Add flyte proxy app: authenticated localhost proxy into a Flyte App

[mhotan]

Overview

Adds a flyte proxy app NAME command — an authenticated localhost reverse proxy into a no-auth-inside Flyte App. It reuses the CLI's own authenticator to inject a fresh bearer on every request, so a local HTTP client (an MCP server, curl, a browser) can reach an app-edge-gated app with no token handling. Think kubectl port-forward for Flyte Apps.

Details

• Resolves the target by app name (App.get) or an explicit --url.
• Binds loopback on both 127.0.0.1 and ::1 so localhost — which resolves to ::1 first on macOS — can't be silently shadowed by a wildcard listener from another process (e.g. a container runtime binding *:PORT with SO_REUSEPORT). Previously an IPv4-only bind still "succeeded" while every request landed on the other listener and reset; now a genuine conflict fails loudly with EADDRINUSE. A non-loopback --address binds a single family and warns (the proxy is ambient authority for the caller's identity).
• Streams responses unbuffered and preserves Content-Encoding — SSE-safe for MCP streamable-HTTP.
• Drops inbound Authorization/hop-by-hop headers, injects the fresh bearer on the standard authorization header (the SDK auth flow may emit it under a gRPC-style metadata key the HTTP app edge never reads), and refreshes + retries once when the edge bounces an expired token to /login.
• --emit-mcp-config prints a generic HTTP-MCP client block pointed at the local port.
• WebSocket upgrade proxying is deferred to a follow-up.

Also included

• App-serde: materialize delayed parameters on the pod_template path. When an app parameter is a delayed value (e.g. RunOutput), it is materialized to a resolved File/Dir URI at deploy time. The container path already did this; the pod_template serialization path did not — so an app that set a pod_template serialized the raw, unresolved getter and the serve entrypoint couldn't download it. Threads parameter_overrides through _get_k8s_pod / _serialized_pod_spec so the pod command carries the resolved value. Independent of the proxy — can be split into its own PR on request.

Testing

• Unit tests (tests/cli/test_proxy.py): inbound-auth/hop-by-hop stripping, Content-Encoding preservation, JWT identity display, and the MCP-config block.
• Validated end-to-end against a running app with an upstream mcp-grafana server driven by the official MCP Python client SDK (ClientSession): initialize -> 65 tools -> list_datasources and query_prometheus returned live data through the proxy with no real credential in the MCP config.
• Dual-stack bind verified: with a wildcard listener holding the port, curl localhost:PORT reaches the proxy (not the other listener); a conflicting specific bind fails with EADDRINUSE.

🤖 Generated with Claude Code

[mhotan]

This is actually a bug fix I think I found.

[kumare3]

can we import lazily?

[mhotan]

Done in 39fd218 — moved import flyte.remote as remote into the app command's name-resolution branch, so the heavy remote client stack only loads when resolving an app by name (the --url path skips it entirely). Module top now imports only stdlib + rich_click (needed for the decorators).