v2026.09.21
karpathy/nn-zero-to-herov2026.09.21Sep 20, 2026by knu
Upgrading?
AI Summary
This release improves background job recovery with a new delayed_job_master process and watchdog, adds configurable timeouts, and updates the Ruby and Docker runtimes.
Key Highlights
- Optional `delayed_job_master` process for single-threaded workers.
- Watchdog for master-managed workers to handle timeouts.
- Configurable execution lock, network open, and I/O timeouts.
- Ruby upgraded to 4.0.7 and Docker base to Ubuntu Noble.
- Proxy routing support for individual Agents.
New Features
- Background job watchdog for master-managed processes.
- Configurable `AGENT_EXECUTION_LOCK_TIMEOUT`.
- Configurable outbound network timeouts.
- Proxy routing via `AGENT_PROXY` and `use_agent_proxy`.
Full Release Notes
# v2026.09.21 This release improves recovery from stalled background jobs, adds opt-in proxy routing for individual Agents, and updates the Ruby and Docker runtimes. ## Background jobs - Add an optional `delayed_job_master` process that forks single-threaded workers on demand. Set `DELAYED_JOB_WORKERS` to control concurrency and `DELAYED_JOB_MAX_MEMORY` to recycle workers after jobs that exceed the memory limit. The existing threaded runner remains the default; see the separate-process example in [Procfile](https://github.com/huginn/huginn/blob/v2026.09.21/Procfile). - Add a watchdog for master-managed workers. It monitors the full reservation, execution and cleanup cycle independently of the database. When `DELAYED_JOB_MAX_RUNTIME` is exceeded, it requests cancellation and allows `DELAYED_JOB_SHUTDOWN_GRACE` seconds (default: 10) for cleanup before forcibly terminating the worker. Abandoned jobs are recovered through the normal failure and retry policy. This watchdog applies to master-managed workers, not the threaded runner or directly executed scheduler work. [#3755](https://github.com/huginn/huginn/pull/3755) - Limit Agent execution-lock acquisition to 30 seconds by default, configurable with `AGENT_EXECUTION_LOCK_TIMEOUT`. Background jobs that cannot acquire the lock fail and can be retried. [#3745](https://github.com/huginn/huginn/pull/3745) - Apply configurable outbound connection and I/O timeouts through `OUTBOUND_NETWORK_OPEN_TIMEOUT` and `OUTBOUND_NETWORK_TIMEOUT`, capped below the background job runtime limit. [#3745](https://github.com/huginn/huginn/pull/3745) - Fix "Run now" for failed jobs by clearing exhausted attempts and stale reservation state before requeuing. [#3755](https://github.com/huginn/huginn/pull/3755) ## Proxy routing Trusted-user instances can set `AGENT_PROXY` and enable `use_agent_proxy` on selected Agents, including JavaScriptAgent's `Agent.fetch` and `Agent.fetchAll`. The Docker images support `START_SMOKESCREEN=true` to start the bundled filtering proxy without forcing all requests through it. Existing `ENABLE_SMOKESCREEN` and `OUTBOUND_PROXY` enforcement takes precedence and is unchanged. See [outbound request configuration](https://github.com/huginn/huginn/blob/v2026.09.21/doc/manual/outbound-requests.md). ## Runtime and Docker fixes - Upgrade Ruby to 4.0.7 and the Docker base to Ubuntu Noble, fixing a `mini_racer` load error during asset precompilation. [#3753](https://github.com/huginn/huginn/pull/3753) - Make single-process Docker workers wait for pending migrations before starting. [#3746](https://github.com/huginn/huginn/issues/3746) - Fix FtpsiteAgent initialization on Ruby 4 when FTP timeout options were mistaken for the host. - Refresh dependencies and keep `json` below 3.0 pending dependency compatibility. Docker images are available for amd64 and arm64 on GHCR and Docker Hub, tagged `v2026.09.21`. **Full Changelog**: https://github.com/huginn/huginn/compare/v2026.09.09...v2026.09.21
Related Repositories
rasbt/LLMs-from-scratch★ 107,072
Implement a ChatGPT-like LLM in PyTorch from scratch, step by step
lyogavin/airllm★ 40,206
AirLLM 2.8T LLM inference with single 4GB GPU
NirDiamant/RAG_Techniques★ 29,960
This repository showcases various advanced techniques for Retrieval-Augmented Generation (RAG) systems. Each technique has a detailed notebook tutorial.
microsoft/OmniParser★ 25,539
A simple screen parsing tool towards pure vision based GUI agent
NirDiamant/GenAI_Agents★ 24,658
50+ tutorials and implementations for Generative AI Agent techniques, from basic conversational bots to complex multi-agent systems.