copilot-api-node20 0.44.0-preview.9 → 0.44.0

package/README.md

@@ -9,7 +9,7 @@
 
 A proxy server that turns a GitHub Copilot subscription into fully compatible OpenAI and Anthropic API endpoints. Use Claude Code, Claude Cowork, Codex CLI, and any tool that speaks the API -- all through the Copilot plan you already pay for, no separate API keys required.
 
-Your organisation runs its models in **Microsoft Foundry / Azure OpenAI** instead? The same proxy fronts your Azure deployments with `--provider foundry`. See [Microsoft Foundry](#microsoft-foundry--azure-openai).
+Want Claude Code on GPT and other non-Claude models? Pick one from your Copilot plan, or serve your own **Microsoft Foundry / Azure OpenAI** deployments with `--provider foundry`, signing in with an API key or Entra ID. Add `--fallback copilot` and Copilot answers whenever Foundry throttles, fails or times out. See [Microsoft Foundry](#microsoft-foundry--azure-openai).
 
 ```
   Clients                        Proxy                         Backends
@@ -59,7 +59,9 @@ Your organisation runs its models in **Microsoft Foundry / Azure OpenAI** instea
 | **Full Claude Code support** | All Claude models, streaming, extended thinking, and the complete **1M token context window** |
 | **Claude Cowork support** | Run Claude's desktop app through Copilot — **one-command setup** via [Tailscale](https://tailscale.com) for zero-trust-hassle HTTPS |
 | **Full Codex CLI support** | Complete **OpenAI Responses API** implementation purpose-built for Codex CLI |
-| **Microsoft Foundry / Azure OpenAI** | Serve Claude Code and Codex from your own Azure deployments (GPT, Claude, and others) with `--provider foundry`: no GitHub account needed |
+| **Claude Code on non-Claude models** | GPT and other models from Copilot or Foundry, translated to the Anthropic API, with tools, streaming and keepalive pings while the model thinks |
+| **Microsoft Foundry / Azure OpenAI** | Your own Azure deployments with `--provider foundry`: API key or Entra ID (`az login`, managed identity, service principal), no GitHub account needed |
+| **Copilot fallback** | `--fallback copilot`: Copilot answers when Foundry rate-limits, errors or times out, and finishes a Claude Code turn on an OpenAI deployment whose stream fails before any output; then traffic moves back to Foundry |
 | **Web search** | Optional web search for Claude Code and Codex sessions, via **Tavily and/or Exa** with automatic failover |
 | **Dual API compatibility** | OpenAI Chat Completions, Anthropic Messages, and OpenAI Responses APIs side by side |
 | **HTTPS with auto-generated certs** | Self-signed TLS certs with configurable SANs -- works across VMs, containers, and LAN |
@@ -99,13 +101,13 @@ npx copilot-api-node20@latest start \
 Start the proxy with Codex mode:
 
 ```bash
-npx copilot-api-node20@latest start --codex --model gpt-5.3-codex
+npx copilot-api-node20@latest start --codex --model gpt-5.6-sol
 ```
 
 Configure `~/.codex/config.toml`:
 
 ```toml
-model = "gpt-5.5"
+model = "gpt-5.6-sol"
 model_provider = "local"
 
 [model_providers.local]
@@ -186,7 +188,7 @@ Point the proxy at your own Azure resource instead of GitHub Copilot. No GitHub
 
 ```bash
 export FOUNDRY_ENDPOINT=https://<resource>.openai.azure.com
-export FOUNDRY_API_KEY=<key>
+export FOUNDRY_API_KEY=<key>      # or omit it and `az login` for Entra ID
 npx copilot-api-node20@latest start --provider foundry
 ```
 
@@ -206,7 +208,9 @@ The model you ask for is the model you get: the proxy never renames or substitut
 | OpenAI (`gpt-*`, `o*`, codex) | `/openai/v1/responses` (translated for Claude Code) |
 | Others (DeepSeek, Grok, …) | `/openai/v1/chat/completions` |
 
-Hitting Foundry's tokens-per-minute limit? Add `--fallback copilot` and a request Foundry rate-limits (429) is served by your GitHub Copilot subscription instead, until Foundry's `retry-after` passes. See [Copilot fallback](./docs/foundry.md#copilot-fallback-on-rate-limits).
+**Sign-in:** by default, with `FOUNDRY_API_KEY` set the proxy uses the key. Without it, it uses Entra ID through the Azure SDK, so `az login`, a managed identity or a service principal all work (`FOUNDRY_AUTH=key|entra` picks one explicitly). The identity needs **Cognitive Services OpenAI User** (GPT) or **Azure AI User** (Claude) on the resource. See [Authentication](./docs/foundry.md#authentication).
+
+**Have a Copilot subscription too?** Add `--fallback copilot`. When Foundry rate-limits (429), fails (408, 500, 502, 503 or 504), times out or drops the connection, Copilot answers the same request. For Claude Code on an OpenAI deployment, a Foundry stream that fails before any text or tool call reaches it is finished by Copilot as well. The model stays on Copilot for Foundry's `retry-after` (initially 10 seconds without one), then goes back. See [Copilot fallback](./docs/foundry.md#copilot-fallback).
 
 Codex works with the usual `config.toml` (see [Codex CLI](#codex-cli)) with `model` set to a deployment name. See **[docs/foundry.md](./docs/foundry.md)** for configuration, routing details, reasoning effort, and troubleshooting (including `NODE_EXTRA_CA_CERTS` behind TLS-inspecting corporate proxies).
 
@@ -250,15 +254,22 @@ curl http://localhost:4141/v1/chat/completions \
 
 All endpoints are served at `http://localhost:4141` by default (`https://` when `--https` or `--claude-cowork` is used). The server binds to `127.0.0.1` (localhost only) by default. To expose to LAN/network/Docker, pass `--hostname 0.0.0.0`.
 
-| Method | Path | Compatibility |
+| Method | Path | Purpose |
 |:--|:--|:--|
+| `GET` | `/` | Basic server-running response |
+| `GET` | `/health` | Proxy, token, and upstream health status |
 | `POST` | `/v1/chat/completions` | OpenAI Chat Completions |
+| `POST` | `/chat/completions` | OpenAI Chat Completions without the `/v1` prefix |
 | `POST` | `/v1/responses` | OpenAI Responses API |
+| `POST` | `/responses` | OpenAI Responses API without the `/v1` prefix |
 | `POST` | `/v1/messages` | Anthropic Messages |
 | `POST` | `/v1/messages/count_tokens` | Anthropic Token Counting |
 | `POST` | `/v1/embeddings` | OpenAI Embeddings |
+| `POST` | `/embeddings` | OpenAI Embeddings without the `/v1` prefix |
 | `GET` | `/v1/models` | OpenAI Models |
-All OpenAI-compatible routes are also available without the `/v1/` prefix.
+| `GET` | `/models` | OpenAI Models without the `/v1` prefix |
+| `GET` | `/usage` | GitHub Copilot usage and quota information |
+| `GET` | `/token` | Copilot token debug endpoint; disabled unless started with `--show-token` |
 
 ---
 
@@ -287,19 +298,23 @@ All OpenAI-compatible routes are also available without the `/v1/` prefix.
 
 | Flag | Alias | Default | Description |
 |:--|:--|:--|:--|
-| `--provider` | | `copilot` | Upstream: `copilot`, or `foundry` for Microsoft Foundry / Azure OpenAI (reads `FOUNDRY_ENDPOINT` / `FOUNDRY_API_KEY`; see [docs/foundry.md](./docs/foundry.md)). Also `COPILOT_API_PROVIDER` |
-| `--fallback` | | | With `--provider foundry`: `copilot` re-sends a Foundry 429 to GitHub Copilot and keeps that model there until Foundry's `retry-after` passes (see [docs/foundry.md](./docs/foundry.md#copilot-fallback-on-rate-limits)). Also `COPILOT_API_FALLBACK` |
+| `--provider` | | `copilot` | Upstream: `copilot`, or `foundry` for Microsoft Foundry / Azure OpenAI (reads `FOUNDRY_ENDPOINT`, plus `FOUNDRY_API_KEY` or an Entra ID login; see [docs/foundry.md](./docs/foundry.md)). Also `COPILOT_API_PROVIDER` |
+| `--fallback` | | | With `--provider foundry`: `copilot` answers from GitHub Copilot when Foundry rate-limits, fails or times out (and, for Claude Code on an OpenAI deployment, when a Foundry stream fails before any output), and keeps that model there for Foundry's `retry-after`, or initially 10 s without one (see [docs/foundry.md](./docs/foundry.md#copilot-fallback)). Also `COPILOT_API_FALLBACK` |
 | `--port` | `-p` | `4141` | Port to listen on |
 | `--hostname` | `-H` | `127.0.0.1` | Hostname/IP to bind to. Use `0.0.0.0` for LAN/Docker exposure |
 | `--claude-code` | `-c` | | Enable Claude Code compatibility mode |
 | `--codex` | | | Enable Codex CLI compatibility mode |
 | `--model` | `-m` | | Primary model to use |
 | `--small-model` | `-s` | | Lightweight model for fast tasks (Claude Code) |
+| `--manual` | | `false` | Require manual approval before forwarding requests |
 | `--rate-limit` | `-r` | | Minimum seconds between requests |
-| `--timeout` | `-t` | `600000` | Request timeout in ms (default 10 min) |
+| `--wait` | `-w` | `false` | Wait instead of returning an error when the configured rate limit is hit |
+| `--timeout` | `-t` | `600000` | Request timeout in ms (default 10 min). A streamed Claude Code turn on a non-Claude model waits at most 4 min for the upstream to start answering |
+| `--connect-timeout` | | `15000` | Connectivity probe timeout in ms; overrides `COPILOT_CONNECT_TIMEOUT_MS` |
 | `--account-type` | `-a` | `individual` | `individual`, `business`, or `enterprise` |
 | `--github-token` | `-g` | | Provide a GitHub token directly (skips OAuth) |
 | `--show-token` | | | Display tokens on fetch and refresh, and enable the local `/token` debug endpoint |
+| `--error-reporting` | | `true` | Send sanitized server-error reports. Disable with `--no-error-reporting` or `COPILOT_API_ERROR_REPORTING=0` |
 | `--smoke-test` | | `true` | Run one self-test generation after startup and print PASS/FAIL. Disable with `--no-smoke-test` or `COPILOT_API_NO_SMOKE_TEST=1` |
 | `--tavily-api-key` | | | **Deprecated** — prefer `TAVILY_API_KEY` env var. Tavily API key for web search |
 | `--web-search-provider` | | | Web search providers in primary→fallback order, e.g. `exa,tavily` (keys via `TAVILY_API_KEY` / `EXA_API_KEY`) |