Risk 01
Multi-file edits with no pre-review
Cascade stages changes across many files before the developer sees the diff. Without a response-scan layer, insecure or license-flagged code can land in the editor in bulk.
← Integrations
AI coding
Windsurf
Windsurf, under policy.
Windsurf's Cascade agent plans, edits, and executes across your codebase. Raidu sits between Cascade and the model, redacts outbound context, gates every tool call, scans responses, and records each decision.
The tool
Windsurf
Codeium's agentic IDE with the Cascade agent.
Windsurf is a full IDE, not an extension. Cascade reads your repo, writes code across files, and executes commands. That is the value. That is the governance surface.
Without governance
A full IDE, a full risk surface.
Cascade is not a suggestion engine. It is an agent. Four failure modes security teams see when Cascade ships without a governance layer.
Risk 02
Shell and command execution
Cascade runs scripts, tests, and installs. A misfired command touches infrastructure. Without per-tool policy, the first signal is an incident.
Risk 03
Broad repo context outbound
To plan across files, Cascade reads widely. Without redaction, secrets and business logic leak to whichever provider the team configured. Nothing on your side proves what did or did not leave.
Risk 04
No unified evidence for audit
Windsurf's internal logs are not chained so that changes show, and not designed for SOC 2 or HIPAA evidence packages. Regulators ask for one exportable bundle, not a screenshot of a panel.
With Raidu
How Raidu governs Windsurf.
Windsurf supports custom model endpoints. Point Cascade at Raidu and every plan, tool call, and response flows through the same runtime that governs the rest of your AI stack.
01
Redact outbound context
Checkpoint 02 · Before LLM
Every file Cascade opens, every shell output, every instruction is scanned before it leaves for the provider. Secrets, PII, and flagged internals are deterministically tokenized. The agent still reasons with the data. The data does not leave raw.
02
Gate tool calls with policy
Checkpoint 03 · Before Tool
Allowlist shell commands, path-glob file writes, require approval for destructive operations. Policy is shared with your other coding tools, so Cascade and Cursor follow the same rules.
03
Scan every response
Checkpoint 05 · Agent Response
Model output is checked for unsafe code, license risks, hallucinated packages, and exfiltration patterns before the editor renders the diff. Clean output streams through. Blocked output is logged with a reason.
04
Tamper-evident evidence chain
Post-execution
Every Cascade turn is a tamper-evident record linked to the previous one. Prompts, tools, diffs, responses, policy version, user identity. Hash-chained, write-once, WORM retained. Exportable for SOC 2.
Integration
Two fields in Windsurf settings.
Windsurf supports OpenAI-compatible endpoints for custom models. Point the endpoint and key at Raidu and Cascade is governed from that moment.
Windsurf · settings
json
{
"cascade.modelProvider": "openai",
"cascade.openaiBaseUrl": "https://raidu.your-company.internal/openai",
"cascade.openaiApiKey": "raidu_xxx",
"cascade.model": "claude-sonnet-5-5",
"cascade.customHeaders": {
"x-raidu-policy": "coding.eng.v7"
}
}
// Every Cascade turn returns x-raidu-record-id in the response headers.Questions
Questions engineering leaders ask before rolling out Cascade.
Does Raidu require a Windsurf fork? +
No. Windsurf supports custom OpenAI-compatible endpoints for Cascade. Pointing it at Raidu is a settings change.
Can Raidu require approval on Cascade's multi-file edits? +
Yes. Policy can require approval for changes that match destructive patterns, changes in regulated paths, or any change above a line threshold. Approvals are recorded in the tamper-evident chain.
What is the overhead? +
You choose how deep the checking goes for each use case. Cascade feels the same to the developer; the difference is the visibility.
Does Raidu work with Windsurf's built-in Codeium models? +
Raidu governs model calls that route through the OpenAI-compatible path you configure. For Codeium's hosted-only models without endpoint override, Raidu governs complementary coding tools in the same tenant and enforces unified policy.
Can my auditor verify records without environment access? +
Records stay in your own cloud account. You give your auditor an export of the records with their hash chain, and any change to an entry breaks the chain. Your auditor does not need access to your environment.
Does this work for regulated repos with stricter rules? +
Yes. Raidu policies are per-repo, per-branch, and per-path. A regulated repo can require a specific model, require human approval, block shell execution, and mandate stricter redaction, while other repos stay on the default.
Bring Windsurf into a governed runtime.
Thirty minutes. Sandbox tenant. A real Cascade task with every tool call checked, every message on the record, and a tamper-evident evidence bundle at the end.