diff --git a/README.md b/README.md
index 68cccb2..56197fd 100644
--- a/README.md
+++ b/README.md
@@ -15,6 +15,7 @@ Forked from [nasrulhadi/proxy-commandcode](https://github.com/nasrulhadi/proxy-c
 | Usage | `finish-step`/`finish` usage mapped to OpenAI `usage` (sync body; extra stream chunk with empty `choices`) |
 | Finish reason | `length` / `tool_calls` / `content_filter` mapped from the upstream reason |
 | Request shape | Tool-call `input` sent as an object; consecutive tool results merged; `developer` role folded into `system`; image parts use `{image, mediaType}`; `temperature` and `max_completion_tokens` passed through; `x-session-id` and `x-cli-environment` headers sent |
+| System prompt | A request with no (or a blank) system message gets a one-line default (`PCMC_DEFAULT_SYSTEM`, default `You are a helpful assistant.`); otherwise the gateway injects the ~7.3K-token `cmd` agent prompt on every call |
 | CLI version | `x-command-code-version` tracks `command-code@latest` on npm (checked at start and every 6 h, default `1.66.0`); set `PCMC_VERSION` to pin |
 | Runtime | Logs to stdout only (no `proxy.log`, no ANSI without a TTY); Windows `netstat`/`taskkill` port handling removed; graceful `SIGTERM`; `Dockerfile` + `npm test` |
 
diff --git a/server.js b/server.js
index 98ad2b7..32a9b8b 100644
--- a/server.js
+++ b/server.js
@@ -19,6 +19,11 @@ const DEFAULT_CC_VERSION = '1.66.0';
 const PINNED_VERSION = process.env.PCMC_VERSION || '';
 const VERSION_REFRESH_MS = 6 * 60 * 60 * 1000;
 const DEBUG = process.env.PCMC_DEBUG === '1';
+// With no system prompt the gateway substitutes the ~7.3K-token `cmd` agent
+// prompt (CLI tools, commit rules, taste file), which costs tokens on every call
+// and makes the model act as a coding agent. A blank system is rejected by some
+// models, so always send a real one.
+const DEFAULT_SYSTEM = (process.env.PCMC_DEFAULT_SYSTEM || '').trim() || 'You are a helpful assistant.';
 
 let ccVersion = PINNED_VERSION || DEFAULT_CC_VERSION;
 
@@ -135,7 +140,7 @@ function transform(oaiBody) {
   }));
 
   const params = {
-    model, system: systemText || undefined, messages, tools: tools.length > 0 ? tools : undefined,
+    model, system: systemText.trim() ? systemText : DEFAULT_SYSTEM, messages, tools: tools.length > 0 ? tools : undefined,
     max_tokens: oaiBody.max_completion_tokens || oaiBody.max_tokens || 32000,
     // The gateway rejects stream:false; always stream upstream and buffer here.
     stream: true,
diff --git a/test/transform.test.js b/test/transform.test.js
index e989fbb..9077728 100644
--- a/test/transform.test.js
+++ b/test/transform.test.js
@@ -12,6 +12,16 @@ test('builds the strict envelope and always streams upstream', () => {
   assert.deepEqual(body.params.messages, [{ role: 'user', content: [{ type: 'text', text: 'hi' }] }]);
 });
 
+test('sends a default system prompt when none or a blank one is given', () => {
+  for (const messages of [
+    [{ role: 'user', content: 'q' }],
+    [{ role: 'system', content: '  ' }, { role: 'user', content: 'q' }],
+  ]) {
+    const body = JSON.parse(transform({ messages }));
+    assert.equal(body.params.system, 'You are a helpful assistant.');
+  }
+});
+
 test('maps system/developer to params.system', () => {
   const body = JSON.parse(transform({ messages: [
     { role: 'system', content: 'a' }, { role: 'developer', content: [{ type: 'text', text: 'b' }] }, { role: 'user', content: 'q' },