Context Compression

1---
2name: context-compression
3description: This skill should be used when the user asks to "compress context", "summarize conversation history", "implement compaction", "reduce token usage", or mentions context compression, structured summarization, tokens-per-task optimization, or long-running agent sessions exceeding context limits.
4---
5 
6# Context Compression Strategies
7 
8When agent sessions generate millions of tokens of conversation history, compression becomes mandatory. The naive approach is aggressive compression to minimize tokens per request. The correct optimization target is tokens per task: total tokens consumed to complete a task, including re-fetching costs when compression loses critical information.
9 
10## When to Activate
11 
12Activate this skill when:
13- Agent sessions exceed context window limits
14- Codebases exceed context windows (5M+ token systems)
15- Designing conversation summarization strategies
16- Debugging cases where agents "forget" what files they modified
17- Building evaluation frameworks for compression quality
18 
19## Core Concepts
20 
21Context compression trades token savings against information loss. Three production-ready approaches exist:
22 
231. **Anchored Iterative Summarization**: Maintain structured, persistent summaries with explicit sections for session intent, file modifications, decisions, and next steps. When compression triggers, summarize only the newly-truncated span and merge with the existing summary. Structure forces preservation by dedicating sections to specific information types.
24 
252. **Opaque Compression**: Produce compressed representations optimized for reconstruction fidelity. Achieves highest compression ratios (99%+) but sacrifices interpretability. Cannot verify what was preserved.
26 
273. **Regenerative Full Summary**: Generate detailed structured summaries on each compression. Produces readable output but may lose details across repeated compression cycles due to full regeneration rather than incremental merging.
28 
29The critical insight: structure forces preservation. Dedicated sections act as checklists that the summarizer must populate, preventing silent information drift.
30 
31## Detailed Topics
32 
33### Why Tokens-Per-Task Matters
34 
35Traditional compression metrics target tokens-per-request. This is the wrong optimization. When compression loses critical details like file paths or error messages, the agent must re-fetch information, re-explore approaches, and waste tokens recovering context.
36 
37The right metric is tokens-per-task: total tokens consumed from task start to completion. A compression strategy saving 0.5% more tokens but causing 20% more re-fetching costs more overall.
38 
39### The Artifact Trail Problem
40 
41Artifact trail integrity is the weakest dimension across all compression methods, scoring 2.2-2.5 out of 5.0 in evaluations. Even structured summarization with explicit file sections struggles to maintain complete file tracking across long sessions.
42 
43Coding agents need to know:
44- Which files were created
45- Which files were modified and what changed
46- Which files were read but not changed
47- Function names, variable names, error messages
48 
49This problem likely requires specialized handling beyond general summarization: a separate artifact index or explicit file-state tracking in agent scaffolding.
50 
51### Structured Summary Sections
52 
53Effective structured summaries include explicit sections:
54 
55```markdown
56## Session Intent
57[What the user is trying to accomplish]
58 
59## Files Modified
60- auth.controller.ts: Fixed JWT token generation
61- config/redis.ts: Updated connection pooling
62- tests/auth.test.ts: Added mock setup for new config
63 
64## Decisions Made
65- Using Redis connection pool instead of per-request connections
66- Retry logic with exponential backoff for transient failures
67 
68## Current State
69- 14 tests passing, 2 failing
70- Remaining: mock setup for session service tests
71 
72## Next Steps
731. Fix remaining test failures
742. Run full test suite
753. Update documentation
76```
77 
78This structure prevents silent loss of file paths or decisions because each section must be explicitly addressed.
79 
80### Compression Trigger Strategies
81 
82When to trigger compression matters as much as how to compress:
83 
84| Strategy | Trigger Point | Trade-off |
85|----------|---------------|-----------|
86| Fixed threshold | 70-80% context utilization | Simple but may compress too early |
87| Sliding window | Keep last N turns + summary | Predictable context size |
88| Importance-based | Compress low-relevance sections first | Complex but preserves signal |
89| Task-boundary | Compress at logical task completions | Clean summaries but unpredictable timing |
90 
91The sliding window approach with structured summaries provides the best balance of predictability and quality for most coding agent use cases.
92 
93### Probe-Based Evaluation
94 
95Traditional metrics like ROUGE or embedding similarity fail to capture functional compression quality. A summary may score high on lexical overlap while missing the one file path the agent needs.
96 
97Probe-based evaluation directly measures functional quality by asking questions after compression:
98 
99| Probe Type | What It Tests | Example Question |
100|------------|---------------|------------------|
101| Recall | Factual retention | "What was the original error message?" |
102| Artifact | File tracking | "Which files have we modified?" |
103| Continuation | Task planning | "What should we do next?" |
104| Decision | Reasoning chain | "What did we decide about the Redis issue?" |
105 
106If compression preserved the right information, the agent answers correctly. If not, it guesses or hallucinates.
107 
108### Evaluation Dimensions
109 
110Six dimensions capture compression quality for coding agents:
111 
1121. **Accuracy**: Are technical details correct? File paths, function names, error codes.
1132. **Context Awareness**: Does the response reflect current conversation state?
1143. **Artifact Trail**: Does the agent know which files were read or modified?
1154. **Completeness**: Does the response address all parts of the question?
1165. **Continuity**: Can work continue without re-fetching information?
1176. **Instruction Following**: Does the response respect stated constraints?
118 
119Accuracy shows the largest variation between compression methods (0.6 point gap). Artifact trail is universally weak (2.2-2.5 range).
120 
121## Practical Guidance
122 
123### Three-Phase Compression Workflow
124 
125For large codebases or agent systems exceeding context windows, apply compression through three phases:
126 
1271. **Research Phase**: Produce a research document from architecture diagrams, documentation, and key interfaces. Compress exploration into a structured analysis of components and dependencies. Output: single research document.
128 
1292. **Planning Phase**: Convert research into implementation specification with function signatures, type definitions, and data flow. A 5M token codebase compresses to approximately 2,000 words of specification.
130 
1313. **Implementation Phase**: Execute against the specification. Context remains focused on the spec rather than raw codebase exploration.
132 
133### Using Example Artifacts as Seeds
134 
135When provided with a manual migration example or reference PR, use it as a template to understand the target pattern. The example reveals constraints that static analysis cannot surface: which invariants must hold, which services break on changes, and what a clean migration looks like.
136 
137This is particularly important when the agent cannot distinguish essential complexity (business requirements) from accidental complexity (legacy workarounds). The example artifact encodes that distinction.
138 
139### Implementing Anchored Iterative Summarization
140 
1411. Define explicit summary sections matching your agent's needs
1422. On first compression trigger, summarize truncated history into sections
1433. On subsequent compressions, summarize only new truncated content
1444. Merge new summary into existing sections rather than regenerating
1455. Track which information came from which compression cycle for debugging
146 
147### When to Use Each Approach
148 
149**Use anchored iterative summarization when:**
150- Sessions are long-running (100+ messages)
151- File tracking matters (coding, debugging)
152- You need to verify what was preserved
153 
154**Use opaque compression when:**
155- Maximum token savings required
156- Sessions are relatively short
157- Re-fetching costs are low
158 
159**Use regenerative summaries when:**
160- Summary interpretability is critical
161- Sessions have clear phase boundaries
162- Full context review is acceptable on each compression
163 
164### Compression Ratio Considerations
165 
166| Method | Compression Ratio | Quality Score | Trade-off |
167|--------|-------------------|---------------|-----------|
168| Anchored Iterative | 98.6% | 3.70 | Best quality, slightly less compression |
169| Regenerative | 98.7% | 3.44 | Good quality, moderate compression |
170| Opaque | 99.3% | 3.35 | Best compression, quality loss |
171 
172The 0.7% additional tokens retained by structured summarization buys 0.35 quality points. For any task where re-fetching costs matter, this trade-off favors structured approaches.
173 
174## Examples
175 
176**Example 1: Debugging Session Compression**
177 
178Original context (89,000 tokens, 178 messages):
179- 401 error on /api/auth/login endpoint
180- Traced through auth controller, middleware, session store
181- Found stale Redis connection
182- Fixed connection pooling, added retry logic
183- 14 tests passing, 2 failing
184 
185Structured summary after compression:
186 
187```markdown
188## Session Intent
189Debug 401 Unauthorized error on /api/auth/login despite valid credentials.
190 
191## Root Cause
192Stale Redis connection in session store. JWT generated correctly but session could not be persisted.
193 
194## Files Modified
195- auth.controller.ts: No changes (read only)
196- middleware/cors.ts: No changes (examined)
197- config/redis.ts: Fixed connection pooling configuration
198- services/session.service.ts: Added retry logic for transient failures
199- tests/auth.test.ts: Updated mock setup
200 
201## Test Status
20214 passing, 2 failing (mock setup issues)
203 
204## Next Steps
2051. Fix remaining test failures (mock session service)
2062. Run full test suite
2073. Deploy to staging
208```
209 
210**Example 2: Probe Response Quality**
211 
212After compression, asking "What was the original error?":
213 
214Good response (structured summarization):
215> "The original error was a 401 Unauthorized response from the /api/auth/login endpoint. Users received this error with valid credentials. Root cause was stale Redis connection in session store."
216 
217Poor response (aggressive compression):
218> "We were debugging an authentication issue. The login was failing. We fixed some configuration problems."
219 
220The structured response preserves endpoint, error code, and root cause. The aggressive response loses all technical detail.
221 
222## Guidelines
223 
2241. Optimize for tokens-per-task, not tokens-per-request
2252. Use structured summaries with explicit sections for file tracking
2263. Trigger compression at 70-80% context utilization
2274. Implement incremental merging rather than full regeneration
2285. Test compression quality with probe-based evaluation
2296. Track artifact trail separately if file tracking is critical
2307. Accept slightly lower compression ratios for better quality retention
2318. Monitor re-fetching frequency as a compression quality signal
232 
233## Integration
234 
235This skill connects to several others in the collection:
236 
237- context-degradation - Compression is a mitigation strategy for degradation
238- context-optimization - Compression is one optimization technique among many
239- evaluation - Probe-based evaluation applies to compression testing
240- memory-systems - Compression relates to scratchpad and summary memory patterns
241 
242## References
243 
244Internal reference:
245- [Evaluation Framework Reference](./references/evaluation-framework.md) - Detailed probe types and scoring rubrics
246 
247Related skills in this collection:
248- context-degradation - Understanding what compression prevents
249- context-optimization - Broader optimization strategies
250- evaluation - Building evaluation frameworks
251 
252External resources:
253- Factory Research: Evaluating Context Compression for AI Agents (December 2025)
254- Research on LLM-as-judge evaluation methodology (Zheng et al., 2023)
255- Netflix Engineering: "The Infinite Software Crisis" - Three-phase workflow and context compression at scale (AI Summit 2025)
256 
257---
258 
259## Skill Metadata
260 
261**Created**: 2025-12-22
262**Last Updated**: 2025-12-26
263**Author**: Agent Skills for Context Engineering Contributors
264**Version**: 1.1.0
265 
266