| name | context |
| description | Context window manager. Snapshots session state, compacts conversation, and resumes via hook. Use when context exceeds 70% or a large new task begins. |
Context — Context Window Management
When to Trigger
- Context window > 70% capacity
- Long conversation with many tool calls
- Before starting a large new task in an existing session
- User asks to "clean up" or "start fresh"
Process
1. Assess
- Estimate current context usage
- Identify what's consuming the most context (large file reads, verbose outputs)
2. Preserve
Before compacting, ensure critical state is saved:
- Current task and progress →
$HARNESS_DIR/sessions/
- Key decisions made → note in conversation
- Files being worked on → list explicitly
3. Compact
Suggest or trigger compaction with a clear summary:
Summary for compaction:
- Working on: [task description]
- Files modified: [list]
- Current status: [what's done, what remains]
- Key decisions: [important choices made]
- Next step: [what to do after compaction]
4. Resume
After compaction, the resume hook will reload:
- Session snapshot from
$HARNESS_DIR/sessions/
- Project memory from
$HARNESS_DIR/memory/
- Evolved skills from
$HARNESS_DIR/evolved/
Anti-Rationalization
| Excuse | Rebuttal | What to do instead |
|---|
| "I still have context left" | Quality degrades well before the hard limit. 70% is the trigger. | Compact proactively. Better to lose 5 min than lose coherence. |
| "Compacting will lose important context" | That's why you preserve first. The resume hook restores it. | Save state → compact → resume loads snapshot + memory + skills. |
| "I'll just re-read the files" | Re-reading 5 large files wastes 30%+ of your fresh context. | Summarize key facts before compacting. Read only what you need after. |
Evidence Required
Before compacting, confirm ALL of these:
Compacting without a summary = guaranteed context loss.
Red Flags
- Continuing with 90%+ context without compacting (quality degrades)
- Compacting without saving state (lose track of progress)
- Re-reading large files after compaction (defeats the purpose)