🌏 中文版
Today's Highlight
No new framework today. Four unrelated projects — from GitHub, Alibaba, OpenAI, and DeepSeek — turn out to be answering the same question when you line them up: how narrow should an agent's write access be before you let it near production? Only DeepSeek's prototype runs the other direction, showing what it looks like to loosen the leash instead of tightening it.
Trending Repos
github/gh-aw ⭐ 5,636
GitHub · Go · MIT
- What it is: GitHub's own project for writing CI automation that can actually use judgment. You write frontmatter for events and permissions, plain-language instructions in the body, and
gh-awcompiles it into a real GitHub Actions workflow that Claude Code, Codex, Copilot, or Gemini executes the reasoning step for. - Why it matters: Hand-written CI rules can't handle tasks that need judgment — "how should this issue be triaged," "is this failure a known flaky test." gh-aw hands the reasoning to an agent, but deliberately splits "read" from "write": the agent job itself is read-only and sandboxed, and anything that actually writes to GitHub (opening an issue, posting a comment, opening a PR) has to go through a separate safe-outputs job with narrowly scoped permissions. It's one of the most carefully worked-out designs yet for "giving an agent write access without giving it the whole token." A good design doesn't mean zero implementation bugs, though — a critical-severity advisory (GHSA-8h78-hpm7-29gg) was patched just in early November: versions 0.83.3–0.85.3 wrote the raw stdout/stderr of a credential-bearing git command into the safe-outputs artifact, and did so before GitHub Actions' secret masking kicked in, so the token sat unmasked in a downloadable artifact. Fixed in 0.85.4.
- Stack: Go CLI (a
ghextension) + GitHub Actions + a swappable AI engine. - Adoption difficulty: Medium —
gh extension install github/gh-awis quick, but you need to actually understand the permission model before turning on writes.
alibaba/open-code-review ⭐ 51,279
GitHub · Go · Apache-2.0
- What it is: Alibaba open-sourced the code review tool it runs at its own internal scale. The architecture stacks a deterministic rule layer under an LLM agent layer: NPE, thread-safety, XSS, SQL injection and similar issues that static rules can catch get caught by rules first, and the LLM agent only handles the parts that need context a rule engine can't express. Both layers support OpenAI- and Anthropic-compatible backends.
- Why it matters: Most "AI code review" tools in this wave are LLM-only, which means they're prone to confident-sounding false positives with no deterministic backbone underneath. open-code-review inverts that — the rule engine is the spine, the LLM is the add-on — making it one of the few tools that openly designs around "which judgments should be a rule, and which should be a model call." It currently only posts line-level comments; it doesn't touch or merge code itself, so its write footprint is narrow.
- Stack: Go core + an npm-distributed CLI (
@alibaba-group/open-code-review), with an OpenSSF Best Practices badge. - Adoption difficulty: Low — an npm install and it runs; the README documents pairing it directly with Claude Code.
openai/codex-security ⭐ 13,503
GitHub · TypeScript · Apache-2.0
- What it is: OpenAI packaged its Codex security scanning into a standalone CLI and TypeScript SDK. It scans a repo, a path, or a single git diff, validates whether a finding is real, generates a patch, can draft a SECURITY.md, exports SARIF/JSON/CSV, and integrates with Linear, Jira, and GitHub.
- Why it matters: This turns "an agent that finds and fixes vulnerabilities" into a pipeline-ready product rather than a chat feature — a direct point of comparison with what Claude Code's
/security-reviewdoes. Both are a scan-validate-patch loop; codex-security's distinguishing choice is leaning hard into external integration (issue trackers, CI, report formats). The patches it generates still need a human to apply them — nothing merges automatically. The repo itself was only created this past July, so compared to the first two tools (both already battle-tested at production scale), it's still building its track record. - Stack: TypeScript CLI/SDK, distributed via npm (
@openai/codex-security), with a GitHub Action available. - Adoption difficulty: Low —
npm install -g,cs login,cs scan .and you're running.
deepseek-ai/dsh-superagent-prototype ⭐ 23,128
GitHub · TypeScript · MIT
- What it is: DeepSeek's own research prototype — a proactive coding agent bound to a single GitHub repo, built on "DeepSeek Harness." It listens to GitHub (and optionally Feishu/Lark); when someone on the allowlist requests work, it opens its own work item, operates inside its own session/directory, and reports back when done.
- Why it matters: It's rare for a major lab to publish, this directly, a research prototype aimed at "a more autonomous agent system" — the repo explicitly calls itself a prototype, has issues and the wiki disabled, and its own docs admit some described behavior "isn't implemented in this prototype." Unlike the first three, which are production tools being open-sourced, this one reads more like DeepSeek showing which direction of autonomy it's exploring next — not something you'd drop into a real workflow yet.
- Stack: TypeScript, DeepSeek Harness's bundle format.
- Adoption difficulty: High — this is effectively a prototype/bundle for DeepSeek Harness, not a general-purpose, install-and-go tool.
Today's four picks line up nicely on one chart: which tool keeps its write access narrowest, and which one is trading guardrails for autonomy.
A qualitative comparison based on the designs and current state described in this article, not a quantitative score.
Notable Releases
No major framework release today. LangGraph 1.2.22, CrewAI 1.15.41, Composio CLI 0.6.1, and llama_index 0.14.26 all shipped within the last 48 hours, but all are patch-level dependency bumps or small bugfixes. browser-use 1.0.5 added a self-evaluating "judge" feature, but it's not substantial enough to warrant its own section.
Today's Takeaway
I used to assume open-sourcing agent tooling was mostly framework and platform vendors pitching new ideas. Today, four very different companies open-sourced tools they actually run internally, at almost the same time, and the emphasis in all four wasn't "how capable is this agent" — it was "how narrow is the write access." That might mean agent reasoning capability is already good enough, and the real competition from here is who can tighten the automated write scope down to something people are willing to hand over. gh-aw's own credential-leak advisory, patched just last month, is a good reminder that "the design scopes permissions tightly" and "the implementation actually has no holes" are two different claims.