translator: guard nil response body in OpenAI ResponseBody (#2240)

**Description**

A JSON `null` response body decodes into a nil `*ChatCompletionResponse`
without
an error, because the decode target is a `**ChatCompletionResponse`
(`Decode(&resp)`).
Some OpenAI-compatible backends return a `null` body with a 200 status
(e.g. an
upstream that aborts the response after committing headers under load).
In the
non-streaming path `ResponseBody` then dereferenced the nil response at
`resp.Usage.PromptTokens`, panicking the ext-proc with a SIGSEGV (addr
0x80).

`RedactBody` already guards `resp == nil`, so the nil case is known; the
main path
just missed it. This treats a nil response as an empty response: report
zero token
usage and fall back to the request model, mirroring the streaming path's
behavior
when no usage is present, rather than crashing.

A regression test feeds a literal `null` body and asserts no panic and
zero usage.

Per the project's generative-AI policy: this change was prepared with
the
assistance of an AI coding tool (Claude Code). The submitter understands
and takes
full ownership of the change.

**Related Issues/PRs (if applicable)**

Fixes #2273

**Special notes for reviewers (if applicable)**

N/A

Signed-off-by: Kunwar Srivastav <kshivamsrivastav@gmail.com>
Co-authored-by: Ignasi Barrera <ignasi@tetrate.io>

Author: kunwarshivam

internal/translator/openai_openai.go

@@ -147,6 +147,14 @@ func (o *openAIToOpenAITranslatorV1ChatCompletion) ResponseBody(_ map[string]str
 	if err := json.NewDecoder(body).Decode(&resp); err != nil {
 		return nil, nil, tokenUsage, responseModel, fmt.Errorf("failed to unmarshal body: %w", err)
 	}
+	// A JSON `null` body decodes into a nil *resp without an error (the decode
+	// target is a **ChatCompletionResponse), which some upstreams return with a
+	// 200 status. Treat it as an empty response so we report zero usage and fall
+	// back to the request model, mirroring the streaming path, instead of
+	// dereferencing nil below.
+	if resp == nil {
+		resp = &openai.ChatCompletionResponse{}
+	}
 
 	// Redact and log response when enabled
 	if o.debugLogEnabled && o.enableRedaction && o.logger != nil {

internal/translator/openai_openai_test.go

@@ -350,6 +350,16 @@ data: [DONE]
 			_, _, _, _, err := o.ResponseBody(nil, bytes.NewBuffer([]byte("invalid")), false, nil)
 			require.Error(t, err)
 		})
+		t.Run("null body", func(t *testing.T) {
+			// A literal JSON `null` decodes to a nil response without an error;
+			// it must not panic and should report zero usage.
+			s := &testotel.MockSpan{}
+			o := &openAIToOpenAITranslatorV1ChatCompletion{requestModel: "gpt-4o-mini"}
+			_, _, usedToken, responseModel, err := o.ResponseBody(nil, bytes.NewBuffer([]byte("null")), false, s)
+			require.NoError(t, err)
+			require.Equal(t, tokenUsageFrom(0, -1, -1, 0, 0, -1), usedToken)
+			require.Equal(t, "gpt-4o-mini", responseModel)
+		})
 		t.Run("valid body", func(t *testing.T) {
 			s := &testotel.MockSpan{}
 			var resp openai.ChatCompletionResponse