Detailed Analysis
Anthropic's Claude Code command-line interface experienced a significant but short-lived service disruption on April 24–25, 2026, when version 2.1.120 introduced a crash-on-startup bug that rendered the tool non-functional for a subset of users attempting to resume prior sessions. The incident began around 01:45 UTC on April 25 and was resolved within approximately 48 to 60 minutes, with Anthropic acknowledging it through an official status update posted to status.claude.com. The root cause was traced to JavaScript runtime errors — specifically exceptions such as "g9H is not a function" and "F1H is not a function" — triggered when users invoked the `--resume` or `--continue` flags to pick up prior Claude Code sessions. Beyond the CLI, the incident also degraded adjacent services including claude.ai's web interface (exhibiting partial page loads and spinning states), console.anthropic.com, and api.anthropic.com, suggesting the fault had some systemic reach despite its narrow technical origin.
Anthropic's response followed a standard rollback protocol: the company automatically reverted affected clients to version 2.1.119, with the fix propagating to most users on their next automatic update check. Users who needed an immediate resolution could manually execute `claude install 2.1.119` to downgrade without waiting for the automated rollback. A further workaround circulated in community discussions — launching a new session via `claude --new` rather than attempting to resume an existing one — allowing productive use of the tool while the fix propagated. The speed of the response and the availability of both automated and manual remediation paths reflect a mature incident-management process, limiting the practical impact of the outage to under an hour for most affected users.
The incident is notable as a case study in the fragility that can accompany rapid versioning in developer-facing AI tools. Claude Code has been iterating through minor versions at a high cadence, and the v2.1.120 regression follows a pattern of instability that was also observed around v2.1.119, with some community members rolling back further to v2.1.117 to avoid related issues. This points to a broader challenge in the agentic AI tooling space: CLI tools like Claude Code, which are deeply integrated into developer workflows and often invoked in long-running or resumable sessions, carry a higher risk profile from faulty releases than stateless web interfaces, because a startup crash eliminates the user's ability to access in-progress work entirely. The `--resume` and `--continue` flags — precisely the features that make Claude Code powerful for sustained, complex coding tasks — became the vector through which the regression manifested.
Within the broader context of Anthropic's operational history in early 2026, this incident is relatively minor. A more significant set of API and claude.ai errors was reported around April 15, 2026, and the company has faced intermittent reliability scrutiny as usage of both its consumer and developer-facing products has scaled. The Claude Code CLI in particular has grown into a critical piece of infrastructure for software engineers who use it as an AI-native development environment, making uptime and version stability increasingly consequential. Anthropic's decision to issue automatic rollbacks rather than pushing a hotfix forward suggests a conservative philosophy around production stability — prioritizing known-good states over rapid patching — which is consistent with the safety-first framing the company applies to its model deployment practices.
The episode also illustrates the increasingly important role of community-driven observability in AI product incidents. The Reddit performance megathread referenced in the status post, along with GitHub issues filed within minutes of the crash becoming widespread, functioned as a real-time distributed bug report that likely accelerated Anthropic's diagnosis. As agentic AI tools become more deeply embedded in professional workflows, the feedback loops between developer communities and AI companies are compressing, enabling faster triangulation of faults. This dynamic — where users are simultaneously consumers, testers, and diagnosticians — is becoming a defining feature of how frontier AI tooling evolves in production.
Read original article →