OpenClaw
OpenClaw extracts reusable @openclaw/ai runtime package
OpenClaw extracted its provider transports into a standalone @openclaw/ai package, and acpx 0.12.0 ships Grok Build plus lossless queue shutdown.
read --wire →
$ tail -f topics/ai.log
AI / ML news, every morning.
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.
Your week in AI / ML, every Monday.
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 →
Recent wires
OpenClaw
Fix config-health migration warning, session drift hard invalidations, MCP stderr crash
Upgrades from legacy config-health to shared SQLite no longer surface a permanent migration warning, and sessions resume on prompt drift instead of dying.
read --wire →
OpenClaw
Ambient transcript watermark now scoped per session, DeepSeek V4 reasoning fixed on OpenRouter
Two production fixes ship: Telegram group transcripts survive session resets without losing older messages, and OpenRouter-routed DeepSeek V4 requests stop sending conflicting reasoning controls.
read --wire →
OpenClaw
Telegram webhook durability and reply-target RPC calls land
This week's patch batch hardens Telegram webhook updates against crashes, makes iMessage RPC replies fail loudly instead of falling back, and fixes vote casting on native Messages polls.
read --wire →
OpenClaw
ClawSweeper passes security fix for agent skill MCP boundary drift
OpenClaw merged a security fix that flags agent skill MCP boundary drift, and removed the direct skill scaffold helper to keep durable skill work inside workshop proposals.
read --wire →
OpenClaw
OpenClaw ships seven security fixes across HTTP, Browser, and Exec surfaces
Seven fixes over two repos harden core surfaces: browser URL checks, glob allowlists, admin scope, and slow-read DoS, with two maintenance commits in the clownfish repo.
read --wire →
OpenClaw
OpenClaw bounds provider response reads against memory exhaustion
Five fixes cap unbounded JSON responses across APNs, Runway, OpenAI, Together, PixVerse, and fal, closing a class of OOM and hang risks on media and push paths.
read --wire →
OpenClaw
OpenClaw doctor exposes plugin registry findings; fixes AbortError misclassification
OpenClaw's doctor command now surfaces structured plugin-registry findings with dry-run repair effects, while a model-fallback fix stops provider-side AbortErrors from being misread as user cancellations.
read --wire →
OpenClaw
Bounded SSE and Google Chat reads prevent OOM; Codex waits for native tool completion
OpenClaw patches memory-exhaustion vectors in SSE transport and Google Chat while fixing a race that caused Codex to send failed tool results for tools still running.
read --wire →
OpenClaw
Exec workdir checks fail early, Telegram reply loops capped
OpenClaw exec now rejects invalid explicit workdirs before running commands, and Telegram reply-session initialization is serialized to stop replay loops.
read --wire →
OpenClaw
HTTP agent traffic now emits model.usage diagnostics, auth failures surface in channels
The agent runtime now reports token and cost diagnostics for all HTTP ingress paths, and channel users get a real error when a provider rejects a request.
read --wire →
OpenClaw
Heartbeat hook now fires on harness prompts
OpenClaw fixed a gap where plugins using heartbeat_prompt_contribution lost context on heartbeat turns when running on harness runtimes, and workboard CLI cards got a cleanup.
read --wire →
OpenClaw
Gateway install warns on wrapper overwrite, status hints for missing-secret diagnostics
Gateway service installs and restarts now warn before rewriting customized macOS LaunchAgent env wrappers, and JSON daemon output surfaces those warnings.
read --wire →
OpenClaw
Telegram lane recovery fixes stuck inbound events
OpenClaw shipped fixes that unblock Telegram inbound events stuck behind live-held claims and image-heavy sessions that silently failed to compact.
read --wire →
OpenClaw
Crabbox hardens secret redaction, redirect guards, and artifact validation
Crabbox shipped five fixes that close security and correctness gaps across providers, runners, and artifact handling.
read --wire →
OpenClaw
Crabbox revokes Code viewer sessions on logout, adds Scaleway SSH leases
Crabbox now revokes Code viewer sessions server-side on GitHub portal logout and adds a direct Scaleway SSH-lease provider, with Cloudflare Access key fetches bounded behind admin auth.
read --wire →
OpenClaw
CLI typo suggestions, iOS notification fix, axios internals cleanup
OpenClaw improves CLI error handling with close-command suggestions, fixes iOS notification settings behavior, and stops poking at axios private internals in the Feishu extension.
read --wire →
OpenClaw
OpenClaw prunes stale agent APIs, Clownfish tightens security review routing
OpenClaw dropped five stale agent APIs while Clownfish hardened its security review pipeline against false positives and Clownfish docs replaced a vulnerable parser.
read --wire →
OpenClaw
Docs translation CI throttled, queue fixes prioritize fresh canaries
Docs translation CI now stays under GitHub's job limits while queue and publish fixes tighten behavior on fresh artifacts.
read --wire →
OpenClaw
Authenticated adapter ingress, release hardening across OpenClaw
Crabbox shipped an authenticated reverse proxy for loopback services and ClawSweeper cut 0.3.0, while OpenClaw fixed a desktop quota pill regression and a 700ms debounce that was hiding Discord tool emojis.
read --wire →