Claude Code 2.1.133: Parallel Session Reliability — No More 401 Crashes, Effort Bleed, or Missing Subagent Skills
Claude CodeClaude Code 2.1.133 delivers a focused set of multi-session reliability fixes for developers running parallel Claude Code instances. A race condition that caused all parallel sessions to crash with 401 errors after a token refresh has been resolved. The /effort command no longer bleeds its setting across concurrent sessions — changing effort in one terminal stays isolated to that session. Additionally, subagents can now properly discover project, user, and plugin skills via the Skill tool, and memory pressure management has been improved so warm-spare background workers are released under low-memory conditions.
Sources & Mentions
3 external resources covering this update
Stability Improvements for Power Users Running Multiple Sessions
Claude Code 2.1.133 lands a cluster of fixes that directly affect developers who run multiple concurrent Claude Code sessions, use multi-agent workflows, or rely on Remote Control from claude.ai. While individually these look like bug fixes, together they represent a meaningful reliability upgrade for the heaviest Claude Code users.
401 Crash After Token Refresh — Fixed
The most disruptive fix in this release addresses a race condition that caused all parallel Claude Code sessions to dead-end at a 401 Unauthorized error following a token refresh. The scenario: when one session triggered a refresh of the shared OAuth refresh token, a concurrent credential write in another session could clear the new token before it was used. Every other session relying on the shared credentials would then receive a 401 and have no path forward except a manual re-login.
This race has been eliminated. Parallel sessions can now refresh tokens concurrently without corrupting each other's credentials.
Effort Level No Longer Leaks Between Sessions
Claude Code's /effort command — which controls how much thinking budget the model applies to a task — previously had a cross-session contamination bug. Setting /effort to high in one terminal could unexpectedly change the effort level in a separate concurrent session. Additionally, an IDE effort change could be silently dropped if it arrived at the wrong moment.
Both issues are fixed. Each session's effort level is now fully isolated. A related issue where the CLAUDE_CODE_EFFORT_LEVEL environment variable override wasn't reflected in the /effort picker (a separate fix in v2.1.132) compounds the improvement: effort configuration is now consistent and predictable across the board.
Subagents Can Now Find Skills
Multi-agent workflows relying on the Skill tool hit a regression where subagents couldn't discover project-level, user-level, or plugin-level skills. This made delegated tasks significantly less capable — a subagent that can't find the available skills can't invoke them. The fix restores full skill discoverability for all subagent types.
Memory Pressure Management
Claude Code now proactively releases warm-spare background workers when memory pressure rises. Long-running sessions with multiple active tools and subagents could previously accumulate significant memory overhead from pre-warmed workers that were kept alive speculatively. Under memory pressure, these workers are now released, reducing the chance of out-of-memory conditions on developer machines during extended coding sessions.
Remote Control Cancellation Fixed
Users managing Claude Code remotely via claude.ai's Remote Control interface encountered a bug where clicking Stop or Interrupt didn't fully cancel the underlying CLI session. The session would appear stopped from the remote interface, but queued messages would never advance after interrupting a stuck tool or prompt. The fix ensures remote stop/interrupt is now equivalent to pressing Esc locally.