Claude Code v2.1.307 Adds OpenTelemetry Safety Events, API Version Command, and MCP Retry Safeguards

542 words
3 minutes
Claude Code v2.1.307 Adds OpenTelemetry Safety Events, API Version Command, and MCP Retry Safeguards
Published: 2026-10-23
By LockDraft
Agents
anthropic
/
AI Agents
/
Developer Tools
/
ai-news
/
claude-code
/
tool-update

What changed#

Anthropic has released Claude Code v2.1.307, introducing new commands, refined observability telemetry, and various bug fixes across local environments, Model Context Protocol (MCP) integrations, and self-hosted runners.

Observability and API Updates#

  • New API versioning command: Added /claude-api api-version to transition Claude API integration code to API version 2026-10-21.
  • OpenTelemetry safety events: Added the claude_code.safety_event OpenTelemetry event. This event fires whenever a tool execution is halted, prompted, or returned due to permission rules, auto mode decisions, safety checks, execution hooks, or sandbox restrictions.
  • Telemetry fixes: Resolved an issue where OTLP telemetry sent from Claude Desktop sessions behind a Claude apps gateway was rejected when OTEL_EXPORTER_OTLP_HEADERS was defined. Additionally, fixed an issue where CLAUDE_STREAM_FIRST_BYTE_TIMEOUT_MS was ignored in apps gateway sessions.

Model and Execution Fixes#

  • Model alias handling: Resolved model-not-found errors when CLAUDE_CODE_DISABLE_1M_CONTEXT was enabled alongside model aliases containing the [1m] suffix, such as opus[1m].
  • Header transmission: Stopped sending the prompt-caching-scope beta header when working behind custom ANTHROPIC_BASE_URL gateways or on Foundry platforms, where the header had no effect.
  • Extended SDK model switch timeouts: Increased the model switch timeout in setModel() from 5 seconds to 15 seconds to prevent agent refusals when API confirmations lag.
  • Tool recovery: Improved resilience against malformed tool calls by attempting up to two retries before terminating a turn with an error.

MCP Server and Cloud Session Improvements#

  • Idempotent HTTP MCP retries: HTTP MCP tool calls that drop mid-connection are now retried only if the tool is designated as read-only or idempotent, preventing unintended side effects from duplicate execution.
  • macOS and sleep fixes: Keychain lock errors on macOS in claude -p mode now report specific error causes and automatically reconnect upon unlock without requiring re-authentication. MCP server timeouts resulting from system sleep on macOS or Linux are also resolved.
  • Cloud and runner persistence: Fixed issues where session restarts in cloud and self-hosted runners reverted /autocompact windows or restored conversations previously wiped with /clear.
  • Self-hosted runner isolation: Sessions running with --use-anthropic-git-proxy no longer commit left-over .claude/settings.local.json files. Additionally, stopping a runner or requeuing a session during its checkout hook no longer leaves orphaned processes running or runner slots locked.

Why it matters#

For engineering teams running AI agents in enterprise workflows or CI/CD pipelines, tracking agent behavior and ensuring transactional safety are critical. The introduction of the claude_code.safety_event OpenTelemetry event provides structured visibility into security blocks and permission prompts, making it easier to audit agent policy compliance alongside existing metrics.

Furthermore, restricting automatic network retries strictly to idempotent HTTP MCP tools prevents duplicate write operations or side effects during network interruptions. These updates build on recent stability efforts across the platform, such as strict hook controls and gateway filters and earlier command safeguards and protocol updates.

What to do next#

  • Update local and runner installations: Upgrade to Claude Code v2.1.307 to pick up the runner process leak fixes and tool call recovery enhancements.
  • Target the new API version: Execute /claude-api api-version within your workspace to shift your Claude API code targeting to 2026-10-21.
  • Ingest safety telemetry: Update your OpenTelemetry ingestion pipelines to listen for claude_code.safety_event if you track sandbox triggers, permission blocks, or auto-mode decisions.

Limitations#

Automatic retries for interrupted HTTP MCP connections are limited strictly to tools marked as read-only or idempotent. Non-idempotent tool calls that fail midway due to closed socket connections will still fail without an automatic retry, requiring explicit error handling by the agent or operator.

Sources#

Share Article

If this article helped you, please share it with others!

Ollama v0.40.9 Adds Local Audio Synthesis, d3 Decision Models, and Cross-Platform Flint
Optimizing SLM Vocabularies Around Inference Pricing Cuts Decode Overhead
Related PostsSmart
1
Claude Code Adds Deep Plugin Mods, Command Safeguards, and MCP Protocol Updates
AgentsAnthropic updates Claude Code with extensible Mods, dangerous command safeguards, Sonnet 5.5 advisor pairing, and MCP protocol updates.
2
Claude Code v2.1.293 Adds Claude Haiku 5.5 Support, Mod Deferred Tools, and MCP Memory Fixes
AgentsClaude Code v2.1.293 adds Claude Haiku 5.5 support, new mod options (isDeferred, agentType), and fixes memory leaks and context compaction bugs.
3
Anthropic Updates Claude Code with Gateway Identity Claims and File Write Safeguards
AgentsClaude Code v2.1.312 brings custom OIDC user ID claims, concurrency limits on execution hooks, and safeguards against truncating large project files.
4
Claude Code v2.1.292 Adds Sub-Agent Effort Controls, Mod Caching, and Critical Security Patches
AgentsAnthropic updates Claude Code with sub-agent effort controls, mod prompt caching, new plugin install flags, and important security patches.
5
Anthropic Releases Claude Code v2.1.295 with Strict Hook Controls, Gateway Filters, and MCP Fixes
AgentsClaude Code v2.1.295 adds strict hook failure options, enterprise gateway model routing, OSC 7501 terminal status, and MCP reconnection fixes.
Random PostsRandom

Comments

Share your thoughts and discuss with everyone

Table of Contents
1
What changed
Observability and API Updates
Model and Execution Fixes
MCP Server and Cloud Session Improvements
2
Why it matters
3
What to do next
Limitations
4
Sources