Claude Code 2.1.317 expands controls and notifications

On November 2, 2026, at 22:25 BRT, Anthropic published Claude Code v2.1.317. The release updates the agent, Agent SDK and gateway, with no new model. See the official list.

The list includes permission warnings, blocking on hook failures and new operational data, plus fixes for sessions, MCP and execution.

In this article you will see:

• The new confirmations before sharing data and connectors
• The behavior of onFailure: "block" in Agent SDK hooks
• The new reports and attributes for teams
• The fixes that affect sessions, MCP and automations

Claude Code 2.1.317 asks for confirmation before two sensitive actions

The release adds a warning to /schedule when a new cloud task receives unsolicited connectors. The interface explains how to remove them, preventing unexpected integrations from going unnoticed.

This does not block all configuration or review permissions. The warning applies only to unsolicited connectors during the creation of a scheduled cloud task.

Another confirmation appears before a web or shell tool call sends the account email address to the network. This makes explicit data that could leave the environment during an agent initiated action.

The warning increases transparency, but it does not detect all sensitive data. The team should still check destination and content before sending.

Image: Anthropic, capture from the official Claude Code page.

Version 2.1.316 and its subagent controls covered previous changes; v2.1.317 brings the warnings and fixes described here.

Agent SDK can block an action when a hook fails

The Agent SDK now accepts onFailure: "block" in hook callbacks. If the callback fails, the corresponding action is blocked instead of continuing as if validation had completed.

A hook is a callback executed at defined points in the flow. Teams can insert their own checks in it before or after agent actions.

The option creates a fail closed mode: if a required check fails, the action does not proceed automatically. This helps avoid a policy depending on an unavailable callback.

The release does not state that all hooks now block by default. The change is the availability of the onFailure: "block" configuration; the team needs to define where to apply it and test the expected behavior.

This is not a complete security mechanism. The option covers callbacks configured in the SDK and does not replace access controls, validation or human review.

Claude Gateway gains data to observe costs and sessions

The Claude apps gateway API gains GET /v1/organizations/analytics/cost, to query past costs by period and group results by total, developer, day or identity group.

For this, gateways with admin: record daily the identity groups of each developer. Retention follows admin.identity_retention_days, which administrators should review.

v2.1.317 includes anthropic.plan, anthropic.seat_type, anthropic.provider and anthropic.custom_base_url in metrics, events and spans of OpenTelemetry. These attributes relate telemetry to account and provider settings.

OpenTelemetry standardizes signals such as metrics and distributed traces. The note lists the new attributes, but does not say that every environment exports them by default.

There is also a fix for claude -p on a self hosted runner. Variable CLAUDE_CODE_DISABLE_NESTED_SESSION_ID=1 makes execution inside a session use the identifier and transcription of that session, as in previous versions.

For context on MCP isolation in auxiliary calls, see the TokenEconomizer article. This update concerns Anthropic's gateway.

Claude Code 2.1.317 fixes MCP, authentication and automation issues

The release fixes project and local MCP servers that could disconnect after the first turn in some Claude Desktop sessions, affecting integrations used during the conversation.

claude auth status could exit before saving a slow token update, causing the next run to end the session. The version fixes this ordering.

In /schedule, results now show the next run time and the task link again. Another fix allows listing and removing stored tasks after disabling the feature.

A safety check in automatic mode stops retrying failures on HTTP 403 or 422 gateways. The 403 case with apiKeyHelper remains unchanged.

If a full disk or quota prevents writing a transcript or configuration, claude -p now surfaces an error and returns exit code 1, in text and JSON formats as well.

The release lists fixes for long sessions, compaction, proxy telemetry, memory and headless startup. See the full list for the other details.

How to evaluate Claude Code 2.1.317 without disrupting your workflow

Identify whether your environment uses scheduled tasks, connectors, Agent SDK hooks, administrative gateway or self hosted runners. The changes do not affect all teams the same way.

On a non critical task, validate connector warnings and email delivery. Confirm the responsible person understands why it happened and knows how to review or remove the integration.

If you use Agent SDK callbacks, test one successful case and one failure case. Confirm that onFailure: "block" blocks the intended action, without assuming the option is already enabled.

Administrators should review cost API configuration, telemetry attributes, who can access reports, which groups are logged and the retention period.

To update, follow the official documentation and your team's procedures. The release does not introduce a new install command, so validate instructions before reusing old commands.