A Claude Code proxy stops cache rebuilds that cost ~22% of your bill
Claude Code's 5-minute cache TTL bills you on long subagent runs; a wave of weekend tools now targets it and the 5-hour window, plus an MCP video editor.
Copy markdown
Warm the cache, skip the 1.25× rebuild
claude-thermos is a local reverse proxy you launch with `uvx claude-thermos` in place of `claude`: when your main agent sits idle past the 5-minute prompt-cache TTL waiting on a long subagent, it replays the last request at max_tokens 1 to keep the cached prefix hot. The repo pegs the rebuilds it prevents at ~22% of the measured token bill — a 0.1× cache read instead of a full 1.25× re-encode.
A shelf of 5-hour-window hacks
Different pain, same weekend: claude-warmup fires a tiny scheduled Haiku message each morning so Claude Code's 5-hour usage window anchors to a convenient reset instead of your first real prompt — you don't get more tokens, just a better reset time. llm-primer keeps pre-warmed sessions waiting in tmux, and claude-code-session-bot warms sessions over Telegram while you're away.
Palmier Pro: a video editor your agent drives
This open-source (GPLv3) Swift editor for macOS exposes an MCP server on 127.0.0.1:19789, so Claude Code, Claude Desktop, Codex or Cursor can edit the timeline and drop in Seedance or Kling clips directly. The editor core is free and needs macOS 26 (Tahoe) on Apple Silicon — enough to wire up an agent that cuts a rough edit from a folder of footage.
'Harness engineering is not enough'
A front-page HN essay, 'Why Software Factories Fail,' argues the harness around coding agents — tests, context, guardrails — isn't sufficient on its own, pushing back on Martin Fowler's harness-engineering framing. A useful gut-check before you pour a weekend into agent plumbing instead of the codebase it runs against.