Problem
Pre-run review truncates the full prompt to MAX_TOTAL_CONTEXT_SIZE by slicing from the start. This can cut off the task/instructions and tail sections (stages/reason), sending the model a wall of code without guidance.
Impact
Review quality degrades or becomes incoherent on large workspaces; critical instructions can be dropped.
Suggested Fix
Truncate per-file inputs and keep the system prompt/task sections intact. Preserve the final instruction block and only reduce large code sections.
References
- src/goldfish/pre_run_review.py MAX_TOTAL_CONTEXT_SIZE handling
Problem
Pre-run review truncates the full prompt to MAX_TOTAL_CONTEXT_SIZE by slicing from the start. This can cut off the task/instructions and tail sections (stages/reason), sending the model a wall of code without guidance.
Impact
Review quality degrades or becomes incoherent on large workspaces; critical instructions can be dropped.
Suggested Fix
Truncate per-file inputs and keep the system prompt/task sections intact. Preserve the final instruction block and only reduce large code sections.
References