Tokenminning in Windsurf
Windsurf (Devin Desktop) re-sends indexed snippets, pinned files, rules, MCP schemas, and Cascade history every turn. Waste shows up as long sessions, frontier defaults, and always_on rules.
Hub for the Windsurf cluster — quota, Cascade, rules, and MCP. Cognition ecosystem overlap: Devin. For the general stack see Where to start.
Quick checklist
- Check the usage meter — Quota & billing.
- Route routine work to SWE-1.5 (free on quota plans).
- Prefer
glob/manualrules overalways_on. - Disable MCP you are not using (100-tool cap).
- New Cascade conversation per task; trim
.codeiumignore.
Typical impact: 50–80% quota savings on SWE routing; 20–50% trimming rules and MCP; 30–60% less context growth from shorter sessions.
How Windsurf bills
Quota-based plans refresh daily/weekly allowances. Token cost varies by model; SWE-1.5 is free on current self-serve plans. Enterprise may use ACUs — see Quota guide.
Cascade sends: prompt, RAG snippets, pins, rules, MCP schemas, conversation history. Command and Tab are separate from Cascade quota.
Guides
| Guide | What you’ll learn |
|---|---|
| Quota & billing | Daily/weekly quota, cache reads, enterprise ACUs |
| Cascade | Sessions, pinning, .codeiumignore, @ mentions |
| Rules | Activation modes, caps, AGENTS.md |
| MCP | Disable servers, schema overhead, 100-tool limit |
Model routing
| Tier | Windsurf choice |
|---|---|
| Routine | SWE-1.5 / SWE-1.6 (free) |
| Multi-file | Sonnet / mid-tier GPT |
| Deep debug | Opus / thinking — last resort |
Stay on one frontier model per task so prompt caching applies.
Related
- IDEs
- Devin — shared Cognition ecosystem patterns
- Windsurf Cascade docs