Compiled entirely from public activity on meta.discourse.org, X, and GitHub.
💬 meta.discourse.org
Sam responded to concerns about the AI caption bot, noting that the subsystem is actively being reworked. He also confirmed that administrators will get controls to rebake captions.
🐦 On social
Social activity focused on practical AI engineering: local benchmarks of Gemma, Qwen, and Laguna, agent planning and shared-browser workflows, and dynamically translated Discourse summaries. A particularly active reply thread worked through Ruby web-server trade-offs for metrics ingestion, revising the recommendation as experiments exposed dependency, protocol, and security concerns. Replies also shared reproducible evaluations and implementation links, turning discussions into concrete technical follow-ups.
- 10 posts and 15 replies captured in the last 7 days
- Top by engagement: I have a repro here: https:// wasnotwas.com/writing/sa…, sadly got similar results on a novel benchmark, I benched Gemma 4 / Qwen 3.6 / Laguna XS on my 4090. L…
Most engaged tweets:
- reply to @PengmingWang — I have a repro here: https:// wasnotwas.com/writing/same-s core-different-brains-coding-agents-on-an-rtx-4090/ … source of eval is on github
0 likes · 0 reposts · 0 replies - reply to @populartourist — sadly got similar results on a novel benchmark
0 likes · 0 reposts · 0 replies - post — I benched Gemma 4 / Qwen 3.6 / Laguna XS on my 4090. Laguna has a bit of catching up to do.
0 likes · 0 reposts · 0 replies - reply to @HamelHusain — graph engineering is the future, strong agree
0 likes · 0 reposts · 0 replies - post — I think I just stumbled upon a less addictive way of consuming X content…. my agent and I share a specific browser session which both of us can drive.
0 likes · 0 reposts · 0 replies
🛠️ GitHub — Sam’s Commits
samsaffron/term-llm
Over the last seven days, Sam focused on making long-running web and terminal sessions more reliable, hardening concurrent workflows, stream lifecycle handling, response durability, and shutdown behavior. He expanded planning support with persisted plans and a live web plan panel, while improving usability through sticky prompt navigation, responsive menus, and smoother terminal resizing. He also broadened integrations with a Cursor Agent CLI provider and fixes for DeepSeek/vLLM, installation, and release tooling.
Key commits:
2a7d2fd— add sticky user prompt navigation2a6f38e— fix: debounce alt-screen resize reflow6562305— fix: preserve adjacent tool grouping6eaeddd— fix: make web responses durable across lifecycle racesd0c6e39— add Cursor Agent CLI provider
discourse/discourse
Sam focused heavily on Discourse AI, making summaries locale-aware, preserving output language, adding GPT‑5.6 support, improving vLLM compatibility, and ensuring agents finish cleanly when token budgets are exhausted. He also added an audited workflow escape hatch for posting despite normal permission restrictions and closed a security gap that exposed raw email content for inaccessible posts. Smaller test-stability fixes rounded out the week.
Key commits:
01632b1— SECURITY: PostsController#raw_email leaks raw emails of inaccessible to mods (#41962)406c5eb— DEV: fix flaky test (#41902)4177377— FEATURE: Add locale-aware AI summaries (#41893)89189bf— DEV: deflake spec (#41866)c9145ca— FIX: Improve vLLM completion handling (#41862)
discourse/dv
Sam expanded dv’s AI-agent support by integrating the Grok CLI, including prompt and interactive modes, XAI credential passthrough, and synchronized authentication/configuration files. He also added versioned configuration migration rules and comprehensive tests so existing installations gain the Grok mappings without unexpectedly restoring settings users deliberately removed.
Key commits:
b92fc46— feat(agents): add Grok CLI support
discourse/discourse_docker
Sam focused on modernizing and hardening Discourse’s container base image. He added a pinned custom build of libheif with HEVC/AV1 support, upgraded ImageMagick, libvips, and Redis, and then rolled the resulting base image out across the default launcher and service templates.
Key commits:
8e01602— Bump default base image to discourse/base:2.0.20260726-0220 (#1091)ae00b9f— DEV: build and pin libheif 1.23.1 (#1090)
SamSaffron/dotfiles
Sam focused on making his agent-driven development environment more reliable: syncing Grok credentials into Discourse containers, enabling planning support, and extending response timeouts for long-running tasks. He also fixed shared-browser startup by sequencing Hyprland environment propagation with the proxy restart. Smaller maintenance included refreshing Neovim dependencies and adding an MIT license.
Key commits:
1768bb3— get the shared browser working17e1bb7— raise timeoutf9fff36— missing planning tooldb8f7a6— docs: add MIT license952e4d0— updates
🤖 Jarvis — Public Repo Work
Agent-authored public commits, typically guided by Sam during implementation work.
SamSaffron/term-llm
Under Sam’s direction, Jarvis focused primarily on making term-llm’s web sessions faster and more reliable, introducing revision-aware transcript synchronization and tightening ownership/reconciliation of streamed, persisted, cross-tab, and resumed responses. The work also hardened long-running sessions and transport behavior, fixing WebRTC connection-slot leaks, provider connection churn, Ollama reasoning replay, and retry artifacts in chat history. Smaller UI improvements streamlined diff controls and preserved mobile access to the model picker.
Key commits:
3015ae3— fix: Abandoned WebRTC handshakes permanently consume connection slots (#965)4c1b3cf— fix: make terminal response ownership absolute (#964)0b12d65— restrict web transcript persistence to client intent (#963)4b443eb— fix: project one compaction boundary (#962)54904b3— retire stale persisted response overlays (#961)
sam-saffron-jarvis/weird-agent-bench
Sam directed Jarvis to publish the initial Weird Agent Bench release: a reproducible harness comparing coding agents across six standard and deliberately unusual Ruby, Go, and Rust tasks. The work packaged clean-room grading with held-out tests, automated run/grade scripts, documented local-model configurations and methodology, and published results comparing Gemma, Qwen, and Laguna models on an RTX 4090.
Key commits:
9c59d63— Initial public benchmark release
sam-saffron-jarvis/jarvis-browser-proxy
Sam directed Jarvis to harden browser-proxy startup and upgrades, especially around Hyprland/Wayland: desktop-session variables are now refreshed reliably, Lua-based Hyprland configs are supported, and upgrades run non-interactively. The proxy was also made more resilient by adopting already-running Chromium processes and treating a healthy CDP endpoint as authoritative, preventing duplicate browser launches and improving status accuracy.
Key commits:
5ad873f— fix: adopt running browser processes (#4)a7da23a— fix: configure browser startup from Hyprland Lua (#3)86d95c0— fix: make upgrades non-interactive (#2)1ebb200— fix: refresh Wayland environment on desktop login (#1)
⤴️ GitHub — Pull Requests
10 PRs this week:
- ✅ discourse/discourse_docker#1091 (diff) — Bump default base image to discourse/base:2.0.20260726-0220 merged
- ✅ discourse/discourse_docker#1090 (diff) — DEV: build and pin libheif 1.23.1 merged
Build libheif with HEVC and AV1 support for ImageMagick and libvips instead of relying on Debian backports. Package and hold the runtime library with a guard package to prevent accidental downgrades, and verify the expected version in the final image…
- ✅ discourse/discourse#41962 (diff) — SECURITY: PostsController#raw_email leaks raw emails of inaccessible to mods merged
Prevent unauthorized access to raw emails by ensuring the mod has permission to view the underlying post before granting access to its raw email data. This addresses an issue where members of privileged groups could bypass visibility restrictions to …
- 🟢 discourse/discourse#41963 (diff) — UX: Stream workflow execution progress opened
Publish admin-only execution and step updates over MessageBus so workflow lists and detail views reflect active, waiting, and completed runs without manual reloads. Show live elapsed times and loading indicators, then refresh authoritative details wh…
- ✅ SamSaffron/term-llm#947 (diff) — fix: Memory mining rescans every completed session every 30 minutes closed
- Added a batched SQLite session-store capability that returns the maximum persisted message sequence for up to 500 requested sessions per query. - Added a single-query memory mining-state listing and use it to filter fully mined sessions before load…
- ✅ SamSaffron/term-llm#946 (diff) — fix: Mining limit permanently starves older unmined sessions closed
- Stop truncating completed-session candidates before mining state and pending messages are checked. - Apply
--limitto sessions that actually reach an extraction attempt, so fully mined sessions no longer consume the allowance. - Add a regression …
- Stop truncating completed-session candidates before mining state and pending messages are checked. - Apply
- ✅ discourse/discourse#41902 (diff) — DEV: fix flaky test merged
- ✅ discourse/discourse#41893 (diff) — FEATURE: Add locale-aware AI summaries merged
Store summaries by locale so topics can serve and regenerate the appropriate localized version. Generate gists for source and translated locales, and make backfills handle each locale independently.
- ✅ SamSaffron/term-llm#939 (diff) — fix: Exited MCP servers remain permanently marked ready closed
- Start a watcher on each successfully initialized MCP
ClientSessionusing the SDK’sClientSession.WaitAPI. - When the active session terminates unexpectedly, clear that exact client’s running session and cached tools, mark the server failed wit…
- Start a watcher on each successfully initialized MCP
- ✅ SamSaffron/term-llm#938 (diff) — fix: Chat starts MCP servers before installing their sampling handler closed
Install the interactive chat MCP sampling provider immediately after loading MCP configuration and before enabling any configured MCP servers. This preserves the existing manager and client behavior while ensuring startup-created clients receive a sa…
🐛 GitHub — Issues
No issue activity this week.
👀 GitHub — Reviews
4 reviews this week:
- discourse/discourse-shared-edits#158 — Bump rack from 3.2.5 to 3.2.6 approved
- discourse/discourse#41966 — SECURITY: PostsController#raw_email leaks raw emails of inaccessible to mods [ba approved
- discourse/discourse#41965 — SECURITY: PostsController#raw_email leaks raw emails of inaccessible to mods [ba approved
- discourse/discourse#41964 — SECURITY: PostsController#raw_email leaks raw emails of inaccessible to mods [ba approved