Open Source Reports
May 2026

Another busy month for LMPlayground - four releases shipped to Google Play, this time focused on running well across more devices. The UI grew adaptive layouts for tablets, foldables, and Chromebooks, and runtime CPU variant selection brings up to 2× faster generation by dlopen’ing the best ARM feature level at startup. Inference now also lives in its own isolated :llama process, so a native crash can’t take the whole app down. The model lineup gained GPT-OSS 20B (with Harmony reasoning), and oversized models now memory-map and run slowly instead of refusing to load. llama.cpp tracked from b8929 to b9411 across the month.

BonjourBrowser shipped 3.0.3, a critical fix that restored mDNS service discovery on Android 16+ - the post-3.0.0 raw-socket resolver had been silenced by the Wi-Fi multicast filter and a port collision with the system mDNS daemon.

The two Spark-related projects each shipped two releases, tracking the Spark CLI from 1.1.0 to 1.2.1. The 1.1.0 releases crossed from read-only into triage/write territory: spark-cli-skills added 11 triage recipes and an 8-role persona library, while spark-claude-extension gained draft/comment/action tools. The 1.2.1 releases then added attachment reading, saved templates, and shared drafts to both.

LMPlayground

• 1.5.3 — Run llama.cpp in a separate :llama process for crash isolation. Hidden foreground-service notification showing the loaded model name and RAM usage. Softened the low-RAM gate from a hard refusal to an override dialog. Updated llama.cpp b8929 → b9009.
• 1.5.4 — Friendly “Couldn’t load model” dialog when a GGUF is corrupt or uses an unrecognized architecture, instead of a native crash. Caught inference-process death at the AIDL proxy and surfaced it as a recoverable error. Updated llama.cpp b9009 → b9089 (RMS_NORM + MUL fusion CPU speedup, KleidiAI v1.24.0).
• 1.6.0 — Tablet, foldable, and Chromebook support — a permanent session sidebar at ≥ 840 dp, hinge-aware sidebar width via androidx.window, and freeform windowing. Runtime CPU variant selection via GGML_CPU_ALL_VARIANTS — dlopens the best ARM feature level at startup (dotprod, i8mm/SVE) for up to 2× faster generation. Localized Play Store tablet screenshots in 28 locales. Updated llama.cpp b9089 → b9192.
• 1.6.1 — Added GPT-OSS 20B (native MXFP4) with the Harmony “analysis” channel rendered as a collapsible thinking section. Large-model support — models over the RAM budget now memory-map and run (slower) instead of failing to start. Leaner downloads — stream directly into the model folder and atomically rename on completion, halving peak disk use. Updated llama.cpp b9192 → b9411.

BonjourBrowser

• v3.0.3 — Fixed mDNS service discovery on Android 16+ where the home screen showed no services. The raw-socket resolver now holds a MulticastLock (with the new CHANGE_WIFI_MULTICAST_STATE permission) and binds to an ephemeral source port so responders send unicast replies per RFC 6762 §5.1. Reworked the open-source licences screen, replacing the ANR-prone WebView with a lightweight TextView.

spark-cli-skills

• 1.1.0 — Expanded the use-spark base skill to Spark CLI 1.1.0 triage-level write commands (drafts, team comments, email/contact actions). Added 11 triage recipes (inbox-zero, vacation-catchup, new-sender-review, draft-batch, meeting-followup, shared-inbox-triage, delegate-and-track, priority-tuning, newsletter-cleanup, notification-hygiene, label-organize) and an 8-persona library (exec-assistant, founder, team-lead, project-manager, sales-rep, support-agent, meeting-manager, freelancer). Introduced per-account access levels (read-only / triage).
• 1.2.1 — Updated the base skill to Spark CLI 1.2.1: new attachment, templates, and template commands. search gained a keyword-less list mode (paged table across all folders/accounts). draft gained team-shared drafts, collaboration flags, and template support. thread and meeting now accept Spark deep links.

spark-claude-extension

• 1.1.0 — Added write access on top of 1.0.0’s read-only tools: draft (new/reply/forward/update with attachments and markdown), comment (team chat with auto-share), action (archive, pin, snooze, move, label, assign, share, mark done, reclassify), and contact-action. Per-account/shared-inbox access levels (read-only / triage). Added a privacy policy and tool annotations.
• 1.2.1 — Surfaced 1.2.1 tools: attachment reading directly through MCP (images, audio, text, PDFs), templates and template listing, draft template application and sharing, and the new keyword-less search list mode across every folder and account.