CONFIRMED. Anthropic released Claude Code 2.1.282 on September 24, and the important change is not a new model. It is a broad effort to make long-running agents less fragile. The release fixes cases where resumed sessions, model switches or network failures could alter history and leave an agent with incomplete context.
The main group of fixes concerns continued sessions. Claude Code now handles histories containing undecryptable web-search results, extended-thinking blocks lost after commands such as /model and /rename, and conversations that resend earlier messages in changed form. It also recovers from invalid redacted_thinking blocks and falls back to another model when compaction is refused.
In practice, this targets a worst-case agent failure: the interface looks alive, but resuming silently changes what the system knows. For a small Brazilian company using Claude Code for automation, document analysis or software maintenance, context loss can cost more than an API call. It can create rework, human review and duplicate actions in poorly protected workflows.
The release also strengthens governance. A new startup notice and /status and claude doctor entries show ignored or disabled telemetry variables. Managed permissions now enforce locks despite mistyped boolean values, and partially invalid settings no longer invalidate an entire block. Symlink paths reaching sensitive macOS locations and Bash rules with mid-pattern operators are fixed too.
For teams running their own infrastructure, two details stand out. The Claude apps gateway adds store.readiness_grace_seconds, keeping readiness available through a short Postgres outage such as failover. Claude Code can also use Chrome beside an exclusive managed-mcp.json when policy permits. Small changelog entries, but relevant for availability and integration conflicts.
The caveat: this update does not replace idempotency, retry limits, observability or permission review. The public npm registry still listed 2.1.278 as latest, suggesting this release may arrive first through Anthropic's own channel. That does not invalidate GitHub, but supports controlled rollout and channel checks.
My reading is straightforward: 2.1.282 is maintenance with above-average production impact. It is not a new Claude capability story, but a reliability correction for teams already depending on persistent agents. The gain appears when teams measure resumptions, compaction, 400 errors, permissions and gateway availability before and after upgrading.
Sources
Official release: https://github.com/anthropics/claude-code/releases/tag/v2.1.282
Official changelog: https://code.claude.com/docs/en/changelog
Independent registry: https://registry.npmjs.org/@anthropic-ai/claude-code/latest