Skip to main content
Every model can only read a limited amount of conversation at once (its context window). Long tasks fill it up. Sixth watches how full it is and compacts the conversation automatically, so a long task keeps going instead of failing or asking you to start over.

Seeing how full the context is

The task header shows a small context ring next to the task title, with a percentage. The ring turns amber from 85% onwards. Click the ring (or the task title) to open the task details. Under Context there’s a bar showing tokens used against the model’s context window, for example 76.4k / 200k, along with the task’s input, output and cache token totals.

Automatic compaction

When the previous response used more than 92% of the context window, Sixth compacts before sending the next request:
  1. The chat shows Compacting history with the current usage.
  2. Sixth summarises the older part of the conversation, keeping what matters: decisions, files touched, problems solved, what’s left to do.
  3. The summary replaces those older turns in what the model sees, and the task continues.
If a summary can’t be made (for example when you use your own API key), Sixth drops the oldest turns instead, so the task can always continue. If a provider rejects a request because the input is too large, Sixth compacts and retries the request on its own. While that runs, the chat shows Compacting conversation context with progress, then Context compacted — resuming.
Compaction only changes what the model reads. Your chat history stays as it was, so you can still scroll back through the full conversation.

Compacting manually with /smol

Type /smol (alias /compact) to compact on demand. It’s useful when you’ve finished one phase of work and want to free up room before the next. Sixth writes a detailed summary of the conversation so far, and once it’s accepted, that summary replaces the older history. See /smol.

Starting fresh with /newtask

When the work has naturally moved on to something new, /newtask starts a new task that carries over a summary of the current one. See /newtask.