At 5ce5888b9f149beaace393957a55ea8ee46c9f71, the OpenAI decoder can use the
SDK’s custom base URL and API key, but the localization CLI restricts --model
to a fixed set of IDs. This prevents using a compatible hosted model through
the existing backend without changing source.
Would a small generic custom-model path be welcome? The proposed change would
keep the existing defaults, allow an arbitrary ID when the OpenAI backend is
selected, and make the context/output budget explicit. It would reuse
agentless/util/model.py and api_requests.py; no new provider transport is
needed. A CLI-to-decoder test should verify that the custom ID and base URL
reach an actual serialized chat request.
Tsubasa’s tsubasa-fast and tsubasa-pro are the concrete use case, each with a
32,768-token context limit. Initial qualification would use single-sample
requests. Multi-sample batching, the complete localization/repair prompt budget,
and actual SWE-bench repair quality have not been tested.
This proposal concerns configuration and request compatibility, not a benchmark
claim. OpenAI Codex assisted with source inspection.
At
5ce5888b9f149beaace393957a55ea8ee46c9f71, the OpenAI decoder can use theSDK’s custom base URL and API key, but the localization CLI restricts
--modelto a fixed set of IDs. This prevents using a compatible hosted model through
the existing backend without changing source.
Would a small generic custom-model path be welcome? The proposed change would
keep the existing defaults, allow an arbitrary ID when the OpenAI backend is
selected, and make the context/output budget explicit. It would reuse
agentless/util/model.pyandapi_requests.py; no new provider transport isneeded. A CLI-to-decoder test should verify that the custom ID and base URL
reach an actual serialized chat request.
Tsubasa’s
tsubasa-fastandtsubasa-proare the concrete use case, each with a32,768-token context limit. Initial qualification would use single-sample
requests. Multi-sample batching, the complete localization/repair prompt budget,
and actual SWE-bench repair quality have not been tested.
This proposal concerns configuration and request compatibility, not a benchmark
claim. OpenAI Codex assisted with source inspection.