News

agenteseautomao

STRONG SIGNAL: Claude Code 2.1.281 fixes memory, retry and remote-session failures

The official Claude Code 2.1.281 release bundles fixes that reduce silent agent failures: lost reasoning history, retries after network errors, permissions, MCP OAuth and remote sessions. It is a verifiable repository signal, not a new-product announcement.

STRONG SIGNAL. Claude Code 2.1.281, published in Anthropic's official repository on September 23, concentrates reliability fixes in the most expensive parts of running agents: recovering context, retrying calls without loops, and keeping remote sessions alive.

The most sensitive fix affected sessions with an earlier advisor result that the API could no longer read. The failure could surface as one broken request on every turn and repeated loss of earlier reasoning. The release says the history is now repaired once instead of continuing to poison the conversation.

The retry watchdog also received important fixes. Sessions using CLAUDE_CODE_RETRY_WATCHDOG could fail on the first 5xx or dropped connection after a run of 429 or 529 waits. In another case, a long Retry-After from a 5xx made the client sleep without a clear cap or feedback. For teams running agents in automation, this is less glamorous than a new feature but much closer to real operational damage.

The list also covers the Artifact tool in Remote Control sessions, MCP OAuth tokens lost on macOS when the keychain was locked, attachment downloads, specific PDF-page reads, and proxy errors returned as HTML. Slack channels where Claude stopped replying inside threads also received a fix.

For Brazilian small and midsize companies, the practical effect is fewer hard-to-diagnose failures in support, document analysis, coding automation and MCP integrations. The update does not remove the need for retry limits, observability and application-side idempotency. It improves the client; it does not magically turn a fragile architecture into a reliable operation.

My reading is simple: this is a maintenance release with above-average operational impact. It does not prove a new model capability and should not be marketed as a product launch. Teams using Claude Code in production have a concrete reason to upgrade in a controlled environment and measure context loss, 5xx recovery, MCP behavior and remote-session stability.

Sources

Official release: https://github.com/anthropics/claude-code/releases/tag/v2.1.281
Official changelog: https://code.claude.com/docs/en/changelog
Independent coverage: https://news.ycombinator.com/item?id=49814947

WA