Merge d1886d2908 into 8a1f3568ba

terraform/provider/CHANGELOG.md

@@ -61,6 +61,7 @@ longer signal it.
 - **key**: A config-supplied `key` value (write-only) is now forwarded to `/key/generate`; previously it was silently dropped and the proxy generated a random key instead
 - **security**: The `litellm_key` data source and `litellm_key_block` resource normalize raw `sk-` keys to their SHA-256 token hash before building request URLs and resource IDs, so plaintext keys no longer land in reverse-proxy access logs, Terraform plan output, or state IDs
 - **unified_access_group**: create now accepts any 2xx response instead of requiring exactly HTTP 200; `POST /v1/unified_access_group` legitimately returns 201, so creation previously succeeded on the proxy but failed in the provider, leaving the group out of state and forcing a `terraform import` to recover on the next apply's 409. Same fix and shape as the one already applied to `mcp_server`/`model`/`key`/`organization_member`; `handleResponse` (shared by several other resources) was the one status-check helper that fix didn't reach. The legacy `litellm_access_group` resource calls the unrelated `/access_group/new` endpoint, which already returns 200, so it was never affected
+- **model**: `mode` on `litellm_model` now accepts `responses`, `realtime`, `image_edit`, `ocr`, `search`, `video_generation` and `evaluation`, matching the modes `model_prices_and_context_window.json` already assigns to shipped models (for example `responses` on `azure/codex-mini` and `realtime` on `amazon.nova-2-sonic-v1:0`); previously the enum only allowed `completion`, `embedding`, `image_generation`, `chat`, `moderation`, `audio_transcription`, `audio_speech` and `rerank`, so setting a model's `mode` to any of the missing values failed at plan time with `expected mode to be one of [...]`. `guardrail` and `vector_store` are deliberately left out, since those are modeled as the dedicated `litellm_guardrail` and `litellm_vector_store` resources rather than as a `litellm_model` mode
 
 ### Changed
 

terraform/provider/docs/resources/model.md

@@ -137,6 +137,13 @@ The following arguments are supported:
   * `audio_transcription`
   * `audio_speech`
   * `rerank`
+  * `responses`
+  * `realtime`
+  * `image_edit`
+  * `ocr`
+  * `search`
+  * `video_generation`
+  * `evaluation`
 
 * `tpm` - (Optional) integer. Tokens per minute limit for this model.
 

terraform/provider/litellm/resource_model.go

@@ -110,6 +110,13 @@ func resourceLiteLLMModel() *schema.Resource {
 					"audio_transcription",
 					"audio_speech",
 					"rerank",
+					"responses",
+					"realtime",
+					"image_edit",
+					"ocr",
+					"search",
+					"video_generation",
+					"evaluation",
 				}, false),
 			},
 			"input_cost_per_million_tokens": {

terraform/provider/litellm/resource_model_test.go

@@ -242,3 +242,22 @@ func TestResourceLiteLLMModelUpdateSkipsPatchWhenDisplayNameUnchanged(t *testing
 		t.Fatalf("update failed: %v", err)
 	}
 }
+
+func TestResourceLiteLLMModel_ModeValidation(t *testing.T) {
+	validateMode := resourceLiteLLMModel().Schema["mode"].ValidateFunc
+
+	valid := []string{
+		"completion", "embedding", "image_generation", "chat", "moderation",
+		"audio_transcription", "audio_speech", "rerank",
+		"responses", "realtime", "image_edit", "ocr", "search", "video_generation", "evaluation",
+	}
+	for _, mode := range valid {
+		if _, errs := validateMode(mode, "mode"); len(errs) != 0 {
+			t.Errorf("mode %q: expected no validation error, got %v", mode, errs)
+		}
+	}
+
+	if _, errs := validateMode("not_a_real_mode", "mode"); len(errs) == 0 {
+		t.Error(`mode "not_a_real_mode": expected a validation error, got none`)
+	}
+}