feat(llm-proxy-openai): new package — Anthropic ↔ OpenAI Chat Completions translator

Lets ComputerAgent target any OpenAI-Chat-Completions–compatible backend
(Lyzr Studio, vLLM, LiteLLM, Together, Ollama, etc.) without engine code
changes. Engines whose SDKs speak Anthropic Messages internally —
claude-agent-sdk and deepagents (via LangChain ChatAnthropic) — only need
ANTHROPIC_BASE_URL pointed at the proxy. gitagent doesn't need this proxy
because gitclaw natively speaks both Anthropic and OpenAI via
GITCLAW_MODEL_BASE_URL + provider:model@baseUrl syntax.

Bidirectional translation covers:
  - request: system field, content blocks (text + tool_use + tool_result),
    tools (Anthropic input_schema → OpenAI function.parameters),
    tool_choice (auto/any/specific-tool)
  - response: text + tool_calls, finish_reason mapping
  - streaming: full Anthropic event sequence (message_start →
    content_block_start/delta/stop → message_delta → message_stop) with
    text_delta + input_json_delta for tool call argument streaming

Exposes both a CLI (npx @computeragent/llm-proxy-openai with env vars) and
a programmatic startProxy() API returning a handle with .close().

Default `max_tokens` is dropped on the upstream request — Lyzr Studio
blanks the response when it's set; opt back in with FORWARD_MAX_TOKENS=1
for backends that need it.

End-to-end verified: claude-agent-sdk + deepagents both wrote real files
via Write/write_file tool calls round-tripped through the proxy → Lyzr →
back. Lives in /tmp during development; this commit promotes it.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: kapaleshreyas

packages/llm-proxy-openai/README.md

@@ -0,0 +1,102 @@
+# @computeragent/llm-proxy-openai
+
+Anthropic Messages ↔ OpenAI Chat Completions **translator proxy**.
+
+Accepts `POST /v1/messages` in Anthropic's Messages format and forwards to any OpenAI-Chat-Completions–compatible endpoint, translating both the request and the response (including streaming and tool calls).
+
+Used by ComputerAgent so that engines whose underlying SDKs speak Anthropic Messages — `claude-agent-sdk` and `deepagents` (via `ChatAnthropic`) — can target OpenAI-compat backends like Lyzr Studio, vLLM, LiteLLM, Together, Ollama, etc.
+
+`gitagent` does **not** need this proxy — gitclaw natively speaks both protocols via `GITCLAW_MODEL_BASE_URL` + `provider:model@baseUrl` syntax.
+
+## Run as a CLI
+
+```bash
+UPSTREAM_BASE=https://your-openai-compat-host \
+UPSTREAM_PATH=/v1/chat/completions \
+UPSTREAM_TOKEN=sk-... \
+PORT=8788 \
+  npx @computeragent/llm-proxy-openai
+```
+
+Then point your ComputerAgent run at the proxy:
+
+```bash
+curl -N -X POST http://127.0.0.1:18800/run \
+  -H 'content-type: application/json' -H 'accept: text/event-stream' \
+  -d '{
+    "source": "github.com/<org>/<gap-repo>",
+    "harness": "claude-agent-sdk",
+    "envs": {
+      "ANTHROPIC_BASE_URL": "http://127.0.0.1:8788",
+      "ANTHROPIC_API_KEY":  "via-proxy"
+    },
+    "message": "Reply: PING"
+  }'
+```
+
+## Run programmatically
+
+```ts
+import { startProxy } from "@computeragent/llm-proxy-openai";
+
+const proxy = await startProxy({
+  port: 8788,
+  upstream: {
+    base: "https://agent-dev.test.studio.lyzr.ai",
+    path: "/v4/chat/completions",
+    token: process.env.LYZR_TOKEN!,
+    modelOverride: "697a4a76496e0831bdde546c", // optional; override client model
+  },
+});
+
+// …
+await proxy.close();
+```
+
+## Environment variables (CLI)
+
+| Var | Required | Default | Notes |
+|---|---|---|---|
+| `UPSTREAM_BASE` | ✓ | — | Origin only, e.g. `https://your-host` |
+| `UPSTREAM_TOKEN` | ✓ | — | Bearer token for the upstream |
+| `UPSTREAM_PATH` | | `/v1/chat/completions` | Some hosts use `/v4/chat/completions` |
+| `UPSTREAM_MODEL` | | — | Force this model on every upstream request (overrides client's Anthropic `model`) |
+| `UPSTREAM_AUTH_SCHEME` | | `Bearer` | Replace with e.g. `Token` if your backend wants something else |
+| `PORT` | | `8788` | Bind port |
+| `FORWARD_MAX_TOKENS` | | `0` | Set to `1` to forward `max_tokens` — some backends (Lyzr) blank the response when this is set, so it's off by default |
+
+## What the proxy translates
+
+| Direction | Anthropic shape | OpenAI shape |
+|---|---|---|
+| Request | `system` field | first `system` message |
+| Request | `messages[].content` text blocks | `messages[].content` strings |
+| Request | assistant `tool_use` blocks | `messages[].tool_calls` |
+| Request | user `tool_result` blocks | `messages[]` with `role: "tool"` + `tool_call_id` |
+| Request | `tools[]` (Anthropic `input_schema`) | `tools[]` (OpenAI `function.parameters`) |
+| Request | `tool_choice: {type:"auto" \| "any" \| "tool"}` | `tool_choice: "auto" \| "required" \| {type:"function",function:{name}}` |
+| Response | `id`, `usage`, content blocks | reassembled from `choices[0].message` |
+| Response | `tool_use` content blocks | from `tool_calls` |
+| Response | `stop_reason: "tool_use" \| "max_tokens" \| "end_turn"` | mapped from `finish_reason` |
+| Streaming | `message_start` → `content_block_start/delta/stop` → `message_delta` → `message_stop` | reassembled from `chat.completion.chunk` deltas (text + `tool_calls.function.arguments` partial JSON) |
+
+## Endpoints
+
+- `GET /health` → `{ ok: true, upstream }`
+- `POST /v1/messages` → translated and forwarded to `<UPSTREAM_BASE><UPSTREAM_PATH>`
+
+Anything else → `404`.
+
+## Limitations
+
+- **Text + tool calls only.** No image inputs, no audio. Anthropic's content blocks for those types are passed through as JSON strings if encountered, which most OpenAI-compat backends will reject.
+- **No retry / circuit-breaker.** A single upstream timeout fails the request.
+- **Stateless.** No request-id correlation, no metrics export. Bring your own observability.
+- **`session_id` is informational.** The proxy adds it to outgoing OpenAI requests for backends that track it (e.g. Lyzr Studio), but it's a string derived from the timestamp — proper session continuity comes from the client replaying `messages[]`, not from server-side state in the upstream.
+
+## Verified compatibility
+
+| Upstream | Backend | Tool calls round-trip? |
+|---|---|---|
+| Lyzr Studio (`/v4/chat/completions`) | claude-agent-sdk | ✓ |
+| Lyzr Studio (`/v4/chat/completions`) | deepagents | ✓ |

packages/llm-proxy-openai/package.json

@@ -0,0 +1,50 @@
+{
+  "name": "@computeragent/llm-proxy-openai",
+  "version": "0.1.0",
+  "description": "Translator proxy: accepts POST /v1/messages (Anthropic Messages format) and forwards to /v1/chat/completions (OpenAI-compat). Lets engines that speak Anthropic Messages (claude-agent-sdk, deepagents) target any OpenAI-Chat-Completions backend — Lyzr Studio, vLLM, LiteLLM, Together, etc.",
+  "license": "MIT",
+  "type": "module",
+  "main": "./dist/index.js",
+  "types": "./dist/index.d.ts",
+  "bin": {
+    "anth-to-openai-proxy": "./dist/cli.js"
+  },
+  "exports": {
+    ".": {
+      "types": "./dist/index.d.ts",
+      "import": "./dist/index.js"
+    }
+  },
+  "files": [
+    "dist",
+    "src",
+    "README.md"
+  ],
+  "scripts": {
+    "build": "tsc -p tsconfig.json",
+    "typecheck": "tsc -p tsconfig.json --noEmit",
+    "test": "vitest run --passWithNoTests",
+    "clean": "rm -rf dist .turbo *.tsbuildinfo",
+    "start": "node dist/cli.js"
+  },
+  "dependencies": {},
+  "devDependencies": {
+    "@types/node": "^22.0.0",
+    "typescript": "^5.5.0",
+    "vitest": "^2.0.0"
+  },
+  "keywords": [
+    "anthropic",
+    "openai",
+    "proxy",
+    "llm",
+    "computeragent",
+    "lyzr"
+  ],
+  "homepage": "https://github.com/open-gitagent/ComputerAgent",
+  "repository": {
+    "type": "git",
+    "url": "https://github.com/open-gitagent/ComputerAgent.git",
+    "directory": "packages/llm-proxy-openai"
+  }
+}

packages/llm-proxy-openai/src/cli.ts

@@ -0,0 +1,49 @@
+#!/usr/bin/env node
+/**
+ * CLI for `@computeragent/llm-proxy-openai`.
+ *
+ * Wires env vars to programmatic options + handles SIGINT/SIGTERM cleanly.
+ *
+ *   UPSTREAM_BASE=https://your-host \
+ *   UPSTREAM_PATH=/v1/chat/completions \
+ *   UPSTREAM_TOKEN=sk-... \
+ *   UPSTREAM_MODEL=optional-model-override \
+ *   PORT=8788 \
+ *     anth-to-openai-proxy
+ *
+ * Or via `npx @computeragent/llm-proxy-openai`.
+ */
+import { startProxy } from "./proxy.js";
+
+const port = Number(process.env.PORT ?? 8788);
+const upstreamBase = process.env.UPSTREAM_BASE;
+const upstreamToken = process.env.UPSTREAM_TOKEN;
+
+if (!upstreamBase) {
+  console.error("ERROR: UPSTREAM_BASE is required (e.g. https://agent-dev.test.studio.lyzr.ai)");
+  process.exit(2);
+}
+if (!upstreamToken) {
+  console.error("ERROR: UPSTREAM_TOKEN is required (Bearer token for the upstream)");
+  process.exit(2);
+}
+
+const handle = await startProxy({
+  port,
+  upstream: {
+    base: upstreamBase,
+    path: process.env.UPSTREAM_PATH,
+    token: upstreamToken,
+    ...(process.env.UPSTREAM_MODEL ? { modelOverride: process.env.UPSTREAM_MODEL } : {}),
+    ...(process.env.UPSTREAM_AUTH_SCHEME ? { authScheme: process.env.UPSTREAM_AUTH_SCHEME } : {}),
+  },
+  forwardMaxTokens: process.env.FORWARD_MAX_TOKENS === "1",
+});
+
+const shutdown = async (sig: string) => {
+  console.error(`[proxy] received ${sig}, shutting down`);
+  await handle.close().catch(() => {});
+  process.exit(0);
+};
+process.on("SIGINT", () => void shutdown("SIGINT"));
+process.on("SIGTERM", () => void shutdown("SIGTERM"));

packages/llm-proxy-openai/src/index.ts

@@ -0,0 +1,42 @@
+/**
+ * `@computeragent/llm-proxy-openai` — Anthropic Messages ↔ OpenAI Chat
+ * Completions translator. Accepts the Anthropic Messages wire shape that
+ * `claude-agent-sdk` and `deepagents` speak natively, and forwards to any
+ * OpenAI-Chat-Completions endpoint (Lyzr Studio, vLLM, LiteLLM, Together,
+ * Ollama, etc.).
+ *
+ * Two ways to use:
+ *
+ *   # CLI:
+ *   UPSTREAM_BASE=https://your-openai-compat-host \
+ *   UPSTREAM_PATH=/v1/chat/completions \
+ *   UPSTREAM_TOKEN=sk-... \
+ *   PORT=8788 \
+ *     npx anth-to-openai-proxy
+ *
+ *   # Programmatic:
+ *   import { startProxy } from "@computeragent/llm-proxy-openai";
+ *   const proxy = await startProxy({
+ *     port: 8788,
+ *     upstream: {
+ *       base: "https://your-openai-compat-host",
+ *       path: "/v1/chat/completions",
+ *       token: "sk-...",
+ *     },
+ *   });
+ *   // … later:
+ *   await proxy.close();
+ *
+ * Then point any ComputerAgent run at the proxy:
+ *
+ *   envs: {
+ *     ANTHROPIC_BASE_URL: "http://127.0.0.1:8788",
+ *     ANTHROPIC_API_KEY:  "via-proxy",  // anything; the upstream token is in the proxy
+ *   }
+ *
+ * Tool calls round-trip in both directions — verified end-to-end against
+ * Lyzr Studio with claude-agent-sdk + deepagents both writing real files
+ * through their tool surfaces.
+ */
+export { startProxy } from "./proxy.js";
+export type { ProxyOptions, ProxyHandle, UpstreamConfig } from "./proxy.js";