@voltagent/server-core@2.1.11

Patch Changes

• #1183 b48f107 Thanks @omeraplak! - feat: persist selected assistant message metadata to memory

  You can enable persisted assistant message metadata at the agent level or per request.

  const result = await agent.streamText("Hello", {
    memory: {
      userId: "user-1",
      conversationId: "conv-1",
      options: {
        messageMetadataPersistence: {
          usage: true,
          finishReason: true,
        },
      },
    },
  });

  With this enabled, fetching messages from memory returns assistant UIMessage.metadata
  with fields like usage and finishReason, not just stream-time metadata.

  REST API requests can enable the same behavior with options.memory.options:

  curl -X POST http://localhost:3141/agents/assistant/text \
    -H "Content-Type: application/json" \
    -d '{
      "input": "Hello",
      "options": {
        "memory": {
          "userId": "user-1",
          "conversationId": "conv-1",
          "options": {
            "messageMetadataPersistence": {
              "usage": true,
              "finishReason": true
            }
          }
        }
      }
    }'

• Updated dependencies [b48f107, 195155b]:

  • @voltagent/core@2.6.14

@voltagent/core@2.6.14

Patch Changes

• #1183 b48f107 Thanks @omeraplak! - feat: persist selected assistant message metadata to memory

  You can enable persisted assistant message metadata at the agent level or per request.

  const result = await agent.streamText("Hello", {
    memory: {
      userId: "user-1",
      conversationId: "conv-1",
      options: {
        messageMetadataPersistence: {
          usage: true,
          finishReason: true,
        },
      },
    },
  });

  With this enabled, fetching messages from memory returns assistant UIMessage.metadata
  with fields like usage and finishReason, not just stream-time metadata.

  REST API requests can enable the same behavior with options.memory.options:

  curl -X POST http://localhost:3141/agents/assistant/text \
    -H "Content-Type: application/json" \
    -d '{
      "input": "Hello",
      "options": {
        "memory": {
          "userId": "user-1",
          "conversationId": "conv-1",
          "options": {
            "messageMetadataPersistence": {
              "usage": true,
              "finishReason": true
            }
          }
        }
      }
    }'

• #1167 195155b Thanks @octo-patch! - fix: use OpenAI-compatible adapter for MiniMax provider

@voltagent/core@2.6.13

Patch Changes

• #1172 8cb2aa5 Thanks @omeraplak! - fix: tighten prompt-context usage telemetry
  • redact nested large binary fields when estimating prompt context usage
  • preserve circular-reference detection when serializing nested prompt message content
  • exclude runtime-only tool metadata from tool schema token estimates
  • avoid emitting cached and reasoning token span attributes when their values are zero

@voltagent/ag-ui@1.0.7

Patch Changes

• #1137 bb6e9b1 Thanks @corners99! - feat(ag-ui): add ACTIVITY_SNAPSHOT and ACTIVITY_DELTA event support

@voltagent/core@2.6.12

Patch Changes

• #1169 25b21d0 Thanks @omeraplak! - feat: add estimated prompt context telemetry for observability
  • record estimated prompt-context breakdown for system instructions, conversation messages, and tool schemas on LLM spans
  • expose cached and reasoning token usage on LLM spans for observability consumers
  • add tests for prompt-context estimation helpers

@voltagent/core@2.6.11

Patch Changes

• #1168 2075bd9 Thanks @omeraplak! - fix: emit LLM judge token and provider cost telemetry on eval scorer spans

  VoltAgent now records LLM judge model, token usage, cached tokens, reasoning tokens,
  and provider-reported cost details on live eval scorer spans.

  This makes scorer-side usage visible in observability backends and enables downstream
  cost aggregation to distinguish agent costs from eval scorer costs.

• #1163 6f14c4d Thanks @omeraplak! - fix: preserve usage and provider cost metadata on structured output failures

  When generateText receives a successful model response but structured output is not produced,
  VoltAgent now keeps the resolved usage, finish reason, and provider metadata on the resulting
  error path.

  This preserves provider-reported cost data for observability spans and makes the same metadata
  available to error hooks through VoltAgentError.metadata.

@voltagent/server-hono@2.0.8

Patch Changes

• b523a60 Thanks @omeraplak! - feat: publish the latest Hono server workflow route updates

  What's Changed

  • expose the latest workflow execution endpoints in the Hono server package
  • include OpenAPI route definitions for attaching to active workflow streams and replaying workflow executions
  • publish the current @voltagent/server-hono route surface so dev-server installs match the latest repo behavior

@voltagent/core@2.6.10

Patch Changes

• #1155 52bda94 Thanks @omeraplak! - fix: capture provider-reported OpenRouter costs in observability spans

  What's Changed

  • Forward OpenRouter provider-reported cost metadata to both LLM spans and root agent spans.
  • Record usage.cost and usage.cost_details.upstream_inference_* attributes for downstream cost consumers.
  • Document OpenRouter usage accounting and custom onEnd hook-based cost reporting in the observability docs.

@voltagent/ag-ui@1.0.6

Patch Changes

• #1149 19c4fcf Thanks @corners99! - fix: use input instead of args for tool-call parts in message conversion

  When converting CopilotKit assistant messages with tool calls to VoltAgent format,
  the adapter was setting args on tool-call parts. The AI SDK's ToolCallPart
  interface expects input, causing the Anthropic provider to send undefined as
  the tool_use input — rejected by the API with:

  "messages.N.content.N.tool_use.input: Input should be a valid dictionary"

@voltagent/core@2.6.8

Patch Changes

• #1146 c7b4c45 Thanks @omeraplak! - Improve structured-output error handling for Agent.generateText when models do not emit a final output (for example after tool-calling steps).
  • Detect missing result.output immediately when output is requested and throw a descriptive VoltAgentError (STRUCTURED_OUTPUT_NOT_GENERATED) instead of surfacing a vague AI_NoOutputGeneratedError later.
  • Include finish reason and step/tool metadata in the error for easier debugging.
  • Add an unhandled-rejection hint in VoltAgent logs for missing structured outputs.