v2.1.199

Source: Anthropic Claude Code Date: 2026-07-02 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.199

Summary

Twenty-three fixes, the load-bearing cluster all about the harness no longer trusting a background agent’s self-report: subagents that hit a usage limit now surface the error to the parent instead of reporting success; mid-stream server errors keep the partial response instead of discarding it; rate-limited children return partial work rather than silently failing; claude stop is now honored even when it races a background-agent respawn. Supporting fixes cover a Linux daemon that killed itself every ~50s after a corrupted worker record, SendMessage misrouting on reused agent names, hooks silently hiding stderr on exit code 2, and a retry watchdog raised to 300 for non-capacity transient errors.

Implications

Continues the background-agent hardening arc (196→197→198→199). Where v2.1.198 removed asks wholesale and kept one boundary — an agent’s message is never the user’s approval (authority doesn’t delegate) — 199 draws the next one: an agent’s success is never proof of success (verification doesn’t delegate). Feeds the agent autonomy thread: once you stop asking, every place the autonomous layer can quietly be wrong becomes a correctness hole, and the fixes are the bill for the autonomy 198 granted. 200 ≠ success shipped as a bug fix. Reliability work is lagging the autonomy, not leading it.