fix(model-bridge): handle SDK context hints and Responses metadata

src/main/model-bridge/protocol.ts

@@ -43,10 +43,25 @@ export function historyKey(content: Json[]): string {
 function replayKey(prefix: string, content: Json[]): string {
   return createHash('sha256').update(prefix).digest('hex') + ':' + createHash('sha256').update(historyKey(content)).digest('hex')
 }
+function validateContextManagement(value: unknown): void {
+  if (value === undefined || value === null) return
+  requireValue(record(value) && Object.keys(value).every(key => key === 'edits'), 'Unsupported Messages context_management')
+  if (value.edits === undefined) return
+  requireValue(Array.isArray(value.edits), 'Invalid Messages context_management edits')
+  // Adaptive thinking adds "keep all" even for custom models. It is a no-op;
+  // actual server-side clearing/compaction cannot be translated across protocols.
+  for (const edit of value.edits) {
+    requireValue(record(edit) && Object.keys(edit).every(key => key === 'type' || key === 'keep')
+      && edit.type === 'clear_thinking_20251015'
+      && (edit.keep === 'all' || record(edit.keep) && edit.keep.type === 'all' && Object.keys(edit.keep).length === 1),
+    'Unsupported Messages context_management edit: only keeping all thinking is supported')
+  }
+}
 export function convertRequest(body: Json, model: ResolvedModel, context: ConversionContext): Json {
   const { config: c, profile: p } = model, responses = c.api === 'openai-responses'
   requireValue(Array.isArray(body.messages), 'messages must be an array')
-  for (const key of ['mcp_servers', 'container', 'context_management']) requireValue(body[key] === undefined, `Unsupported Messages field: ${key}`)
+  for (const key of ['mcp_servers', 'container']) requireValue(body[key] === undefined, `Unsupported Messages field: ${key}`)
+  validateContextManagement(body.context_management)
   const input: Json[] = []
   const prefix: Json[] = []
   context.responsePrefix = undefined
@@ -56,8 +71,10 @@ export function convertRequest(body: Json, model: ResolvedModel, context: Conver
     if (system.length) input.push({ role: 'system', content: system.map(b => b.text).join('\n\n') })
   }
   for (const message of body.messages) {
-    requireValue(record(message) && ['user', 'assistant'].includes(message.role), 'Invalid message role')
+    requireValue(record(message) && ['user', 'assistant', 'system'].includes(message.role), 'Invalid message role')
     const role = message.role, content = blocks(message.content)
+    // The SDK also appends environment instructions as an in-history system message.
+    if (role === 'system') requireValue(content.every(b => b.type === 'text' && typeof b.text === 'string'), 'Only text system blocks are supported')
     const prior = JSON.stringify(prefix)
     prefix.push({ role, content: JSON.parse(historyKey(content)) })
     if (responses && role === 'assistant') {

src/main/model-bridge/stream.ts

@@ -136,6 +136,9 @@ export async function streamResponse(upstream: Response, response: ServerRespons
             await context.persist?.()
             finish = converted.stop_reason; usage = converted.usage; completed = true; break
           }
+          // Gateway metadata carries no model content. Keep reading to the terminal
+          // event so reasoning history is saved before the SDK starts its next turn.
+          case 'response.metadata':
           case 'response.created': case 'response.in_progress': case 'response.queued':
           case 'response.content_part.added': case 'response.content_part.done':
           case 'response.reasoning_summary_part.added': case 'response.reasoning_summary_part.done':