Context Hygiene & Session Management

Context Hygiene & Session Management

Purpose

AI context windows degrade as conversations grow. After ~30+ exchanges, the AI loses track of earlier constraints, starts hallucinating details, and wastes tokens re-reading irrelevant history.

This skill teaches Anti-Gravity to actively manage session boundaries automatically — securing all progress to disk before the user moves to a clean chat window.

When to Advise a Session Split

Recommend the user open a new chat window when:

1. The task changes significantly. If we finished debugging and are now building a new feature, that is a new session.
2. The conversation exceeds ~30 exchanges. Context quality degrades. Proactively say: "We've covered a lot. I recommend starting a fresh chat for the next piece so I stay sharp."
3. Multiple unrelated tasks are being discussed. Each task deserves its own clean context.
4. A micro-task sequence is complete. After finishing a planned set of tasks, suggest: "Good stopping point. For the next batch, a fresh session will give you better results."

How to Execute a Handoff (The Zero-Friction Method)

Never ask Beloved to manually copy-paste a "Handoff Brief." Instead, perform these two steps automatically:

Step 1: Secure Long-Term Knowledge (Memory)

Before suggesting a split, ensure all permanent knowledge from this session is written to disk:

• Did we make an architectural choice? Write it to .agents/memory/decisions-log.md.
• Did we learn what not to do? Write it to .agents/memory/mistakes-to-avoid.md.
• Did we establish a new convention? Write it to .agents/memory/common-patterns.md.

Step 2: Secure Immediate State (Workflow)

Update .agents/workflow-state.json so the next session knows exactly where to resume. Ensure the notes, key_decisions, and phases fields are perfectly up to date.

Reference format from GLOBAL_MEMORY.md:

{
  "workflow": "build-feature",
  "current_phase": "IMPLEMENT",
  "completion_pct": 64,
  "notes": "Finished the UI layout. Next session must focus on wiring the submit button to the API.",
  "key_decisions": ["Used Flexbox instead of Grid for the main container"],
  "blockers": []
}

Step 3: Inform the User

Once the files are updated, tell Beloved:

"I've secured our progress in workflow-state.json and logged our decisions to Memory. This chat is getting long — I recommend opening a new chat window to continue. The new session will automatically read the state file and pick up exactly where we left off."

When NOT to Split

• In the middle of a single atomic task (don't interrupt flow)
• When the user is in a sprint and splitting would break momentum
• When the conversation is short and focused