Context basics

Each request is built from context, which is the model input window for that turn.

• Conversation history that remains relevant to the active task.
• Current request text, plus any attachments (files, diffs, text, images).
• System/runtime instructions, selected mode/model behavior, and persistent guidance such as memories.
• Recent tool outputs and other execution artifacts needed for continuity.

• Every turn adds new user text, assistant output, and tool results.
• Large attachments and pasted blobs accelerate growth.
• Long multi-topic sessions retain stale details that compete for attention.
• Higher-context turns cost more because larger input windows are reprocessed.

• 70%+: warning zone (quality and cost risk starts rising).
• 90%+: critical zone (strong chance of drift and unnecessary spend).
• 95%+: sending is blocked until pressure is reduced.

Quick rule:
- <70%: continue normally
- 70-89%: tighten prompts + narrow scope
- 90-94%: compact before next major step
- >=95%: compact or start a fresh session

• Auto-compaction defaults to 85% context usage for coding requests.
• Auto-compaction is skipped for max-mode requests.
• You can manually compact to summarize history and free room for new work.
• Max mode prioritizes continuity by using larger context, but usually increases token spend.

• Keep one larger objective per session to avoid topic drift.
• Use focused attachments instead of broad file dumps.
• Compact at milestone boundaries before starting a new sub-task.
• When a task is done, open a fresh session for the next objective.
• If quality drops, restate non-negotiable constraints explicitly after compaction.

Bad: "Let us keep this same session for every task this week."
Better: "We finished feature A. Start a new session for feature B with clean context and updated constraints."