openclaw - 💡(How to fix) Fix feat: Per-request workspace override via HTTP header (x-openclaw-workspace) [2 comments, 2 participants]

🚀 Feature Request

Summary

Add support for a per-request workspace override via an HTTP header (x-openclaw-workspace) on the /v1/chat/completions endpoint.

Problem / Motivation

When deploying OpenClaw as a multi-tenant SaaS backend (e.g., serving thousands of merchants/users from a shared cluster), each tenant needs isolated memory, workspace files, and context — but the current architecture only supports a static workspace path configured per agent in openclaw.json.

This forces one of two unacceptable tradeoffs:

1. One OC instance per tenant — prohibitively expensive at scale (10k+ tenants × N instances each)
2. Shared workspace for all tenants — breaks memory isolation entirely

There is no way today to tell OpenClaw "for this specific request, use workspace /efs/tenants/tenant_001" without modifying agent config and restarting.

Proposed Solution

Support a new HTTP header on /v1/chat/completions:

x-openclaw-workspace: /path/to/tenant/workspace

When this header is present, OpenClaw uses the specified path as the workspace for that request only — overriding the agent's configured workspaceDir for the duration of the run.

This allows the router layer (sitting in front of the OC cluster) to:

1. Map tenant_id → workspace path (e.g., on EFS/NFS)
2. Inject the path via header on each request
3. Let OC handle memory reads/writes in the correct isolated directory

Example Usage

curl -X POST http://openclaw-host:18789/v1/chat/completions \
  -H "Authorization: Bearer TOKEN" \
  -H "x-openclaw-workspace: /efs/tenants/tenant_001" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openclaw:main",
    "messages": [{"role": "user", "content": "Analyze my sales data"}]
  }'

Implementation Sketch

The change is minimal (~15 lines across 2 files):

src/gateway/http-utils.ts — add header resolver:

export function resolveWorkspaceOverride(req: IncomingMessage): string | undefined {
  const raw = getHeader(req, "x-openclaw-workspace")?.trim();
  return raw || undefined;
}

src/gateway/openai-http.ts — read header and pass to agentCommand:

const workspaceOverride = resolveWorkspaceOverride(req);
// pass workspaceOverride into buildAgentCommandInput → agentCommand

src/commands/agent.ts — apply override in workspace resolution:

if (opts.workspaceOverride) {
  workspaceDir = opts.workspaceOverride;
}

Security Considerations

• The path should be validated (must be absolute, no path traversal ../)
• Should only be accepted when gateway auth is satisfied (already enforced by existing auth middleware)
• Optionally: add a config flag gateway.http.allowWorkspaceOverride: true to require explicit opt-in

Use Cases

Alternatives Considered

• One agent config per tenant in openclaw.json: Doesn't scale beyond hundreds of agents; hot-reload performance degrades
• Router injects full memory as system prompt: Works but loses OC's native memory_search/memory_get tooling and file write-back

I'm willing to contribute a PR for this. Happy to discuss the implementation approach before coding.