Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
Workspace Files Controller
Per-SharedFileSystem HTTP file API. Each SharedFileSystem (SFS) gets its own workspace-files-controller (wfc) instance that mounts the SFS PVC at /workspace and exposes list/stat/download plus upload/replace/mkdir/move/delete over HTTP. It is the write path for a team workspace: host-context-controller mounts the same PVC read-only into the mcp-host pods that reference it, so agents read the workspace but never mutate it — humans and admins write through this service (via control-api).
How It Works
- host-context-controller's SharedFileSystem reconciler provisions, per SFS: a PVC (
sfs-<hash>-files, default20Gi, default access modeReadWriteOnce), a single-replicaRecreateDeployment (wfc-<hash>), a ClusterIP Service on the wfc port, and ingress/egress NetworkPolicies. AllWSF_*env below is injected by that reconciler, exceptWSF_JWT_ISSUERandWSF_JWT_AUDIENCE, which are never injected — the service uses its in-code defaults. - The wfc pod's root init container seeds
spec.directories[]and chowns the fresh PVC to the configured UID/GID (default 1000:1000); the serving container then runs non-root with a read-only root filesystem, all capabilities dropped, andautomountServiceAccountToken: false. - With the default RWO PVC, the wfc is single-node; mcp-host consumer pods co-locate with it via podAffinity. Multi-node
ReadWriteManyrequires an RWX-capable StorageClass set explicitly viaspec.accessModes. - Stack: Express 5, multer (in-memory multipart uploads), jose (JWT verification), pino/pino-http logging.
API
All non-streaming responses use the envelope { ok: true, data } / { ok: false, error: { code, message } } with stable error.code values.
| Route | Scope | Behavior |
|---|---|---|
GET /v1/files?path=<rel> |
files:read |
List directory entries (capped, truncated flag) |
GET /v1/files/stat?path=<rel> |
files:read |
File/directory metadata |
GET /v1/files/download?path=<rel> |
files:read |
Stream file contents as attachment |
POST /v1/files/upload (multipart) |
files:write |
Create new file — 409 already_exists if present |
PUT /v1/files/replace (multipart) |
files:write |
Whole-file replace — 404 if missing |
POST /v1/files/mkdir {path} |
files:write |
mkdir -p; idempotent on existing directories |
POST /v1/files/move {from, to} |
files:write |
Rename within the mount; refuses to clobber |
DELETE /v1/files?path=<rel>&recursive |
files:write |
Delete file or directory (409 not_empty without recursive=true) |
GET /healthz, GET /readyz |
none | Liveness / readiness (mount present and writable) |
/healthz and /readyz are registered before auth — kubelet probes are unauthenticated. Everything under /v1 requires a valid browsing JWT.
Configuration
| Variable | Explanation | Default |
|---|---|---|
WSF_PORT |
HTTP listen port. | 8086 |
WSF_MOUNT_PATH |
Absolute path of the mounted SFS PVC. | /workspace |
WSF_SHARED_FILESYSTEM_NAME |
SFS identity this instance serves; the JWT sharedFileSystem claim must match. |
required |
WSF_SHARED_FILESYSTEM_NAMESPACE |
SFS namespace; the JWT sharedFileSystemNamespace claim must match. |
required |
WSF_JWT_PUBLIC_KEY |
PEM-encoded RSA public key for verifying browsing JWTs. | required |
WSF_JWT_ISSUER |
Required JWT iss claim. |
control-api |
WSF_JWT_AUDIENCE |
Required JWT aud claim. |
workspace-files-controller |
WSF_MAX_UPLOAD_BYTES |
Per-file upload cap; larger uploads get 413 payload_too_large. |
104857600 (100 MiB) |
WSF_MAX_LIST_ENTRIES |
Max entries returned by a single directory listing. | 5000 |
WSF_MAX_PATH_DEPTH |
Max path depth (segments) accepted from clients. | 32 |
Security Model
Auth. Browsing JWTs are short-lived RS256 tokens minted by control-api (signed with the same keypair as RPC tokens; audience separation keeps the token kinds apart). The wfc rejects any token unless all of: iss matches, aud matches, sharedFileSystem and sharedFileSystemNamespace claims equal this instance's SFS, and scopes contains only entries from the closed set files:read / files:write (unknown or duplicate scopes are rejected outright). Scopes are then enforced per route. Audience or SFS mismatch is 403; signature/expiry failures are 401.
Path safety. Every client path is validated lexically (reject .., backslash, ASCII control chars, excess depth; resolve under the mount root) and physically (each existing segment is lstated — symlinks are rejected or omitted everywhere: directory listings silently skip symlink entries, and every other operation (listing roots, move sources, delete targets, …) hard-rejects with path_invalid — with a realpath containment check). Writes re-validate the chain after mkdir/rename to defend against TOCTOU symlink swaps, uploads go to a temp sibling then rename atomically so partial writes never appear at the destination, and the mount root itself can never be written, replaced, moved, or deleted.
Network. The HCC-generated NetworkPolicies allow ingress only from control-api pods on the wfc port, and egress to DNS only — a compromised mcp-host pod cannot reach the writable side of its own workspace.
Development
npm install
WSF_SHARED_FILESYSTEM_NAME=dev WSF_SHARED_FILESYSTEM_NAMESPACE=mcp-host \
WSF_JWT_PUBLIC_KEY="$(cat public.pem)" npm run dev
Testing
npm test
Five vitest suites cover config loading, JWT verification, path safety, and the read/write route behavior (via supertest against the composed Express app).
Deployment
There are no static manifests for this service: instances are created dynamically by host-context-controller when a SharedFileSystem CRD is applied. The image, port, JWT public key ConfigMap, and WSF_MAX_* limits are configured on HCC via CONTEXT_MAPPER_WFC_* env (see deploy/base/control-plane/host-context-controller.yaml and host-context-controller/src/config.ts). See the SharedFileSystem CRD reference for the resource that drives it.