This release includes a some nice contributions form the community:
- @cphlipot improved the sidebar on mobile
- @candrews added model parsing from the websocket stream in v1/realtime
- @heshiming added support for TensorFold metrics parsing
On the backend there was lots of work to finetune LFM3-1.7B for
the new help agent on llama-swap.ai. This includes the training scripts
and data and the new docbot package that has the Help's mcp endpoints.
- PR #1283 ui: dismiss mobile sidebar when user selects an item: closes the mobile sidebar overlay as soon as a nav item is tapped so the selected page comes into focus immediately, because the overlay previously stayed open and made switching pages (eg. activity -> playground) take two taps; no effect on desktop where the sidebar is always visible by @cphlipot
- PR #1304 evals/docs-agent: update cases for current KB and tune the Docs Agent: adds 66 training and 29 holdout eval cases, fixes four wrong assertions, raises
get_docdefault output from 250 to 400 lines (it was cutting off the kubeswap gpu-budget example), corrects the cmdStop,${PORT}macro and kubeswap docs, and points KB pages at the guides holding the answers, lifting gemma-4-12B-r9700's training score from 96.0% to 97.3% by @mostlygeek - PR #1308 evals/docs-agent: add fine-tuning pipeline for a small Docs Agent model: adds the scripts, recorded conversations and run notes used to fine-tune LFM3-1.7B on Docs Agent conversations (lfm3-v2 Q4_K_M scores 95.5% on training cases and 96.1% on the holdout at a 60 s per-question limit), with no weights included, as a reference for tuning future help-agent models by @mostlygeek
- PR #1310 internal/server: route /v1/realtime by model in session.update: fixes 404s before the websocket upgrade for clients that follow OpenAI's documented
/v1/realtimeAPI and connect with?intent=transcriptionbut no?model=, by having llama-swap complete the handshake itself, buffer events until asession.updatenames a model (session.model, then the transcription model keys), and replay them upstream through the normal model chain, reporting failures as OpenAI style error events by @candrews - PR #1312 docbot: new importable package: extracts MCP protocol handling into a stateless
internal/mcpserver.Handlerand adds adocbotpackage that serves llama-swap's docs (and optionally source) as MCP tools over any HTTP route withSystemPrompt()andTools()helpers, a prerequisite for the help agent on llama-swap.ai by @mostlygeek - PR #1313 TensorFold support for pluggable usage handlers: adds a TensorFold usage and metrics extractor with simplified captured-metrics fixtures, rebuilt on the pluggable per-upstream usage handler standard from #1292 after that refactor left the original TensorFold support stale by @heshiming
- dea1363 README.md: add tip for llama-swap.ai: adds a README tip pointing to the browser-based Chat agent powered by an LFM3-1.7B fine tuned on the docs by @mostlygeek
- cef61b3 Revise Chat agent info in README: revises the Chat agent description and links to the fine-tuned model on Hugging Face by @mostlygeek
- c805e62 Update link format for Chat agent in README: links the whole "Chat agent on llama-swap.ai" phrase to the chat docs page by @mostlygeek