fix: remove blunt step-prune fallback that caused amnesia on long turns
The STEP_PRUNE_FALLBACK_THRESHOLD (200 messages) caused applyStepPruning to fall back to applyContextPruning on every step of long turns. That function uses reasoning: 'before-last-message' which strips all chain-of-thought during tool loops, causing the model to lose its plan and repeat completed work. Removed the threshold entirely — step pruning now always uses reasoning: 'none' with the adjusted tool-call window, preserving the model's chain-of-thought regardless of conversation length.
xonecas2026-03-21
30bf69b0c623edaccd115c324a20e902c04ca20a
parent 09719d4