Federico Mariani created CAMEL-23393:
----------------------------------------
Summary: camel-openai: Add token budget enforcement to the MCP
agentic loop
Key: CAMEL-23393
URL: https://issues.apache.org/jira/browse/CAMEL-23393
Project: Camel
Issue Type: Improvement
Components: camel-openai
Reporter: Federico Mariani
The _camel-openai_ agentic loop (OpenAIProducer.processNonStreamingAgentic())
is bounded only by *maxToolIterations* (default 50). This prevents infinite
iteration but does not guard against token/cost exhaustion. A loop could
consume hundreds of thousands of tokens across many iterations while the
iteration count remains within bounds.
Real-world agentic workloads can escalate quickly: each iteration sends the
full accumulated conversation (system + user + assistant + tool results for
every prior iteration) back to the model. By iteration 30, the prompt alone may
exceed 100k tokens even if no single tool returned a large result.
- Add a *maxAgenticTokens* configuration parameter (default: unset /
unlimited). When set, the loop tracks cumulative `prompt_tokens +
completion_tokens` across all iterations and stops when the budget is exceeded.
- Expose cumulative token usage via output headers
(`CamelOpenAIAgenticPromptTokens`, `CamelOpenAIAgenticCompletionTokens`,
`CamelOpenAIAgenticTotalTokens`) so downstream routes can observe total cost
regardless of whether the budget is enforced.
- When the budget is exceeded, throw a descriptive exception (similar to the
existing `maxToolIterations` exceeded behavior) that includes the token count
and iteration number.
The token data is already available — each `ChatCompletion` response includes
`usage.promptTokens()` / `usage.completionTokens()`. The implementation only
needs to accumulate these values across the loop.
*Example usage:*
{code:java}
- to:
uri: "openai:chat-completion"
parameters:
model: "gpt-4o"
maxToolIterations: 20
maxAgenticTokens: 200000
{code}
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
