Claus Ibsen created CAMEL-27312:
-----------------------------------

             Summary: camel-jbang - the MCP server shows the agent runs, model 
calls and tool calls of the GenAI spans
                 Key: CAMEL-27312
                 URL: https://issues.apache.org/jira/browse/CAMEL-27312
             Project: Camel
          Issue Type: Sub-task
          Components: camel-jbang
            Reporter: Claus Ibsen
             Fix For: 4.24.0


CAMEL-27131 records GenAI spans at three levels 
({{camel.aiObservability.level}}): {{basic}} has one span per model call, 
{{detailed}} adds the agent runs ({{invoke_agent}}) with their model calls and 
tool calls ({{execute_tool}}), and {{full}} adds the messages. CAMEL-27135 
shows them in the Camel TUI: the AI usage view counts only the model calls, the 
Spans tab labels the GenAI spans and summarizes an agent run, and {{t}} shows 
the conversation of a span (apache/camel#30521, #30522, #30523).

An AI coding agent that uses the Camel MCP server sees none of this yet:
* {{camel_runtime_ai_trace}} combines the message history, the processor 
statistics and the route structure. It reads no spans, so it has no agent runs, 
no tool calls and no per-call tokens.
* {{camel_runtime_spans}} returns the raw span dump (100 spans by default). The 
GenAI data is there, but the agent has to rebuild the tree itself, and at level 
{{full}} each message attribute is a JSON string of up to 8192 characters per 
text.

*Proposal*
Make {{camel_runtime_ai_trace}} (or a new tool next to it) read the span dump 
({{get_spans}}) and add:
# *Model calls*: per model the calls, input, output and cached tokens, latency 
and finish reasons. Only model operations count, not {{invoke_agent}}, 
{{execute_tool}} or {{create_agent}} (the same rule as the TUI).
# *Agent runs* (level {{detailed}} and above): per run its model, its steps in 
order (each model call with tokens and finish reason, each tool call with its 
name and time), the token totals, and how its time splits into model calls, 
tool calls and the rest.
# *Conversation* (level {{full}}), only when asked with a parameter: the 
messages of one run as turns (system, user, assistant, tool call with 
arguments, tool result named after its tool). The last model call of a run 
already holds the whole run. Content may hold personal data and can be large, 
so it is opt-in and capped.
# *The level the spans show*, worked out from their attributes as the TUI does, 
with a hint when it is {{basic}} (agent runs and tool calls are not traced) or 
when there are no GenAI spans ({{camel run --observe}} adds 
camel-ai-observability when a GenAI component is used).

*Reuse*
The TUI already has this logic in camel-jbang-plugin-tui: 
{{GenAiSpanUsageExtractor}}, {{GenAiSpanView}} and {{GenAiTranscript}}. 
camel-jbang-mcp does not depend on the TUI plugin, so move the span model (no 
TamboUI types; {{GenAiSpanView}} mixes in the rendering) to a module both use, 
such as camel-jbang-core. Then the TUI and the MCP server count and group the 
same way.

Also update the tool description, which lists the AI components.

Example to try it with: apache/camel-jbang-examples#140 ({{ai/agent-tools}}).

_Created by Claude Code on behalf of Claus Ibsen_



--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

Reply via email to