Context Compress

Maximize signal, minimize noise in context windows.

When to Use

• Context window approaching limits
• Starting session with prior work
• Large codebase exploration consuming tokens
• Need to preserve decisions across checkpoints

Strategies

1. Summarize Aggressively

Bad:  Full stack trace (50 lines)
Good: "TypeError at src/api.ts:42 — missing null check on response.data.user"

Bad:  Entire file read (500 lines)
Good: "File has 200 lines, key function at line 85-120, uses X pattern"

Bad:  Running every test individually
Good: "All 47 tests pass, coverage at 82%"

2. Layer Information

Critical (always keep):
  - Current task goal
  - Active blockers
  - Key architectural decisions
  - Modified file paths

Important (summarize when needed):
  - Prior decisions and rationale
  - Test results
  - Exploration findings

Ephemeral (drop when tight):
  - Full file contents (re-read on demand)
  - Verbose command output
  - Intermediate search results

3. Use Memory System

• Store durable facts in MEMORY.md
• Use notes.md for session observations
• Reference memory instead of re-deriving
• Search memory before re-exploring

4. Parallel Context via Subagents

• Spawn subagents for heavy exploration
• Each gets clean context
• Synthesize findings, not raw output
• Don't pass full conversation to every subagent

Token Budget

Rule: If reading it twice, summarize it once.

Session Handoff

When context is tight:

1. Write findings to notes.md or task progress
2. Summarize what's done and what's next
3. Reference specific file paths and line numbers
4. Start fresh session with focused prompt

Anti-Patterns