$ tail -f topics/ai.log
Daily updates from the AI and machine-learning ecosystem - PyTorch, Hugging Face, the Anthropic and OpenAI SDKs, local LLMs, and the agent frameworks.
256 updates across 10 projects this week.
One calm review of what shipped across AI / ML - the commits, releases, and security advisories that matter. Every Monday, with security advisories same-day. Free, unsubscribe in one click.
We'll start you on the top AI / ML projects - refine anytime. · Read a sample issue →
OpenClaw
Coordinated fixes across OpenClaw's runtime and OCM land today: signal-to-commit races closed, macOS permissions survive updates, and docs builds no longer render every locale twice.
read --wire →
OpenClaw
OpenClaw now grants every agent pair session access by default, a default-on widening of visibility that operators must explicitly narrow.
read --wire →
OpenClaw
OCM managed Gateway upgrades no longer get killed by their own shutdown, and oversized QA uploads now reject without crashing the provider.
read --wire →
OpenClaw
This week's ocm fixes ensure that upgrade simulations and migrations report true candidate outcomes even when cleanup or legacy config repair fails.
read --wire →
OpenClaw
Crabbox now requires exact local ownership claims before any automatic cleanup can delete a VM, closing destructive gaps across Tart, Proxmox, and Incus.
read --wire →
OpenClaw
Large history upgrades and Doctor checks speed up, while permission matching is prepared once per filtering operation.
read --wire →
OpenClaw
Crabbox ships native Daytona snapshot checkpoints and forks, plus a lease-lock fix that stops active leases from blocking unrelated runs.
read --wire →
OpenClaw
OpenClaw's cron jobs now record real delivery outcomes, preventing false-success completions that have been silently hiding failed deliveries.
read --wire →
OpenClaw
Crabbox shipped replay-safe checkpoint forks with exact lease IDs plus a wave of claim-fencing fixes for machines, Morph instances, and Semaphore jobs.
read --wire →
OpenClaw
Crabbox hardened XCP-ng, Sprites, Tenki, and Namespace lifecycle operations against cross-account and cross-region VM deletion.
read --wire →
OpenClaw
OpenClaw now records the creator of person-created automations in activity, and CLI conversations survive non-expiry failures.
read --wire →
OpenClaw
Systemd-managed OpenClaw builds could size the V8 heap from host RAM, risking unrelated services; remote checkouts no longer offer editor handoff; the bundled OpenProse plugin is gone.
read --wire →
OpenClaw
OpenClaw's CLI runner now keeps partial assistant text when a Claude live turn is aborted, fixing dropped transcript content.
read --wire →
OpenClaw
Mantis Telegram Desktop proof agent is now programmable at its trust boundaries with declarative fault rules, scripted mock responses, and proxy-side request recording.
read --wire →
OpenClaw
OpenClaw fixed a concurrency bug where session updates could overwrite newer state, and Crabbox hardened cleanup across Vast and Cloudflare to stop orphaning billed instances.
read --wire →
OpenClaw
OpenClaw shipped three behavioral fixes: the New Session picker no longer stays on 'Models unavailable' after a missing catalog owner, ambient owner resolution now honors agents.defaults.systemAgent.agentId on multi-agent rosters, and Skill Workshop revisions stage atomically instead of tearing on interruption. [ref:1][ref:4][ref:5]
read --wire →
OpenClaw
OpenClaw's gateway now loads workspace icons without sharing secrets, while crabbox v0.44.0 adds credential-free provider discovery and a Telegram Desktop recorder.
read --wire →
OpenClaw
OpenClaw fixed a browser bug where screenshot results could name a message tool that policy had removed, risking AI hallucinations.
read --wire →
OpenClaw
OpenClaw consolidated plugin JSON body parsing onto a single canonical bounded reader with strict size, timeout, and cleanup semantics, closing an indefinite-hold vulnerability in unauthenticated Telegram Mini App requests.
read --wire →
OpenClaw
Two audit fixes ensure fresh openclaw installs stop flagging valid agent rosters and aborted chat steering is no longer reported as completed.
read --wire →
