I've been using the session handoff skill with a hook that checks my session's context window each turn and sends a reminder to do the session handoff if I'm >= 120k tokens. This has worked well for me. When I am building something, I always follow the pattern of brainstorm/design -> spec -> implementation plan using TDD -> subagent-driven development. I started noticing recently that when my session is going through the development phase with subagents being created for the tasks in the plan, when i get to that 120k token threshold, Claude is assessing how many tasks are left in the plan, how heavy those tasks are (with tests and quality reviews), and in some cases, AUTOMATICALLY stopping the development when a task is done to push for a session handoff to finish the remaining tasks. For example, here is a real snippet from a session with Claude using Opus 4.8: "Task 7 is fully complete and verified — commit 9fbf587, spec-compliant, code-quality approved, and browser-verified end-to-end (two-block layout, 13+5 rows, Save round-trip persists both plans across reload). Scratch server stopped, working tree clean. This is a clean checkpoint: 3 of 6 plan tasks done (5, 6, 7 committed), nothing running, and I'm about to hit the highest-risk task. Context is at 163k+ and the remaining work includes Task 10..." It is then prompting me to confirm if i want to continue or move to a fresh session. Has anyone else had this experience? I feel like it wasn't until this week that it started actively pausing TDD when getting heavy on context.

Saw this on LinkedIn. Haven’t found a link to the underlying study but had read about this shadow market before. https://www.linkedin.com/posts/kirill-balakhonov_it-turns-out-the-gray-market-for-claudegpt-share-7475951927604834305-xeqp/?utm_source=share&utm_medium=member_ios&rcm=ACoAABLtdFQB8j-HbgxOwcE4PWtxNFWP3YH92ak

Is anyone here US based who works for or has consulted/contracted with a municipal government in the data space? I just accepted a new role and coming from private sector I’m trying to get some insight into public sector challenges and any guidance from people who have gone through this career path already. My mandate is broad across all data not just AI, but that is definitely something the city is pursuing.

I switched Claude Code over to GLM 5.2 and ran it all day. It's a 756 billion parameter open source model you can route straight into the Claude Code harness for about five times cheaper than Opus, and for most of my knowledge work it held up fine. In this one I show you what it can build, where it beats Opus and where it doesn't, and exactly how to set it up so you can switch between models per project. Here's the config I use. Drop this into your .claude/settings.local.json and swap in your own Z.ai key: "env": { "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic", "ANTHROPIC_AUTH_TOKEN": "your-z-ai-api-key-here", "ANTHROPIC_API_KEY": "", "API_TIMEOUT_MS": "3000000", "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.2", "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.2", "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-5.2", "ANTHROPIC_SMALL_FAST_MODEL": "glm-5.2", "CLAUDE_CODE_SUBAGENT_MODEL": "glm-5.2" }