[
https://issues.apache.org/jira/browse/CAMEL-26758?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Claus Ibsen resolved CAMEL-26758.
---------------------------------
Resolution: Fixed
> camel-jbang - AI tools: the third identical refused write gets the fix that
> the refusal gave
> --------------------------------------------------------------------------------------------
>
> Key: CAMEL-26758
> URL: https://issues.apache.org/jira/browse/CAMEL-26758
> Project: Camel
> Issue Type: Improvement
> Components: camel-jbang, camel-yaml-dsl
> Reporter: Claus Ibsen
> Assignee: Claus Ibsen
> Priority: Minor
> Fix For: 4.24.0
>
>
> Follow-up from the local-model benchmark (qwen3.6, r28 to r30). A model
> sometimes sends the same refused file again and again, also when the refusal
> gives the exact lines to write. In r28 to r30 this happened in 23 step runs:
> 13 of the 14 runs with 4 or more sends failed, and all 9 runs with 8 or more
> sends failed. Repairs per mistake (CAMEL-25532, CAMEL-26730) help, but the
> loop then moves to the next mistake: contracts-rest-code-first step 1 looped
> on a concatenation and an extra quote (r28), on the literal \n (r29), and on
> a jsonpath with [] (r30, 15 times in two runs).
> * When the write tool refuses the same content of a file for the third time
> in a row, it now makes the fix that the refusal gave as exact text ("Write
> lines A-B as:" with the lines, or "write the line as ..."), runs the repairs
> of shapes with one reading, and writes the file when no new problem is left.
> The answer lists what changed. This happens only when every problem of the
> write has such text; otherwise the write is refused as before. Not for Citrus
> tests.
> * The jsonpath message for [] gives the line to write, and the message for an
> unclosed double quote ends with the line to write, so the write tool can
> apply both.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
