Fiello Model Transport Proxy
Fiello Model Transport Proxy runs a loopback proxy inside the VS Code
extension host. It forwards custom language-model requests to one configured
HTTPS endpoint while forcing a fresh HTTP/1.1 TLS connection for every upstream
request.
Why this proxy exists
Some custom Chat Completions and Responses API connections in VS Code fail
intermittently when the HTTP/2 connection is closed by the upstream gateway.
Typical symptoms include GOAWAY, ERR_HTTP2_PING_FAILED, and generic network
errors even though the same endpoint remains available over HTTP/1.1.
This extension works around that transport failure by accepting the request on
localhost and forwarding it to the configured HTTPS endpoint over a fresh
HTTP/1.1 connection. It does not change the model request or provide a model
API. The underlying VS Code behavior is tracked in
microsoft/vscode#335219.
Features
- Starts with VS Code and stops with the extension host.
- Listens only on
127.0.0.1:43129.
- Forces upstream ALPN to
http/1.1.
- Retries transient provider SSE errors with progressive delays before exposing
a failed response to the model client.
- Streams responses and retains each active request body only in memory while
retries are possible.
- Keeps API keys and authorization headers in the calling extension's configuration.
- Shows status, uptime, request counts, active requests, and upstream errors in
a VS Code webview.
- Shares one listener safely across multiple VS Code windows.
- Sends no telemetry.
Set the full upstream model endpoint using any of these methods:
- Click
Fiello Model Transport Proxy in the VS Code status bar, enter the endpoint, and
select Save.
- Run Fiello Model Transport Proxy: Configure Upstream from the
Command Palette.
- Open VS Code Settings and set
customModelHttp1Proxy.upstreamUrl.
The dashboard and VS Code Settings also expose
customModelHttp1Proxy.maxProviderRetries. It accepts 0 through 30 and
defaults to 5. Retry delays grow from 1 second to a 15-second cap; a provider
Retry-After value is honored up to 60 seconds.
The endpoint must:
- use HTTPS;
- include the complete path and query string expected by the model API;
- not contain credentials or a URL fragment.
Example:
https://models.example.com/v1/responses?api-version=2026-01-01
Then configure the custom model client to call the local listener. The local
path can mirror the upstream path for readability:
http://127.0.0.1:43129/v1/responses
The proxy sends every non-health request to the complete configured upstream
URL. Request headers, including authorization headers, are forwarded after
hop-by-hop headers are removed.
Status
Click Fiello Model Transport Proxy in the status bar or run
Fiello Model Transport Proxy: Show Status.
The status view provides:
- runtime state;
- local and upstream endpoints;
- uptime and request counters;
- active request count;
- upstream error count and latest error;
- upstream configuration, VS Code Settings, and restart controls.
The local health endpoint is also available at:
http://127.0.0.1:43129/__health
Install
Visual Studio Marketplace
After publication, install Fiello Model Transport Proxy from the
Extensions view in VS Code.
VSIX
Download or build the .vsix, then use Extensions: Install from VSIX...
from the Command Palette.
The command-line equivalent is:
code --install-extension fiello-model-transport-proxy-0.2.2.vsix
Build
Requirements:
- Node.js 20 or newer;
- VS Code 1.85 or newer.
Run:
npm test
npm run package
npm run package creates the standard versioned VSIX installer in the project directory.
Security
- The listener binds only to IPv4 loopback and is not exposed to the local network.
- Only HTTPS upstream endpoints are accepted.
- User information and credentials are rejected in the configured URL.
- Request bodies are held in memory only for the lifetime of an active request
and are not logged or persisted. Credentials are not logged or persisted.
- The extension does not read custom model configuration files and does not
store API keys.
- Every upstream request uses a new connection with HTTP/1.1 ALPN, avoiding
stale pooled HTTP/2 sessions.
Any local process can connect to a loopback port. Keep secrets in authorization
headers and use the operating system and VS Code security controls appropriate
for your environment.
Limitations
- One upstream endpoint is supported per VS Code user profile.
- The local listener uses fixed port
43129.
- This is a transport workaround, not a replacement for an upstream fix in VS
Code, the calling extension, or the model gateway.
- An in-flight request ends when the VS Code window that owns the listener
closes. Another open VS Code window takes ownership for subsequent requests.
Troubleshooting
Set customModelHttp1Proxy.upstreamUrl to the complete upstream HTTPS endpoint.
blocked
Another process owns port 43129. Stop that process, then run
Fiello Model Transport Proxy: Restart.
502 upstream_request_failed
Open the status view and inspect Last error. Confirm DNS, TLS, firewall,
endpoint path, and gateway availability.
The custom model still uses HTTP/2
Confirm that its configured URL starts with http://127.0.0.1:43129/, then
check that the request counter increases in the status view.
License
MIT
