mirror of
https://github.com/agentscope-ai/ReMe.git
synced 2026-11-05 07:39:47 +00:00
The dream extract step packed every changed file into a single material blob and sent it in one LLM call, so a large scan window could exceed the backend context and crash it (issue #659: 19.4K-token prompt -> vkAllocateMemory failure). Split changed paths into size-capped batches (max_chars_per_batch, default 40000), scoped clean_paths to the current batch, merged same-name units across batches, and bisected a batch that fails in transport so the offending input is isolated instead of retried with an even larger prompt. |
||
|---|---|---|
| .. | ||
| fixtures/config | ||
| integration | ||
| unit | ||