[
https://issues.apache.org/jira/browse/CAMEL-25953?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Claus Ibsen resolved CAMEL-25953.
---------------------------------
Resolution: Fixed
> camel-jbang - AI tools: an unchanged write says so, the log marks records
> from before the last reload, and a missed edit shows the file as it is
> ------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: CAMEL-25953
> URL: https://issues.apache.org/jira/browse/CAMEL-25953
> Project: Camel
> Issue Type: Improvement
> Components: camel-jbang
> Reporter: Claus Ibsen
> Assignee: Claus Ibsen
> Priority: Major
> Fix For: 4.24.0
>
>
> From the local-model benchmark of Oct 21-22 2026 (qwen3.6:35b-mlx through the
> camel-jbang-mcp server), five answers of the AI tools that misled the model:
> *1. An unchanged write.* A write of the content the file already has answered
> "overwritten", and models wrote one file unchanged up to 24 times in a row.
> It now answers status unchanged; nothing is written, so a running integration
> does not reload. An edit whose replacement gives the text the file has
> answers the same.
> *2. Log records from before the last reload.* camel_get_log showed ERROR
> records from before the model's own reload with no mark. The model took them
> for errors of its new version and changed a route that worked (the circuit
> breaker example). The records before the last reload are now marked
> beforeLastReload, with a reloadNote first in the answer. camel_get_errors
> already marked such errors, but its note came after 60-80 KB of stack traces,
> and the model read the first few thousand characters. The note is now the
> first key, and says how many exchanges failed since.
> *3. A thread name with brackets.* camel_get_log folded the records of a
> thread whose name has brackets, such as the JMS consumer's [onsumer[orders]],
> into the record before it, so the errors of one route showed as the cause of
> another. The record pattern now reads such names.
> *4. A miss among several edits.* camel_edit_file with two edits, of which the
> second missed, answered "None of the edits was made" with a fileWindow that
> showed the first edit made. The model sent only the second edit again, and
> the first was never written. fileWindow now shows the file as it is, and the
> message says to send all the edits again.
> *5. A line the snippet did not have.* The closest-lines merge of CAMEL-25532
> dropped a line of the file that the model's snippet from memory did not have
> (name: report of a cron endpoint). It now keeps the file's lines that the
> snippet does not have, and replaces only the ones it has.
> _Filed by Claude Code on behalf of davsclaus_
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
