens if all enabled at once. Toggle unused MCPs off in Dashboard → MCPs or scope MCPs to specific Specialists. |
| Deep Agents builtin tool schemas | Built-in file/todo/execution primitives | Minimal fixed footprint (~700 tokens). |
| Skills index (progressive disclosure) | Compact name + 1-line description list of installed skills | Only the compact index (~2k tokens) is in the prompt; full SKILL.md bodies load on demand when invoked. |
Last updated on