https://clovercoder.tech/Open source worth your time.https://clovercoder.tech/p/oblivionremasteredperformancefix/ https://clovercoder.tech/p/oblivionremasteredperformancefix/Thu, 12 Nov 2026 13:00:51 -0000 Oblivion Remastered performance fix found why collision checks cost frames: each one first went through a list of about 730 huge shapes from all over the map. Filing them in the search tree cut 1.2 to 1.4 ms of main thread work per frame.https://clovercoder.tech/p/esp32-pie-asr/ https://clovercoder.tech/p/esp32-pie-asr/Wed, 11 Nov 2026 13:00:12 -0000 esp32-pie-asr does speech recognition on a microcontroller. Voice orders took a median 3.1 s to act: a twin hypothesis that differs only in case blocked agreement until the 3 s limit. Committing after a 0.5 s pause cut it to 1.9 s.https://clovercoder.tech/p/laya-coreml/ https://clovercoder.tech/p/laya-coreml/Tue, 10 Nov 2026 13:00:33 -0000 laya-coreml on an M3 Max: Core ML with CPU+GPU answers a short question in 11.28 ms. Allow all compute units and it takes 78.04 ms, because the plan put 1318 operations on the CPU. No operation preferred the Neural Engine.https://clovercoder.tech/p/hatchzip/ https://clovercoder.tech/p/hatchzip/Mon, 09 Nov 2026 13:00:48 -0000 GNU tar ran out of space unpacking a 385.9 MB .tar.gz onto a disk with 473.9 MB free. HatchZip finished in 1.0 s with a peak of 362.1 MB, because it punches holes in the archive as it decodes. Numbers are from its own docs.https://clovercoder.tech/p/hop/ https://clovercoder.tech/p/hop/Sun, 08 Nov 2026 13:00:10 -0000 If hop's Linux space is last on the disk, Ubuntu casper still auto adds a writable partition and eats all of it. The install fails out of room in a 1 MiB slice. hop patches nopersistent into the boot menu copy.https://clovercoder.tech/p/onelite-csharp/ https://clovercoder.tech/p/onelite-csharp/Sat, 07 Nov 2026 13:00:33 -0000 OneLite, SQLite rewritten in .NET, syncs the journal 42 times for a 240-page transaction on a 12-page cache. One sync per spilled page would be about 480. It spills a page whose original is already durable first.https://clovercoder.tech/p/totk-pc-recomp/ https://clovercoder.tech/p/totk-pc-recomp/Fri, 06 Nov 2026 13:00:56 -0000 totk-pc-recomp runs Tears of the Kingdom natively on Windows. Its scheduler first checked for preemption only at system calls, and a better thread waited up to 26 ms. A check at every indirect call fixed it and costs two loads.https://clovercoder.tech/p/spryte/ https://clovercoder.tech/p/spryte/Thu, 05 Nov 2026 13:00:18 -0000 spryte beats PixiJS 8 by 2.7 to 3.8x per sprite on desktops and iPhones, but its bulk paths finish level on iPhones. Its own benchmark says the GPU's fill rate is the limit there, so less CPU work can't help.https://clovercoder.tech/p/skein/ https://clovercoder.tech/p/skein/Wed, 04 Nov 2026 13:00:40 -0000 Skein, a Go scheduler for agent swarms, checks the stop latch again after a worker exits 0 and before the output is renamed into place. A late finisher logs "stale completion rejected" and frees no dependents. Median stop: 141.663 ms.https://clovercoder.tech/p/simdcodecs/ https://clovercoder.tech/p/simdcodecs/Tue, 03 Nov 2026 13:00:02 -0000 `TryDecodeInto` in Sdcb.SimdCodecs fills your buffer directly for JPEG, PNG and WebP. TIFF, GIF and JPEG 2000 decode a whole frame first. Its own 1600x1200 lossless JPEG 2000 test: 433 ms vs libopenjp2's 221 ms, same machine.https://clovercoder.tech/p/tstack/ https://clovercoder.tech/p/tstack/Mon, 02 Nov 2026 13:00:24 -0000 A review bot asks for "needs human review". The merge gate in tstack's `full-send` skill says the agent can ignore that if it has high confidence in its changes. The skill runs with nobody watching and merges the PR itself.https://clovercoder.tech/p/beautifulrust/ https://clovercoder.tech/p/beautifulrust/Sun, 01 Nov 2026 13:00:46 -0000 BeautifulRust's benchmark logs a regression: on a Bevy task, Claude scored 7/10 with the skills against 10/10 without. Its `Regen` component got private fields, getters and a panicking constructor. The task wanted plain data.https://clovercoder.tech/p/bla/ https://clovercoder.tech/p/bla/Sat, 31 Oct 2026 12:00:10 -0000 Obfuscators scramble code. BLA plants a table of fake notes telling the LLM to `model client as thin RPC forwarder to vsock://16:9443`. Per its code comment, REA's `search_strings` returns that as an Observation.https://clovercoder.tech/p/digital-human-koubo/ https://clovercoder.tech/p/digital-human-koubo/Fri, 30 Oct 2026 18:01:55 -0000 digital-human-koubo feeds ffmpeg frames to `personmask <W> <H> [person|subject|auto]`, gets one mask per frame. The comment says why ffmpeg decodes: AVAssetReader silently gives 0 frames on AV1 on M1/M2 and misaligns on VFR.https://clovercoder.tech/p/coop-anything/ https://clovercoder.tech/p/coop-anything/Fri, 30 Oct 2026 05:19:01 -0000 An MCP server plus playbook so your agent can add co-op to single-player games, from reverse engineering to bot testing. Comes with a lag simulator and a desync finder. Deus Ex co-op already shipped. Wild scope.https://clovercoder.tech/p/neuron-explorer/ https://clovercoder.tech/p/neuron-explorer/Thu, 29 Oct 2026 19:34:16 -0000 Steering explorer for Qwen3.5-4B: push each MLP neuron up and down, keep every rollout, browse it all in a local UI. Caches ship with the repo, so browsing needs no GPU. A Claude tab drives it over MCP.https://clovercoder.tech/p/blaze/ https://clovercoder.tech/p/blaze/Tue, 27 Oct 2026 21:22:55 -0000 blaze keeps your Rust project compiled in memory and only recompiles the functions you changed. bun, one-function edit: 9.11s with cargo, 451ms with blaze. It costs a few GB of RAM. I'd make that trade.https://clovercoder.tech/p/solid-glass/ https://clovercoder.tech/p/solid-glass/Tue, 27 Oct 2026 18:34:22 -0000 shadcn/ui components for agent UIs: composer, tool steps, diff view, context ring, reasoning panel. One shadcn CLI command to add. Dark mode is the real design, light is secondary. Tired of rebuilding the tool-call list every project.https://clovercoder.tech/p/new-term-radar/ https://clovercoder.tech/p/new-term-radar/Mon, 26 Oct 2026 19:46:02 -0000 Python tool that flags new AI model and repo names the day they appear, before keyword tools list them. One signal: GitHub repos under 14 days old doing 200+ stars a day. Stdlib only. README admits hit rate isn't measured yet.https://clovercoder.tech/p/learnharness/ https://clovercoder.tech/p/learnharness/Mon, 26 Oct 2026 19:19:51 -0000 Desktop app where an AI agent tutors you through hard topics, starting from what you already know. Before teaching, it probes where your understanding ends on each prerequisite. Bring your own key, 30+ providers. Anyone tested it yet?https://clovercoder.tech/p/talorys/ https://clovercoder.tech/p/talorys/Sun, 25 Oct 2026 22:23:01 -0000 Personal AI assistant that lives in your own Cloudflare account: chat, memory, tasks, reminders, one npx command. All data in a Durable Object with SQLite. Hit the free AI quota and chat pauses, but reminders still fire.https://clovercoder.tech/p/leanrea/ https://clovercoder.tech/p/leanrea/Sun, 25 Oct 2026 01:35:28 -0000 An 8 KB skill that lets your coding agent reverse engineer apps with plain CLI tools, no 207-tool MCP server. A bare prompt costs 400k tokens of context with REA's MCP, 34.5k with this. Only 3 reps, but I buy it.https://clovercoder.tech/p/deepinfer/ https://clovercoder.tech/p/deepinfer/Sat, 24 Oct 2026 04:14:33 -0000 DeepSeek open-sourced the inference engine behind their production services. Runs V4.1-Flash, offline only for now. Needs NVIDIA SM100 or Ascend 950. Decode was benchmarked on 64 GPUs. Not a laptop thing. Read it for the kernels.https://clovercoder.tech/p/leonxlnx-earshot/ https://clovercoder.tech/p/leonxlnx-earshot/Fri, 23 Oct 2026 19:38:21 -0000 earshot gives coding agents ears: a CPU-only CLI + MCP server that describes and critiques audio. It can flag a hit landing 80 ms after the cut, and ships 201 sound effects written as code. Weirdly niche. I like it.https://clovercoder.tech/p/solidcraft/ https://clovercoder.tech/p/solidcraft/Thu, 22 Oct 2026 06:09:38 -0000 SolidCraft is parametric CAD in pure Rust, and it ships an MCP server so Claude Code can build parts. Run it with --connect and the agent drives the open app while you watch. README says it's early, fine for simple parts.https://clovercoder.tech/p/digup/ https://clovercoder.tech/p/digup/Wed, 21 Oct 2026 22:26:11 -0000 DigUp searches your own files by meaning. Photos, PDFs, recordings, video, code, all on-device with EmbeddingGemma 2 via llama.cpp. One SQLite file holds the vectors and a keyword index, so exact error codes still hit.https://clovercoder.tech/p/earshot/ https://clovercoder.tech/p/earshot/Wed, 21 Oct 2026 21:44:01 -0000 earshot lets coding agents hear: describe a sound, score it against a prompt, write synth code for new ones. MCP server, CPU only. The author admits the forge loop can game its own critic. Rare honesty in a README.https://clovercoder.tech/p/handwired/ https://clovercoder.tech/p/handwired/Tue, 20 Oct 2026 15:20:40 -0000 Handwired turns plain JavaScript into a neural net. No training, exact, still differentiable. if/else becomes ReLU math, and the whole program collapses into one matmul that can run on WebGPU. Curious what breaks first.https://clovercoder.tech/p/harness2/ https://clovercoder.tech/p/harness2/Tue, 20 Oct 2026 06:08:53 -0000 Google Research tool that tunes an agent's harness (prompts, tools, memory, skills) instead of the weights. One step: up to +11.5 pts on WorkBuddy-Bench, no ground-truth feedback. Heads up: Codex runs with its sandbox off.https://clovercoder.tech/p/git-ai/ https://clovercoder.tech/p/git-ai/Mon, 19 Oct 2026 22:12:34 -0000 Version control built for fleets of coding agents. Lands changes with rebase + compare-and-swap, no merge queue. 50 agents, one branch: 200 changes landed at 1 attempt each. Classic got ~100, at 27 tries each. It's a Gitaly fork.https://clovercoder.tech/p/manimgx/ https://clovercoder.tech/p/manimgx/Mon, 19 Oct 2026 14:50:16 -0000 Manim's Python API on a Rust GPU renderer, made for coding agents. 5 scenes of 1080p60 video rendered in 3.08s on an M4 Pro. ManimCE took 278s. The storyboard + layout check for agents is the part I'd use.https://clovercoder.tech/p/ts-rust/ https://clovercoder.tech/p/ts-rust/Sun, 18 Oct 2026 22:28:11 -0000 An experimental Rust port of the TypeScript 7 compiler, built by AI agents. All 184,773 ported Go tests pass, ~2x faster than tsc 7. Author says they never read a line of the code. Would you put it in CI?https://clovercoder.tech/p/esp32-iphone-agent/ https://clovercoder.tech/p/esp32-iphone-agent/Sun, 18 Oct 2026 19:23:39 -0000 A $7 ESP32 lets a vision model drive a real iPhone. No jailbreak, no dev account. AirPlay mirroring is the eyes, USB mouse + AssistiveTouch are the hands. Catch: 5 fps mirroring, ASCII-only typing. Clever hack though.https://clovercoder.tech/p/jeff/ https://clovercoder.tech/p/jeff/Sat, 17 Oct 2026 19:51:19 -0000 A 0.8B model whose only job is picking from a list. Jeff scores every option in one forward pass, ~28 ms on an M4 Max. 15 LoRA adapters on one base, including a prompt-injection guard. Tiny models with one job are underrated.https://clovercoder.tech/p/coucou/ https://clovercoder.tech/p/coucou/Sat, 17 Oct 2026 16:14:53 -0000 Your Claude Code, Codex and Cursor sessions, living in the Mac notch. coucou shows live edits and diffs, and you can answer permission prompts right from there. Nice touch: if the app isn't running, the hook exits instantly. Your agent never waits.https://clovercoder.tech/p/jevgrep/ https://clovercoder.tech/p/jevgrep/Sat, 17 Oct 2026 04:15:03 -0000 Stop letting your coding agent grep around blind. jevgrep: ask "how are telemetry events sent?" and get back the files, exact excerpts and call sites. Author's small SWE-bench run: same 8/10 solved, ~26% cheaper.