v0.3.20-beta

Added kodelet conversation show --format markdown with structured markdown rendering for tool calls and results, plus truncation controls for verbose tool output.
Introduced the read_conversation tool so agents can extract goal-focused context from saved Kodelet conversations without loading the full transcript.
Limited bash tool output returned to the model to roughly 10k tokens, preserving head/tail context while reporting total output lines and clearly marking truncation for very large command results.