v0.2.3

Fixed

• Sanitized Cline request message content before sending to /chat/completions to prevent empty content blocks that can cause 400 errors on Anthropic-backed models (e.g. Claude Sonnet 4.6).
• Normalized Cline message history to text-first blocks for Anthropic compatibility (dropping non-text assistant/tool-call parts and injecting fallback text when needed) to avoid messages.N content is empty failures.
• Hardened Cline context hook activation by inferring active provider from session model_change entries when ctx.model is unavailable.
• Coerced unknown message content shapes to safe fallback text to prevent downstream empty-content serialization.
• Collapsed per-turn Cline context into system + wrapped user transcript to avoid Anthropic tool-protocol/empty-content failures on follow-up turns.
• Made Cline transcript collapsing idempotent by reusing existing <task> content and appending only new turns, preventing recursive self-review loops on follow-up tool calls.
• Preserved tool-call intent in collapsed transcripts by attaching concise tool command summaries (e.g. bash command/path) to <tool_result> blocks.
• Reduced assistant self-repetition by excluding assistant orchestration-only tool-call turns from collapsed history while keeping actual tool outputs.
• De-duplicated repeated identical no-output tool results and injected a system note when the same command returns no output multiple times (including guidance for common git diff main...origin/main misuse).