Reliable conversation compaction for long AI chats

When a conversation outgrows its model's context, a summarizer writes a checkpoint and the conversation continues from there. In production, the default utility model could spend its whole output budget on hidden reasoning and stop before the summary was complete. Compaction then failed the run, and every later message in that conversation failed the same way. Compaction now always lands a checkpoint, and never writes a partial one.

How it works

• Bounded summaries. The summarizer runs at its lowest thinking setting with a 16K-token output cap and a prompt asking for under 4,000 words.
• A fallback chain. If the utility model's summary errors, is cut off, or comes back empty, the run's own model writes it instead, and a context overflow retries once with half the input.
• A last resort. If no summarizer finishes, the checkpoint is a word-for-word excerpt of the start of the conversation and the latest work, clearly labeled, so the conversation keeps working.
• Clean interruptions. Cancel, timeout, and handoff abort the summary without writing anything, and Stop now interrupts a compaction in progress.

Also shipped

• Long workflows keep platform access. A code workflow's platform key now lasts as long as the run itself, instead of expiring after 10 minutes.
• Viewing a workflow never turns it off. Opening a workflow's page mid-edit could disable it; reads no longer change anything.
• Fixes from our latest build eval: 22 builds on Claude Sonnet 5.5, all goals met.