CEREBRO

Coding agents & harnesses

OKF Agent Memory: vendor-neutral, git-native persistent memory spec (OKF v0.2) so agent context survives across sessions instead of resetting with the window.
everything-claude-code: hackathon-winner's full Claude Code config dump, agents/skills/hooks/commands built over 10+ months of daily use, raw material for stealing patterns.
ECC: harness performance-optimization layer (skills, instincts, memory, security) spanning Claude Code, Codex, Opencode, Cursor, security warning to install only from official channels.
ruflo: execution-layer harness bolted onto Claude Code/Codex adding 100+ specialized agents, swarm coordination, and self-learning memory, "agent = model + harness" framing worth stealing.
Git-native memory for Cursor: pure-Go zero-dep MCP tool replacing bloated .cursorrules files with persistent git-backed memory.
humanlayer/skills: includes an improve-claude-md skill that rewrites CLAUDE.md using <important if> blocks for better instruction adherence, directly relevant to your own CLAUDE.md tuning.
shot-scraper video: new command records Playwright-driven video demos from a storyboard.yml, useful for agents proving their own UI work.
8-month agent harness, abandoned: cautionary war-story on over-engineering a custom agent harness, worth a skim for what not to build.

Claude Fable 5 and AI safety fables: Anthropic's Fable 5 release notes plus a walkback on silent classifier-based manipulation of AI-research queries.
GPT-6 Astra on robot arms: head-to-head against Fable 5/5.1 on identical YAM-arm manipulation tasks, community notes $2/block cost still beats labor economics.
GPT-6 Astra via BYOK in Cursor: workaround thread for getting a non-native model's tool-calling and reasoning working inside Cursor's agent.
Per-subagent model configuration in Cursor: request/workaround thread for pinning specific models to specific modes/subagents, same routing problem your cost-aware-pipeline setup already solves.
Stop heredoc, save tokens on Windows: CLAUDE.md snippet to stop Claude burning tokens/failing on heredoc syntax under Windows shells.

Hyperprobe: lets agents debug production without redeploying, live-inspection tool for agent-driven ops.
Reflexio: behavioral-learning layer meant to make agents improve over time from past runs.
Local-first skill router: sub-20ms, zero-token, CPU-only router for picking which Agent Skill to invoke, avoids burning a model call just to route.
Remote MCP servers fail silently in Cursor: diagnosis + monitoring approach for silent MCP connection failures, relevant if you run remote MCP servers.
Cloud Remote Execution vs Cloud Agents: clarifies Cursor's two distinct cloud-execution offerings, easy to conflate.
dif.sh: markdown-defined feature flags that your coding agent installs itself.
GitWarren: review tool for having coding agents review code before you commit.

Export Claude chats to text/md: community thread on extracting Claude conversation history to plain files.
Fix for Claude Desktop hiding history on account switch: tool built to recover chat history that Claude Desktop hides when switching accounts.
Ponytail: tool positioning "new code as last resort," reuse-first code generation, notably same name as your ponytail-mode persona.

mattpocock/skills: opinionated small/composable agent skills for real engineering, explicit pushback against heavyweight process frameworks (GSD/BMAD/Spec-Kit) that take away control.

Claude Code subagents managing an Outlook inbox: thread probing whether anyone's using subagents for inbox triage, adjacent to your gog-inbox-triage setup.