Qwen Code users push for a real "disable auto-compaction" switch as hard tier fires at 85%
edwardyoon · ghdev · 2026-09-24
- A feature request on QwenLM/qwen-code argues that context.autoCompactThreshold only tunes the auto tier of the three-tier ladder (warn/auto/hard), and the hard tier always fires around 85% — there is currently no supported way to run a session that never auto-summarizes.
- The proposal adds an explicit switch like context.autoCompaction.enabled: false that skips all automatic compaction tiers while keeping manual /compact, following the common harness precedent of separating auto and manual triggers.
- The requester pairs Declarative Attention with KV cache offload on their inference backend, evicting long-lived context to an external store — but Qwen Code's forced compaction reintroduces the lossy-summary problem they're trying to eliminate.
More from coding & agent
- AI compliance startup CompAI hits 1,000+ paying business customers in under a year — JosephJacks_ · 2026-09-25
- Sleep Data for Personal AI? Separating Daily Signals from Hard Rules — sujingshen · 2026-09-25
- Markdown Is the New Source Code: How to Manage Runbooks, Prompts and Rules for Agents — arpit_bhayani · 2026-09-25
- The personal agent supercycle needs hard authorization boundaries, not autonomy — sujingshen · 2026-09-25
- Mnemos.Field nears launch: a virtual world where humans and AI agents both register and participate — RileyRalmuto · 2026-09-25
- SemIf open-sources a Jev-style interface: typed option probabilities from a 4B model, no JSON parsing — JeremyCMorgan · 2026-09-25