[
https://issues.apache.org/jira/browse/CAMEL-24886?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18125779#comment-18125779
]
Claus Ibsen commented on CAMEL-24886:
-------------------------------------
Status, Oct 10 2026. The four HTTP rungs are in the step-by-step benchmark, and
the local model (qwen3.6:35b-a3b through the camel-jbang-mcp server) passes
most of them: 69 of 78 steps across the five r7 runs (Oct 9) and the q8 run
(Oct 10), against 0 of 15 one-shot attempts when this ticket was filed.
||rung||steps passed||
|stock-api|17 of 18|
|http-client|15 of 18|
|openapi-server|22 of 24|
|openapi-client|15 of 18|
What is left, by cause:
* openapi-client step 2 (2 losses): the model keeps the sku in an exchange
property, and the rest-openapi call goes out to a literal /stock/{sku}/reserve
(985 and 3124 requests). The validator now reports it: CAMEL-24992.
* http-client step 2 (3 losses): the order id is set inside the split from
${body[orderId]}, where the body is one line, so the log reads ": CAMEL-MUG x
2"; and twice the URI reads ${header.sku} where the route set the exchange
property sku. The second is CAMEL-25516. For the first, a write now answers
with what the route logged after the reload (CAMEL-25513), which shows the
empty order id at once; the next run measures whether that is enough. If not,
checking the fields a step reads against a sample of the input (the shape of
the body inside the split) belongs with the body type flow, CAMEL-24844.
* openapi-server step 1 and step 4, openapi-client step 1: one loss each, no
common cause.
Next: a full ladder run on a build with CAMEL-25513, CAMEL-24992 and
CAMEL-25516.
> camel-jbang - make a local model good at HTTP-based apps: the rest, http,
> rest-openapi rungs of the examples ladder, step by step
> ---------------------------------------------------------------------------------------------------------------------------------
>
> Key: CAMEL-24886
> URL: https://issues.apache.org/jira/browse/CAMEL-24886
> Project: Camel
> Issue Type: Improvement
> Components: camel-http, camel-jbang, camel-rest-openapi, rest
> Reporter: Claus Ibsen
> Assignee: Claus Ibsen
> Priority: Major
> Fix For: 4.24.0
>
>
> HTTP is what most integrations are today: a REST API in front of a route, a
> client calling a service, a contract in OpenAPI. In the camel-jbang-mcp
> benchmark the local model never passed one of the HTTP rungs of the
> camel-jbang-examples ladder in one shot (connect/stock-api,
> connect/http-client, contracts/openapi-server, contracts/openapi-client: 0 of
> 15 attempts across two series), and they are not in the step-by-step
> benchmark yet.
> This ticket tracks making a local model good at HTTP-based Camel apps, the
> same way the other rungs were done: run the four READMEs' "build it step by
> step" sections through the camel-jbang-mcp server, read every failed step
> against the current code, and fix Camel where Camel is at fault (a message
> that does not say the shape, a validator that lets a wrong form through, a
> catalog sample that does not show the whole rest/rest-openapi/http shape, a
> runtime error that names a 404 instead of the option that caused it). Each
> finding gets its own ticket linked here.
> The four rungs and what they exercise:
> * stock-api: rest with get, a {sku} path parameter, jsonpath on a file, a 404
> branch
> * http-client: an http producer with a fixed URI, then toD with a per-line
> URI, throwExceptionOnFailure=false and a choice on the response code
> * openapi-server: contract-first rest from stock-api.json,
> listStock/getStock/reserveStock, request validation with a 400, a 409 for
> short stock
> * openapi-client: rest-openapi with operationId, a body from the message, an
> onException for the 409
> Benchmark side (camel-performance-tests, not Jira work): a probe request the
> harness sends after each step (the README's curl), and a peer app started
> next to the model's app for the two clients.
> Related: CAMEL-24834 (tool groups from the running app), CAMEL-24844 (body
> type flow); the rungs the model already does stepwise are in the round-2 blog
> post.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
