Open Source Radar — November 4: DeepSeek Harness 0.2.2
Today's board is plumbing, not models: the harness an agent runs inside, the memory it keeps between sessions, the skills it loads and the scraper that feeds it. All four pushed code in the last three days.
DeepSeek Harness (TypeScript, 271,877 stars, MIT) shipped v0.2.2-alpha.3 on November 3, and the update is almost entirely about the part of coding agents nobody demos: review. There is now an experimental change-request-and-worktree view that browses local code and PR diffs in a searchable file tree, lets you select a snippet, attach a comment and hand it back to the agent, then tracks the review state. A second tool, apply_patch, does multi-file add, edit, delete and move in one call, off by default, with the result visible in that same review pane. Automation tasks moved up to a host-level manager, so a scheduled job is no longer tied to the session that created it. Chinese tech outlet 智东西 Zhidx reported the release as the harness quietly filling in the last gaps of a code-review loop. It is still an alpha, and the release notes say anyone running the old standalone git-worktree plugin has to remove it by hand before upgrading.
claude-mem (TypeScript, 112,355 stars, Apache-2.0) treats context as something to compress rather than something to re-read. Lifecycle hooks on session start, prompt submit, tool use and stop capture what the agent did, an AI pass turns that into observations and summaries, and the result lands in SQLite with a Chroma vector index behind it — exposed as four MCP search tools that return a compact index first, so a lookup costs roughly 50 to 100 tokens per hit instead of dragging transcripts back into the window. It now installs against Codex, Copilot, Gemini CLI and, as of v15.9.0 on November 2, Devin. Read the changelog before you install: v15.8.3 fixed a setup flow that had been uploading the last 30 days of prompts — with session IDs, project names, folder paths and timestamps — to the vendor's hosted service without asking.
Skills for Real Engineers (Shell, 312,017 stars, MIT) is Matt Pocock's own working set, roughly thirty skills he uses daily and published on the argument that process-heavy frameworks like GSD, BMAD and Spec-Kit fix agent behaviour by taking the process away from you, which makes the bugs in that process impossible to debug. His are deliberately small and composable, and they install as a plugin into Claude Code, Codex, GitHub Copilot and Gemini CLI rather than living in one vendor's format. The ones worth stealing even if you never install it: a two-axis code review that scores a diff against repo standards and against the originating spec in parallel, a bug-diagnosis loop that has to go red on the bug before it is allowed to fix it, and a planning skill for work too big for one session. It picked up 1,164 stars today on a repo last pushed November 1.
Scrapling (Python, 95,580 stars, BSD-3) is the highest-velocity project on the board today at 1,658 stars, and it is the least glamorous: a scraping framework whose v0.5 release on October 17 rewrote fetching, parsing, spiders and the MCP server together. The useful parts for agent work are a stealth fetcher that handles Cloudflare's Turnstile challenge, spiders with pause-and-resume so a long crawl survives a restart, and a one-line conversion of any page into clean Markdown for retrieval. There is also an MCP server and an installable agent skill, so Claude or Cursor can scrape through it without hand-written selectors. Two caveats: v0.5 carries breaking changes, and robots.txt compliance is an opt-in flag rather than a default.
Worth watching this week: whether DeepSeek's review pane gets adopted by other harnesses, or stays a feature of one editor.
