Anthropics' Claude Code is addressing a critical flaw in its auto-compaction feature that previously discarded user-provided data like DOM markup, config files, and logs despite the original transcripts remaining intact on disk. A recent GitHub proposal introduces line-range annotations in compacted summaries to surgically recover lost context. For example, a 8,200-character DOM markup summary now includes `[transcript:lines 847-1023]`, allowing Claude to retrieve specific segments without reloading the entire transcript. This fix targets high-impact use cases where developers rely on precise code analysis or surgical debugging after extended sessions.

The issue emerged as sessions grew beyond context window limits, with auto-compaction irreversibly stripping details from summaries. Users reported hallucinated responses or repetitive requests for re-pasted data. The proposed solution adds minimal overhead: summaries retain compacted text while embedding metadata pointers to the original transcript's exact line ranges. This enables zero-loss recovery when users query specifics, avoiding redundant data ingestion and maintaining token efficiency.

Technical advantages include architecturally lightweight implementation — no new databases or models required — and flat scalability, as costs only apply during targeted recoveries. The fix also lays groundwork for cross-session references, potentially linking past projects' transcripts via session IDs. Community-driven workarounds like manual compaction flags or embedding stores are rendered obsolete by this native solution.

Priority: High. Impact: Critical for developers handling large input payloads. Example: A Chrome extension builder iterating on 8,000+ lines of DOM markup can now query precise element structures post-compaction without data loss. Expert FAQ: _"How does this differ from existing memory solutions?"_ Answer: Unlike external recall systems, the line-range approach leverages the existing transcript as ground truth, eliminating hallucinations and storage bloat while keeping summaries lean.