Optimize LLM Context Compression for Codebases

Context Compression Strategies guides summarization for long sessions and codebases that exceed the working context window. Its optimization target is total tokens per completed task, including the cost of recovering information lost during compression. A smaller request can be more expensive overall if missing paths, errors, or decisions force repeated exploration.

The skill compares anchored iterative summarization, opaque compression, and regenerative full summaries. Anchored summaries preserve explicit sections and merge only newly truncated history into the existing record. Opaque representations prioritize reconstruction and compression while reducing interpretability. Regenerative summaries remain readable but rebuild the record each time, risking detail loss across cycles.

Structured sections act as preservation checklists. Session intent, files modified, decisions, current state, and next steps each receive a dedicated place. The artifact trail distinguishes created, modified, and read-only files and retains function names, variables, and errors. Because file tracking is weak even with structured summaries, the skill proposes a separate artifact index or explicit file-state tracking where necessary.

Use this skill for long-running sessions, large-codebase exploration, forgotten file modifications, summarization design, or compression evaluation. Anchored summaries fit work where file tracking and inspectable preservation matter. Opaque compression fits shorter sessions with low recovery costs and strong token-saving requirements. Regenerative summaries fit clear phase boundaries and situations where reviewing the full context again is acceptable.

Compression triggers include a fixed utilization threshold, a sliding window of recent turns plus summary, importance-based removal, and task-boundary compression. Sliding windows with structured summaries suit predictable coding workflows; consider compression around 70-80% utilization. These are design choices to evaluate in the actual environment.

Inputs include conversation history, task intent, artifact changes, technical identifiers, decisions, constraints, and pending work. Outputs are structured summaries, incremental updates, artifact indexes, compression policies, and evaluation results. Subsequent updates summarize only new truncated content, merge it into established sections, and track the originating compression cycle for debugging.

Functional probes test factual recall, modified-file tracking, continuation planning, and decision reasoning. The evaluation dimensions are accuracy, context awareness, artifact trail, completeness, continuity, and instruction following. Lexical metrics such as ROUGE or embedding similarity alone can miss the one path needed to continue.

Compare compression ratio and quality together: anchored summaries prioritize quality over the smallest representation, and recovery costs affect total work. Illustrative evaluation values pair anchored iterative summarization with 98.6% compression and a 3.70 quality score, regenerative summaries with 98.7% and 3.44, and opaque compression with 99.3% and 3.35. Artifact-trail scores range from 2.2 to 2.5 out of 5.0 across approaches. Monitor re-fetch frequency to see whether compression is actually saving work.

For large systems, the practical workflow first compresses exploration into a research document, then creates an implementation specification with signatures, types, and data flow, and finally implements against that specification. Migration examples and reference changes can seed the pattern by exposing invariants and distinguishing business requirements from legacy workarounds. The debugging example illustrates preservation of a concrete authentication error, a Redis-related cause, read-only versus modified files, test state, and remaining actions.

Related skills cover context degradation, context optimization, evaluation, and memory systems. An evaluation-framework reference supports functional probes, judge-based evaluation, and phased large-system work.