[ 
https://issues.apache.org/jira/browse/CAMEL-26410?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Claus Ibsen resolved CAMEL-26410.
---------------------------------
    Resolution: Fixed

Merged in https://github.com/apache/camel/pull/28717

> camel-jbang - AI tools: a refused write reports the problems behind a YAML 
> indentation the write repairs by itself
> ------------------------------------------------------------------------------------------------------------------
>
>                 Key: CAMEL-26410
>                 URL: https://issues.apache.org/jira/browse/CAMEL-26410
>             Project: Camel
>          Issue Type: Improvement
>          Components: camel-jbang
>            Reporter: Claus Ibsen
>            Assignee: Claus Ibsen
>            Priority: Minor
>             Fix For: 4.24.0
>
>
> The AI write and edit tools repair a YAML shape that has one reading, for 
> example a list item two columns off (CAMEL-25532). The write then goes ahead 
> and says what it repaired.
> But when the repaired content has another problem, the repair is dropped, and 
> the write is refused with the indentation error only. The YAML parser stops 
> at the indentation, so the other checks never run. The model never sees the 
> real problem, and it sends the same lines again.
> In the local-model benchmark (qwen3.6:35b), this gave loops of 16 to 19 
> refused writes in a row. Four such steps were checked:
> * route-aggregator step 1 (three runs): a log two columns off after an 
> aggregate. With the column repaired, the validator says that steps: is at the 
> route level and goes under from:.
> * route-enrich-and-dedupe step 2: a log two columns off after a pollEnrich. 
> With the column repaired, the validator says that the log reads a field of 
> the variable that holds the file as it was read (a GenericFile), and how to 
> unmarshal it in place.
> In both cases the validator already has a clear message, but the model never 
> got it.
> The change: when the write can repair the shape, but the repaired content has 
> other problems, the write is still refused. The answer then gives those 
> problems, and a note that the write repairs the shape by itself. The answer 
> stays as before when a problem of the content as written is still there after 
> the repairs, or when a repair moved lines (the line numbers would not be the 
> model's).



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

Reply via email to