CEREBRO

Coding agents & orchestration

Codex 0.146.0: named/pinned sessions via /new and /clear, switch side conversations without closing them, big agent-session ergonomics bump.
Orca: ADE running Codex/Claude Code/OpenCode/Pi side-by-side in separate worktrees with one unified usage tracker, aimed at multi-agent "fleet" workflows.
/mission for Claude Code: spawns a team of agents from a single mission brief, another entrant in the multi-agent-orchestration-on-top-of-Claude-Code space.
Task Monki: runs coding agents through a full dev-process pipeline (plan → code → review), packaging agent orchestration as SaaS.
ADHD skill: tree-of-thought with pruning for Claude/Codex Agent SDK, fans out divergent reasoning frames in parallel then scores and prunes dead ends before deepening survivors.
affaan-m/ECC: performance-optimization harness for agent skills/instincts/memory/security across Claude Code, Codex, Opencode, Cursor.

Plan mode fix: user report on Plan mode solving premature-action frustration with coding agents, useful anecdotal signal on workflow discipline.
Best CLAUDE.md line: community consensus forming that negative constraints ("don't do X") outperform positive instructions in agent system prompts.
Undo context compression?: thread on recovering pre-compaction context, relevant to anyone hitting Claude Code's auto-summarization mid-task.
MCP is useless, just use a CLI?: recurring debate on whether MCP servers beat plain CLI tool-calling for agent tool use, worth tracking as a design-pattern signal.
MemoryCustodian: repo-native persistent memory for coding agents, competing space to Claude Code's own memory/CLAUDE.md conventions.
mattpocock/skills: small composable model-agnostic agent skills built for real engineering, explicit pushback against process-owning frameworks like GSD/BMAD/Spec-Kit.

GLM-5.2 step change: open-weight model reported crossing a capability threshold for agentic tasks, plus a new paper on open RL recipes for terminal agents.
Codex 0 to 10M users: OpenAI's Akshay Nathan on Codex MAU up >10x since Jan 2026, framing "people who use code but can't write it" as the biggest agentic-interface prize.
Frontier post-training recipe review: Finbarr Timbers podcast on what an Olmo-style recipe needs to reach frontier post-training quality.
Scientific computing in the agentic AI age: field report on coding agents accelerating genomics and scientific software modernization.

book-to-skill: converts technical book PDFs/doc folders into a queryable agent skill for Claude Code/Copilot CLI/Amp.
Denovo: business-manager layer to turn vibe-coded apps into paying-customer products, signal on the vibe-coding-to-monetization pipeline maturing.

SnapState: persistent state layer for AI agent workflows, HN-posted infra play for agent memory/checkpointing.
Hubble: open-source Markdown/HTML notetaking app pitched for humans and agents; HN pushback that a synced .md directory already solves this.
claude-video: /watch skill gives Claude video comprehension by downloading, frame-extracting, and transcribing before handoff to the model.

Homebrew 6.0.0: new tap-trust security mechanism, faster internal JSON API, Linux sandboxing; HN notes the forced-upgrade model is pushing users toward Mise for version pinning.
npm-scan: behavioral supply-chain scanner for npm catching eBPF rootkits, credential stealers, and AI-targeted attacks that audit/Snyk/Socket miss.
JusTTY: native macOS terminal built on Swift + libghostty.
ACM digital library access for LLMs: opinion piece proposing LLM access to ACM's corpus; HN pushback centers on ACM monetizing author content without consent or revenue share.