Context compaction backfires: Billion Context plugin sends coding agent into a loop
Iory1998 · reddit · 2026-10-03
The author uses DeepSeek Harness (supporting local models) for vibe coding, but with 262K context windows, long sessions need memory management. DSH's manual compaction fails after a few rounds, so they tried the Billion Context plugin.
- It's a heavy compression skill that nudges the LLM to compact every few turns, each pass taking 5-10 minutes and stretching sessions significantly
- After many rounds, the model spends most of its working context unpacking compressed text, which triggers re-compression — ending in a loop
- The author concludes compaction isn't the answer and argues memory is the most under-resourced area; sub-agents returning summaries help but the orchestrator's window still needs managing
The post solicits community tips on context/memory management for DSH or other harnesses.
More from coding & agent
- Real-world agent test: one nails church revenue analysis one-shot — MikkoH · 2026-10-03
- Agent workflow: triage Slack bugs, auto-open draft PRs when info suffices — gabrielchua · 2026-10-03
- Meta Muse hits 5M downloads as AI feedback loops could build a fast moat — RihardJarc · 2026-10-03
- Context.dev launches Highlights web-scraping API built for AI agents and LLMs — DanielLockyer · 2026-10-03
- Arize ships hosted MCP server, explains MCP vs CLI vs skills — aparnadhinak · 2026-10-03
- Muse Agent Shops, Pays via Stripe, Books Car Recall Service and Sells on Facebook Marketplace — IanAndrewsDC · 2026-10-03