• Sources: primary, discussion
  • Summary: Caleb's post says over three quarters of 244 Claude Code releases this year changed the system prompt or tool descriptions, that the default Bash tool description is almost 3,000 tokens, and that IDE text selection works for only one connected session. He moved to Pi for its stability and extension model and wrote an extension for multi-tab selection in about 20 minutes.
  • Why it matters: Harness choice now affects token cost, prompt stability and which model subscriptions you can use, so teams standardizing on an agent should weigh those terms alongside model quality.
  • Background: A coding harness is the program around a model that supplies the system prompt, tools, permissions, and context management. Pi is a minimal open source harness that users extend with their own code, while Claude Code ships those features built in and changes them frequently.
  • Comments: The thread reached 177 points and 98 comments. Commenters said Anthropic subscriptions are restricted to its own harness while OpenAI subscriptions work in Pi, and quoted Anthropic support saying Max and Team plans include monthly API credits since 2026-10-07. Others noted Pi has no built-in permission classifier, so users add extensions such as pi-permission-system or run agents in their own sandbox, and one user reported much higher cost through Pi than Claude Code for similar work, pointing to prompt caching.