Ponytail is a plugin that forces coding agents to reuse repo code, stdlib, and native APIs before writing new lines. It took #3 on Product Hunt on Sept 5 and installs on Claude Code, Codex, Cursor, and 14+ other agents.
Announced Oct 21 as a research preview with Microsoft Research and built into GitHub Next's Chopin planner, PDD Kit turns your plan into a Verus spec for a human to sign off, then has the Verus verifier prove the agent's Rust code satisfies it for every input, falling back to tests and computer use agents where a proof isn't possible. It's Rust only for now, with a live demo at GitHub Universe on Oct 28–29 (Automater recap)…
Launched Oct 21 (YC S25), Alter sits between shared internal agents and downstream APIs, keeping credentials in its vault and evaluating default deny, per parameter policies against the end user each request is made for, with approval for risky calls and audit logs. In proxy mode the token never reaches your agent process; the team also published an OpenClaw plugin (docs)…
Released Oct 21 as MIT source, LayerFS mounts a full workspace per agent without copying the base, commits each step into one SQLite store, and lets agents branch, rewind, or merge, refusing same file clashes unless a .lww rule says the newest edit wins. For now it runs through the dsh layerfs plugin for DeepSeek Harness, and the README warns the store has no crash or power loss durability (docs)…
Released Oct 21, GitLab 19.5.1, 19.4.3, 19.3.6, 19.2.9 and backports down to 18.11.13 fix CVE 2026 110264, an integer overflow when compiling a crafted regular expression that could let an unauthenticated attacker run code on the server. It follows the Oct 12 CI/CD regex fix (CVE 2026 97229); GitLab.com is already patched, so self managed instances, including ones your agents or runners connect to, should upgrade now…
Released Oct 21, corridor argos 1 is a Qwen3.5 9B fine tune that reads a coding agent's plan and returns the project specific security requirements that apply, such as which auth middleware or upload helper to use, feeding Corridor's Agent Steering. On Corridor's own SecReT benchmark it scored 0.655 micro F1, the best of 14 models tested, with the lowest mean time to first token at 740 ms…
Spotted something we missed? Start a thread.
