fix: handle streaming usage chunk and soft-deleted provider filtering (#76)

* fix(adapters): handle separate usage chunk in OpenAI streaming responses

Some OpenAI-compatible providers send token usage in a separate chunk
(with choices=[]) after the finish_reason chunk, rather than bundling
them together. Previously, this usage-only chunk was silently skipped
because data.choices[0] was undefined, causing token counts to always
be recorded as -1 for streaming requests. This broke TPM rate limiting
and usage tracking.

Now both patterns are supported:
- Usage bundled with finish_reason in the same chunk
- Usage in a separate subsequent chunk (choices=[])

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(db): filter soft-deleted providers in listUniqueSystemNames

listUniqueSystemNames() only filtered NOT models.deleted but did not
join the providers table or check NOT providers.deleted. When a provider
was soft-deleted, its models (still deleted=false) would appear in the
global model registry, but getModelsWithProviderBySystemName() correctly
filtered them out, causing the UI to show models with no providers.

Add innerJoin on ProvidersTable and NOT providers.deleted filter to
match the behavior of getModelsWithProviderBySystemName().

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(adapters): omit messageDelta in usage-only chunk to avoid overwriting stopReason

Remove explicit stopReason: null from the usage-only message_delta
yield to prevent overwriting a previously set stopReason. This aligns
with the pattern used in openai-responses.ts for usage-only events.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: pescn

backend/src/adapters/upstream/openai.ts

@@ -11,8 +11,7 @@ import type {
   InternalResponse,
   InternalStreamChunk,
   InternalToolDefinition,
-  InternalUsage,
-  ProviderConfig,
+ProviderConfig,
   StopReason,
   TextContentBlock,
   ThinkingContentBlock,
@@ -387,7 +386,7 @@ export const openaiUpstreamAdapter: UpstreamAdapter = {
 
   buildRequest(
     request: InternalRequest,
-    provider: ProviderConfig,
+    provider:  ProviderConfig,
   ): { url: string; init: RequestInit } {
     // Build messages array with system prompt
     const messages: OpenAIMessage[] = [];
@@ -510,6 +509,17 @@ export const openaiUpstreamAdapter: UpstreamAdapter = {
 
       const choice = data.choices[0];
       if (!choice) {
+        // Usage-only chunk (choices=[]) — some providers send usage separately
+        // after the finish_reason chunk when stream_options.include_usage=true
+        if (data.usage) {
+          yield {
+            type: "message_delta",
+            usage: {
+              inputTokens: data.usage.prompt_tokens,
+              outputTokens: data.usage.completion_tokens,
+            },
+          };
+        }
         continue;
       }
 
@@ -581,18 +591,20 @@ export const openaiUpstreamAdapter: UpstreamAdapter = {
       // Handle finish reason
       if (choice.finish_reason) {
         yield { type: "content_block_stop", index: blockIndex };
-        const usage: InternalUsage = data.usage
-          ? {
-              inputTokens: data.usage.prompt_tokens,
-              outputTokens: data.usage.completion_tokens,
-            }
-          : { inputTokens: -1, outputTokens: -1 };
+        // Include usage only if upstream provided it in this chunk.
+        // Some providers bundle usage with finish_reason; others send a
+        // separate usage-only chunk (choices=[]) immediately after.
         yield {
           type: "message_delta",
           messageDelta: {
             stopReason: convertFinishReason(choice.finish_reason),
           },
-          usage,
+          ...(data.usage && {
+            usage: {
+              inputTokens: data.usage.prompt_tokens,
+              outputTokens: data.usage.completion_tokens,
+            },
+          }),
         };
       }
     }

backend/src/db/index.ts

@@ -955,9 +955,14 @@ export async function listUniqueSystemNames(
   const r = await db
     .selectDistinct({ systemName: schema.ModelsTable.systemName })
     .from(schema.ModelsTable)
+    .innerJoin(
+      schema.ProvidersTable,
+      eq(schema.ModelsTable.providerId, schema.ProvidersTable.id),
+    )
     .where(
       and(
         not(schema.ModelsTable.deleted),
+        not(schema.ProvidersTable.deleted),
         modelType ? eq(schema.ModelsTable.modelType, modelType) : undefined,
       ),
     )