## Success Criteria

### Context Selection (25 points)
- Only relevant information is included
- No unrelated files or content
- Information is specific to the task at hand
- Recency and specificity are prioritized

### Context Structure (25 points)
- Task description is clear and first
- Current state is accurately represented
- Requirements are specific and testable
- Supporting context is ordered by relevance

### Summarization Quality (20 points)
- Large files are summarized appropriately (Level 1-3)
- Summaries preserve essential details
- Line numbers are included for code snippets
- Full content is used only when necessary

### Token Efficiency (15 points)
- Context stays within reasonable token budgets
- No duplication of information
- References used instead of full content dumps
- Progressive disclosure is applied for multi-turn tasks

### Multi-Turn Management (15 points)
- Context carries forward relevant information
- Previous turn outputs are included when needed
- Stale information is dropped
- Context grows with task complexity

## Evaluation Scenarios

### Scenario A: File Dump
- Agent receives 500+ lines of unrelated code
- Task description is buried at the bottom
- No structure or organization
- **Result:** FAIL — violates selection, structure, and efficiency criteria

### Scenario B: Well-Structured Context
- Task is clear and first
- Only relevant files included with line numbers
- Summaries used for large files
- Requirements are specific
- **Result:** PASS — meets all criteria

### Scenario C: Missing Context
- Agent receives too little information
- Key files are not included
- Agent has to make assumptions
- **Result:** FAIL — violates selection criteria (insufficient)

### Scenario D: Good First Turn, Bad Second Turn
- Turn 1 has excellent context
- Turn 2 re-includes all of Turn 1's context
- Turn 2 does not include the output from Turn 1
- **Result:** CONDITIONAL — first turn passes, second turn needs improvement

## Red Flags (Automatic Deductions)
- Entire files included when only a function is needed: -10 points
- Task description not in the first section: -5 points
- No line numbers for code references: -5 points
- Contradictory information in context: -10 points
- Context exceeds 80% of available window: -10 points
- Information repeated across sections: -5 points
- Stale information from previous turns not removed: -5 points