2026-05-10

• #0005

  #0005Agents & toolsGitHub Trending Weekly3 weeks ago

  `Ouroboros`: spec-first Agent OS for replayable AI coding

  60radar

  OuroborosAgent OS — spec-first AI coding workflow

  Adds a local-first runtime that turns agent coding into a replayable, policy-bound workflow. Worth testing if you want less prompt thrash and more verifiable runs across multiple coding CLIs.

  • —Uses a fixed loop of interview -> crystallize -> execute -> evaluate -> evolve, pushing requirements capture ahead of code generation.
  • —Targets multiple runtimes including Codex CLI, Claude Code, OpenCode, Hermes, Copilot CLI, and Kiro, reducing lock-in.
  • —The pitch is a 3-stage automated evaluation gate plus an immutable seed spec, so architecture drift and manual QA get constrained early.
  • —Setup is oriented around local tooling and MCP registration, with ooo interview "..." as the entry point instead of freeform prompting.

  Source: github.com/Q00/ouroborosRead original →
• #0004

  #0004Agents & toolsr/ClaudeAI3 weeks ago

  `Autoharness` lets `Claude Code` tune its own agent harness

  50radar

  AutoharnessHarness tuner — validates and keeps only score gains

  A layer above agent setup is emerging: tools that rewrite prompts, scoring, and runtime context, then keep only eval-proven wins. Concrete benchmark lifts make this more than workflow bragging, and the repo is worth skimming if you run custom agents.

  • —On tau2-airline, it reports +40.7% from best-of-N skillbook scoring with an LLM judge; evaluator design is moving into the product surface.
  • —Another +24.1% came from reflector hyperparameter changes like temperature and max subagent calls, so simple harness knobs still have large headroom.
  • —Injecting runtime context on every step added +22.2%; step budget, recent tool calls, and recent results materially changed outcomes.
  • —Flow is minimal: install, point Claude Code at GUIDE.md, propose harness edits, run evals, and keep only score-improving changes.
  • —Because gains are benchmark-specific, the useful takeaway is the loop: mutate harness, eval immediately, and ship only measured improvements.

  Source: www.reddit.com/r/ClaudeAI/comments/1t8cn9y/claude_improvRead original →
• #0003

  #0003Agents & toolsr/ClaudeAI3 weeks ago

  One Escaping Bug Turned `rd /S /Q \` Into a 717 GB Windows Wipe

  40radar

  A cross-shell quoting failure turned a folder delete into C:\ root deletion in under two minutes. Useful reminder: never pipe destructive ops through multiple parsers when a native command plus -WhatIf exists.

  • —The intended target was C:\Users\ADMIN\Desktop\WIP, but cmd /c plus zsh/tmux/PowerShell collapsing changed it to rd /S /Q \.
  • —Damage reached about 717 GB before interruption; only a separate HDD backup prevented real loss, which makes backup isolation the real safety net.
  • —Using PowerShell Remove-Item -Path 'C:\absolute\path' -Recurse -Force -WhatIf would have exposed the broken path before execution.
  • —The practical lesson is operational, not model-specific: echo the fully expanded command and avoid routing deletes through cmd when PowerShell can run them directly.

  Source: www.reddit.com/r/ClaudeAI/comments/1t923er/i_deleted_a_gRead original →
• #0002

  #0002Infra & SaaSGeekNews3 weeks ago

  `AWS` Virginia Outage Resolved After `US-East-1` Overheating Incident

  40radar

  US-East-1AWS region — its primary Northern Virginia region

  A cooling-related failure in `US-East-1` hit major customer-facing services, not just internal AWS dashboards. If anything still depends on a single region, this is another reminder that multi-region failover is cheaper than surprise downtime.

  • —The trigger was data center overheating in US-East-1, turning a routine status alert into a real regional availability event.
  • —Named impact reached services like Coinbase and FanDuel, which shows the blast radius reached end-user transactions, not background jobs only.
  • —AWS marked recovery at 3:29 PM ET Friday. Even a resolved incident matters if your stack, auth, or webhooks are pinned to one region.

  Source: news.hada.io/topic?id=29337Read original →
• #0001

  #0001OtherGeekNews3 weeks ago

  `NixOS` Secrets: Plaintext in `Nix store` Is Readable

  40radar

  NixOSLinux distro — declarative, reproducible system config

  Putting secrets in Nix config, private Git, or git-crypt plaintext still leaks them through Nix store. sops-nix is the practical baseline if you deploy with NixOS; otherwise your server access model is already broken.

  • —Secrets embedded in Nix config can end up in Nix store, which is world-readable on the machine; repo privacy does not help after deploy.
  • —Keeping values in a private repo or decrypting with git-crypt before build still leaves plaintext artifacts in the store, so the leak point moves to runtime.
  • —sops-nix uses .sops.yaml rules and encrypted-file editing flow, giving a cleaner path than ad hoc secret injection for NixOS setups.

  Source: news.hada.io/topic?id=29328Read original →