Open Source Radar — September 14: agents get a QA department
Today's trending board is mostly about control, not capability: a code reviewer that Alibaba ran internally for two years, a verifier that refuses to let a model assert facts, a memory store for agents that forget on session close, a scanned skill registry, and a browser agent that hands you a spreadsheet instead of an essay.
OpenCodeReview (Go, 24,147 stars) — Alibaba's internal AI code review assistant, open-sourced after the company says two years of use across tens of thousands of its own developers surfaced millions of defects. It reads a Git diff, then lets an agent with tool access read whole files, search the repo and check other changed files for context, so feedback lands on specific lines rather than on the diff's surface. Where there's no useful diff — auditing an unfamiliar codebase — it reviews entire files instead. The design point is honesty about what general-purpose agents do badly on review: they quietly skip files in a large changeset, cite the wrong line numbers, and swing wildly with prompt tweaks, so the steps that must not go wrong run as fixed engineering logic and the model only handles judgment. The project claims markedly better precision than Claude Code on the same underlying model while using about one-ninth of the tokens, and it publishes its own review benchmark built from 200 real pull requests across 50 repositories and 10 languages, annotated by more than 80 senior engineers — the project's own numbers, not third-party. Apache-2.0, works with any OpenAI- or Anthropic-compatible endpoint. Use it if you've been running code review through a chat agent and don't trust the coverage.
Reverify (Python, 1,197 stars) — A verifier that makes a deterministic tool the judge of everything a model claims. The model proposes a fact about a binary — a struct layout, a function prologue, what a routine computes — and the answer comes back verified, refuted, or inconclusive with the bytes it actually checked, because claims are resolved by disassembly, pattern matching and CPU emulation rather than by the model's confidence. Reverse engineering is where hallucination is worst, so that's where its measurements come from: on 71 real Windows system files, the author reports the model's textbook answer was wrong 97% of the time and that the gate never accepted a single incorrect claim — its own benchmark, run on every code push, with the checks cross-graded against Capstone, LIEF and Unicorn rather than against its own tests. The same discipline applies to ordinary source: hand it a reference implementation and a candidate rewrite and it runs both over shared inputs, so a refactor is tested rather than trusted, and a refutation arrives with the input and both outputs. It runs as an MCP server inside Claude Code, Cursor and friends, and keeps a per-file ledger of verified and refuted claims on disk, so a session that hits its context limit resumes from what was checked instead of from what the model remembered. MIT-licensed, and scoped to authorized work — software you own, malware you're contracted to analyze, CTFs.
Agent Memory (Python, 1,416 stars) — A long-term memory runtime for agents that treats Markdown files as the only source of truth and the database index beside them as a cache you can delete at any time. Anything that can run a shell command shares the store; agents speaking MCP get the same calls through a server. Retrieval is local and ranked and answers with file paths rather than pasted text, so the agent opens a memory only as deep as the task needs. Two design choices are worth stealing: writes fire at conversation boundaries instead of waiting for the model to remember to save something, and a consolidation pass runs on its own clock outside the session — pruning by value, filing anything destructive as a proposal you approve rather than deleting unattended. No API keys, no embeddings vendor, nothing in the cloud. MIT-licensed, version 0.1, and the most direct open answer to the "my agent forgets everything when the window closes" complaint.
Paw Work (JavaScript, 2,597 stars) — A Chrome extension built around one rule: the output should be a file you can edit, not a paragraph about the page. You select images, tables, text or links on the site already in front of you, describe the outcome you want in a side panel, and get back a spreadsheet, deck, poster, page or long document built on live canvases. Deliberately bounded — it says up front that it won't roam the web, fill out checkouts or pay for things, and that it inspects the page on demand and will ask once when the target format is genuinely unclear. Bring your own model keys; there's no account and no server behind it, and the whole thing runs sandboxed inside the browser. Built on Univer and tldraw, which means the design and slides surfaces carry tldraw's watermark until you license it — one of the few open agent tools that's honest about where the commercial dependency sits. MIT for the code, roughly 2,600 stars since a late-August start.
Agent Skills (TypeScript, 5,863 stars) — A curated, scanned registry of agent skills, pitched against an ecosystem where Snyk's own review of marketplace skills found critical issues in 13.4% of them. The hardening is structural rather than aspirational: no binaries in the library, static analysis in the release pipeline, every skill scanned with Snyk Agent Scan before publication, and installed sets pinned by lockfile and content hash so an update can't silently change what your agent loads. Skills themselves are plain packages of instructions plus templates and reference docs, addressed by category so names don't collide, installed to whichever agent you use — Claude Code, Cursor, Copilot, Codex, Antigravity, Windsurf and a dozen more — and an MCP server exposes the catalog to the agents directly, search first, fetch only what's needed. Maintainer-written skills are CC-BY-4.0 with third-party skills keeping their own licenses. If you've been dropping unreviewed skills into an agent config, this is the version of that habit with a supply chain behind it.
Worth watching this week: whether skill registries consolidate around scanned, hash-pinned installs the way package managers did — and whether code-review agents get judged on recall, not on how confident their comments sound.
Which agent output do you actually trust without re-checking it? Tell us in the comments.
Sources: Open Code Review (GitHub) · Reverify (GitHub) · Agent Memory (GitHub) · Paw Work (GitHub) · Agent Skills (GitHub) · Snyk Agent Scan skills report
