A developer who adopted Claude Code the day it hit the Claude Pro plan has walked away from it, switching to Pi — an open-source coding harness. The defection is one data point, but the grievances read like a product-management case study in how incumbents lose their earliest adopters. The proximate causes are specific and reproducible. Claude Code's VSCode integration allows only one session to connect to the IDE at a time; a second tab cannot receive text-selection events. The bug has been open since at least December 2025 and was closed as a duplicate of an unrelated issue. The official GUI extension fares no better — slash commands broken, no custom status line, subagent outputs jumbled into a single chronological thread. These are not edge cases; they are core developer-experience failures. The deeper complaint is strategic. In 2026 alone, Claude Code shipped 244 releases. Over three-quarters of them modified the system prompt or tool descriptions — changes that alter model behavior in ways that can silently contradict user prompts and repo conventions. The author frames this as a product without a strong "no" function: engineers and PMs adding features for their own workflows without considering downstream breakage for existing users. The token overhead tells its own story. Claude Code's default Bash tool description runs nearly 3,000 tokens, mostly instructions for git operations and GitHub PRs that frontier models already know how to execute. At the scale the author estimates — one million daily active users, 100 million requests per day — trimming even 1,000 tokens from that description could save millions of dollars in annual inference costs. That the overhead persists suggests either organizational inertia or a deliberate choice to optimize for enterprise onboarding over power-user efficiency. Meanwhile, Anthropic announced AGENTS.md support the day after the author decided to leave — a feature that addresses one of the complaints (the CLAUDE.md file fragmentation) but arrives too late to reverse the accumulated frustration. The timing underscores a familiar pattern: incumbents respond to defection signals with point fixes rather than architectural rethinking. Pi's appeal is the inverse of Claude Code's trajectory: stability, minimalism, and extensibility. The author reports writing a multi-tab IDE integration extension in twenty minutes — the exact feature Claude Code has failed to ship for over nine months. Open-source harnesses benefit from the same hindsight advantage that every second mover enjoys, but they also benefit from a structural alignment: their users are their developers, so bloat gets pruned rather than accumulated. The broader signal here is that the LLM coding-harness market is entering its disaggregation phase. The first wave was about getting anything to work at all; the second wave rewards tools that do less, more reliably. Anthropic's challenge is that Claude Code's revenue model likely depends on token throughput — exactly the metric that power users want to minimize. That tension between vendor economics and user efficiency is the real story underneath one developer's migration blog post.
