Skill: Context Compress

What This Skill Does

Mid-session reset for long conversations. When a session grows too long and context becomes expensive, this skill compresses the relevant context into a compact briefing — keeping what matters, discarding noise.

Think of it as generate-handover but for the middle of a session, not the end.

When to Use

• When a session is running long (50+ messages)
• When the conversation has explored multiple tangents
• When you notice repetition or context confusion
• Before switching to a different task within the same session

Do NOT use at session end — use generate-handover for that.

Execution Model

• Always: the primary agent runs this skill directly.
• Token budget: ~2-3k tokens to produce the summary. Goal: save 10-20k+ tokens going forward.
• Output: chat-based briefing (not persisted — it IS the new context).

Workflow

Step 1: Identify Active Context

1. Active task: what are we working on right now?
2. Key decisions made: what was decided and why?
3. Current file state: which files were modified?
4. Pending items: what still needs to be done?
5. Blockers: any unresolved issues?

Step 2: Identify Noise

Flag context that can be discarded:

• Exploratory discussions that led nowhere
• Debugging steps that were dead ends
• Verbose tool output that's been processed
• Repeated explanations and off-topic tangents

Step 3: Generate Compressed Context

## Session Context (Compressed)

### Active Task
<one sentence: what we're doing right now>

### Progress
- Completed: <completed step>
- In Progress: <current step>
- Pending: <pending step>

### Key Decisions
- <decision>: <rationale>

### Modified Files
- `<file>`: <what changed>

### Next Steps
1. <immediate next action>
2. <after that>

Step 4: Confirm with User

Use the question tool: "I've compressed the session context. Anything important I missed?"

Rules

1. Relevance over completeness: include only what's needed to continue the current task.
2. Decisions are critical: always preserve key decisions and their rationale.
3. File state matters: list modified files so the agent doesn't re-read unchanged files.
4. Be honest about what's lost: note when detail is omitted.
5. No built-in explore agent: do NOT use the built-in explore subagent type.