SEZ9 opened a new issue, #11651: URL: https://github.com/apache/seatunnel/issues/11651
### Search before asking - [X] I had searched in the [feature](https://github.com/apache/seatunnel/issues?q=is%3Aissue+label%3A%22Feature%22) and found no similar feature requirement. ### Description Sub-task of #11616 (Stage-1 items 1.2 / 1.3 / 1.4). The accuracy benchmark (#11553) identified three **pattern-level knowledge gaps** where nearly every evaluated model fails — these are not connector-option gaps but missing scenario knowledge in the CLI's skill/example layers: 1. **Conditional routing / predicate split** (3 tasks failed by 5–6 of 7 models; weakest models score 0% in this category): "route amount >= 1000 to files, the rest to console". Models consistently mis-wire the split — reading the source twice, chaining the transforms instead of running them in parallel off the same source label, or writing overlapping predicates (`> 100` / `< 100` drops equality). No existing skill or golden example covers the pattern "one source → parallel SQL transforms with mutually exclusive predicates → different sinks". 2. **PostgreSQL-CDC prerequisites** (CDC category: 25–50% pass on weaker models): generated PG-CDC configs omit `slot.name` / `decoding.plugin.name` and use MySQL-style `db.table` identifiers instead of `schema.table`. The cdc_realtime skill documents MySQL specifics but nothing PostgreSQL-specific. 3. **Wide-DAG wiring** (5+ blocks with transforms mid-chain; 1 task failed by 5/7 models): after inserting a transform, models leave the sink consuming the source's label (silently bypassing the transform), or reuse one label across different chains. ### Usage Scenario Proposed additions (content-only — markdown skills + a few lines in the config-agent prompt template; no code changes): - new `conditional_routing` skill (trigger-routed, with a complete mutually-exclusive-split HOCON pattern) - a PostgreSQL-CDC section in the `cdc_realtime` skill (slot/decoding plugin/schema-qualified names, MySQL-vs-PG table) - wide-DAG wiring rules in the `multi_pipeline` skill and the Multi-Pipeline Routing section of the config-agent system prompt Effectiveness to be verified with before/after benchmark runs on the affected task subset, per the #11616 working principle. ### Related issues #11616, #11553 ### Are you willing to submit a PR? - [X] Yes I am willing to submit a PR! ### Code of Conduct - [X] I agree to follow this project's [Code of Conduct](https://www.apache.org/foundation/policies/conduct) -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
