Windsurf Cascade
Windsurf Cascade is the agentic chat surface that bills quota (or ACUs). Long threads, wide indexing, and over-pinning inflate every turn.
Hub: Tokenminning in Windsurf.
New conversation per task
Every Cascade message includes prior context. Start fresh when:
- You finish one task and begin another
- You switch models (cache miss)
- Cascade loops on a stuck problem
Matches GSC interest in quick-checklist and session hygiene on /ides/windsurf.
Indexing surface
Windsurf indexes the workspace and retrieves via RAG (Context Awareness ).
# .codeiumignore (gitignore syntax)
node_modules/
dist/
.next/
*.lockEnterprise: optional global ~/.codeium/.codeiumignore. Narrow the indexed surface before chasing model changes.
Context pinning
Pin only what retrieval misses :
| Pin type | Risk |
|---|---|
| Single file / function | Low |
| Whole package directory | High token load every turn |
Custom Chat Instructions — keep brief. Close unrelated tabs so Active Document focus stays lean.
@ mentions
Prefer deterministic context:
@filename,@function,@directory,@diff⌘⏎to force codebase context only when needed
See Context hygiene for just-in-time retrieval patterns.
Fast Context
Fast Context uses SWE-grep subagents for faster, narrower retrieval — still depends on .codeiumignore hygiene.
Match model to task
GSC fragment match-the-model-to-the-task — route routine work to free SWE models; see hub model routing and Quota.