Claude Code v2.1.89: StructuredOutput Reliability, Autocompact Loop Fix, and Thinking Summaries Off by Default
Claude CodeClaude Code v2.1.89 delivers a trio of reliability improvements for production agentic use. StructuredOutput parsing is now more robust, preventing malformed responses from crashing agent pipelines. An infinite loop bug in the autocompact flow has been resolved. Thinking summaries are now disabled by default to reduce noise in standard workflows, with users able to opt back in when extended reasoning traces are useful.
Sources & Mentions
2 external resources covering this update
Reliability and Defaults Tuned in v2.1.89
Claude Code v2.1.89 is a targeted reliability release focused on three pain points that surfaced in real-world agentic usage: structured output fragility, a looping bug in the autocompact system, and noisy thinking summaries enabled by default.
StructuredOutput Parsing Made More Robust
Claude Code uses structured output to return machine-readable results from tool calls and agent pipelines. Previously, certain edge-case responses could cause the parser to fail ungracefully, breaking agent workflows in the middle of execution. The v2.1.89 fix hardens the parser to handle malformed or unexpected structured output without crashing, making pipelines more resilient to model variability.
This is particularly relevant for users running Claude Code in headless or scripted contexts — CI pipelines, multi-agent orchestration systems, or automated code review flows — where a single parsing failure can orphan an entire task.
Autocompact Infinite Loop Resolved
The autocompact feature automatically compresses the conversation context when it approaches the model's context window limit, allowing long sessions to continue without the user manually invoking /compact. A bug introduced in a prior release caused autocompact to enter an infinite loop under certain conditions, effectively hanging the session.
This fix is a direct successor to the /compact deadlock fix shipped in v2.1.85 — together, the two releases address the two primary failure modes in long-session context management.
Thinking Summaries Now Off by Default
Claude Code's extended thinking feature produces intermediate reasoning traces — the model's internal chain-of-thought — that can be surfaced as visible summaries in the UI. While useful for debugging and transparency, these summaries add visual noise during routine tasks where users simply want results.
In v2.1.89, thinking summaries are turned off by default. Users who want to see reasoning traces can re-enable them explicitly. This default change aligns Claude Code's behavior with how most users actually work: extended thinking summaries are valuable occasionally, not constantly.
Summary
V2.1.89 does not introduce new features. Instead, it closes gaps in three areas that matter for production reliability: structured output handling, long-session stability, and UI defaults. For teams running Claude Code at scale, this release removes several sources of silent failure and unnecessary noise.