Claus Ibsen created CAMEL-26962:
-----------------------------------

             Summary: camel-jbang - The AI clients send temperature to Claude, 
which the current Claude models reject
                 Key: CAMEL-26962
                 URL: https://issues.apache.org/jira/browse/CAMEL-26962
             Project: Camel
          Issue Type: Bug
          Components: camel-jbang
            Reporter: Claus Ibsen
            Assignee: Claus Ibsen
             Fix For: 4.24.0


The AI clients of the Camel CLI (camel ask, the AI panel, camel explain and the 
AI overview) call the Anthropic API through {{LlmClient}}. Every request it 
sends has {{temperature}}: {{chatAnthropicFormat}} puts it in the body, and the 
AI panel and camel ask set 0.3. The Claude API rejects that for the current 
models:
* Claude Opus 5.5 ({{claude-opus-5-5}}) rejects {{temperature}}, {{top_p}} and 
{{top_k}} with a 400.
* Claude Sonnet 5.5 ({{claude-sonnet-5-5}}) rejects a non-default value with a 
400.
* From API version {{2026-10-21}}, a request with {{temperature}} is a 400 on 
every model, even at the default value. {{LlmClient}} sends 
{{anthropic-version: 2023-06-01}} today.

So a user who picks a current Claude model, with {{camel.tui.ai.model}} or in 
camel ask, gets an error. The default model of {{LlmClient}} is 
{{claude-sonnet-4-6}}, which still accepts it.

Found while planning a benchmark that asks the same questions of a local model 
and of Claude in the AI panel. Not tested against the API (there is no key on 
the test machine): the rules above are from the Claude API documentation.

*Proposal*
* Do not send {{temperature}}, {{top_p}} or {{top_k}} to Anthropic. Keep them 
for the other providers.
* Use a current model as the default for Anthropic, for example 
{{claude-opus-5-5}}.
* Turn on prompt caching with a top-level {{"cache_control": {"type": 
"ephemeral"}}} on the request. The AI panel already keeps its prompt prefix the 
same from question to question (for the cache of a local server), so the tools, 
the system prompt and the history would come from the cache. Today every 
request pays the full input price for all of it. {{LlmClient}} already reads 
{{cache_read_input_tokens}}, so the effect shows in the usage.

*To check while doing this*
* The current models think by default (Claude Opus 5.5 always does). The tool 
loop rebuilds each assistant message from its text and tool calls, so the 
thinking blocks are not sent back. The documentation says to send them back 
unchanged, and that leaving them out can give a 400 in a tool loop.
* If the thinking blocks are sent back, compacting the history becomes an edit. 
For accounts created on or after Aug 31 2026, the API checks that everything 
before a thinking block is unchanged, and rejects an edited history with a 400. 
The AI panel compacts the history of a hosted model a little every turn.
* A move to API version 2026-10-21 needs the {{temperature}} change first: 
https://platform.claude.com/docs/en/api/migration-2026-10-21




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

Reply via email to