Skip to content

The open-source agent harness - the runtime layer that turns an LLM into a working agent.

fix: persist model.message with omitted null content and deferred call_tool wrapper tool_info (#343)

Srajan Asthana889cacaf8c9e
diff --git a/.changeset/model-message-event-shape.md b/.changeset/model-message-event-shape.md
new file mode 100644
index 000000000..418ec26a0
--- /dev/null
+++ b/.changeset/model-message-event-shape.md
@@ -0,0 +1,5 @@
+---
+'@truefoundry/trueforge-core': patch
+---
+
+Persist `model.message` with omitted null content and deferred `call_tool` wrapper `tool_info`, matching the live stream.
diff --git a/packages/trueforge-core/src/core/runtime/AgentThread.ts b/packages/trueforge-core/src/core/runtime/AgentThread.ts
index eb2094739..3e2c6785b 100644
--- a/packages/trueforge-core/src/core/runtime/AgentThread.ts
+++ b/packages/trueforge-core/src/core/runtime/AgentThread.ts
@@ -239,12 +239,16 @@ function buildModelMessageEvent({
id: string;
}): ModelMessageEvent {
// `thinking_blocks` / `source` stay on the context message for replay; strip them from the client event.
- const { role, tool_calls, thinking_blocks, source, ...rest } = assistantMessage;
+ const { role, tool_calls, thinking_blocks, source, content, ...rest } = assistantMessage;
void role;
void thinking_blocks;
void source;
const event: ModelMessageEvent = {
...rest,
+ // Tool-only completions store content: null on context (OpenAI replay) but the
+ // SSE placeholder omits the field. Drop null/empty here so listTurnEvents JSON
+ // matches a folded stream.
+ ...(content && { content }),
tool_calls: tool_calls?.map(toEnrichedToolCall),
type: EventType.MODEL_MESSAGE,
id,
@@ -411,10 +415,15 @@ async function buildModelMessageDeltaEvent({
};
}
-async function buildContextAssistantMessage(
- assistantMessage: RawAssistantMessage,
- toolMapping: Map<string, MappedMCPTool>,
-): Promise<InternalEnrichedAssistantMessage> {
+async function enrichAssistantMessage({
+ assistantMessage,
+ toolMapping,
+ resolveUnderlyingTool,
+}: {
+ assistantMessage: RawAssistantMessage;
+ toolMapping: Map<string, MappedMCPTool>;
+ resolveUnderlyingTool: boolean;
+}): Promise<InternalEnrichedAssistantMessage> {
// Omit tool_calls when empty; OpenAI rejects `tool_calls: []` on replay.
if (!assistantMessage.tool_calls?.length) {
const { tool_calls, ...rest } = assistantMessage;
@@ -435,11 +444,9 @@ async function buildContextAssistantMessage(
const tool_info = await toolInfo.toolSet.toolCallInfo(
{
name: toolInfo.originalToolName,
- arguments: tryParseToolArgs(toolCall.function.arguments),
+ ...(resolveUnderlyingTool ? { arguments: tryParseToolArgs(toolCall.function.arguments) } : {}),
},
- // During context persistence, we have the complete message and hence the tool arguments are available.
- // Hence, we resolve the underlying tool to get the tool information.
- true /* resolveUnderlyingTool */,
+ resolveUnderlyingTool,
);
enrichedToolCalls.push({ ...toolCall, tool_info });
}
@@ -1063,13 +1070,21 @@ export class AgentThread {
usage: result.value.usage,
toolMapping,
});
- const assistantMessage: InternalEnrichedAssistantMessage = await buildContextAssistantMessage(
- result.value.output,
+ const assistantMessage: InternalEnrichedAssistantMessage = await enrichAssistantMessage({
+ assistantMessage: result.value.output,
toolMapping,
- );
+ // During context persistence, we have the complete message and hence the tool arguments are available.
+ // Hence, we resolve the underlying tool to get the tool information.
+ resolveUnderlyingTool: true,
+ });
const finishReason = result.value.finish_reason;
const agentAssistantMessage = buildModelMessageEvent({
- assistantMessage,
+ assistantMessage: await enrichAssistantMessage({
+ assistantMessage: result.value.output,
+ toolMapping,
+ // Same unresolved wrapper as SSE deltas so listTurnEvents matches a folded stream.
+ resolveUnderlyingTool: false,
+ }),
threadId: this.threadId,
finishReason,
usage: modelMessageUsage,