Improved chat template for Laguna XS/S 2.1 adds forced thinking and stability fixes
arbv · reddit · 2026-09-28
Reddit user arbv released an improved chat template for Poolside's Laguna XS and S 2.1 models to fix inconsistent reasoning — by default the models lazily skip thinking exactly when they need it most.
Key changes:
- A forcethinking toggle using a community-discovered prompt trick, plus a reasoningeffort preset (none/auto/max) for easier use in Pi and other harnesses
- A preservethinking toggle: the original template permanently preserves historical reasoning (great for prefix caching and agentic tool loops, but it shreds your context window in normal chat)
- Basic validation against smuggled control tokens across roles, disableable via allowinjection: true
- Defaults match upstream behavior, so it works as a drop-in replacement
The template is open-sourced on Hugging Face, with recommended llama.cpp sampling settings (BF16 and quantized) and a Pi models. snippet in the README. The author also pinged Laguna's post-train lead for a data point.
More from coding & agent
- Coding agents write defensive code with overly broad exception catches, and review agents still LGTM it — srchvrs · 2026-09-28
- Asked Claude for a small scrape, woke up to 'three weeks remaining' at 60 MB/s — generativist · 2026-09-28
- Dad and kids vibe code a Zelda-style open-world game with Opus 5.5 in one day — IanArawjo · 2026-09-28
- PSClaudeCode: a 250-line PowerShell reimplementation of Claude Code's agent loop — dfinke · 2026-09-28
- Dev realizes his 2023 tool already draws AI tool-calling as sequence diagrams — dfinke · 2026-09-28
- Dev uses Claude Opus 5.5 to build animated UX redesign walkthroughs, hailed as a UI focus masterclass — jonathan_wilke · 2026-09-28