Claude Code v2.1.122 Adds Bedrock Service Tier, Fixes MCP Tool Discovery and Bash Mode

Anthropic released Claude Code v2.1.122 with several notable fixes and one new feature. The release is primarily a bugfix rollup, but adds a new environment variable for AWS Bedrock users.
New Bedrock Service Tier
Set ANTHROPIC_BEDROCK_SERVICE_TIER to default, flex, or priority. The value is sent as the X-Amzn-Bedrock-Service-Tier header on each API request.
Key Fixes
- PR URL resume: Pasting a PR URL into
/resumenow finds the session that created that PR (GitHub, GitLab, Bitbucket supported). - MCP server visibility:
/mcpshows claude.ai connectors hidden by a manually-added server with the same URL, plus a hint to remove the duplicate. - MCP tool discovery: Fixed ToolSearch missing MCP tools that connected after session start in nonblocking mode.
- Bash mode exit:
!exit/!quitnow runs as shell commands instead of terminating the CLI. - Image resize max: Images sent to newer models are resized to 2000px max per side (was incorrectly 2576px).
- Model effort:
/modelnow shows the Effort option for Bedrock application inference profile ARNs; those ARNs also receiveoutput_config.effort. - Vertex AI / Bedrock structured outputs: Fixed
invalid_request_error: output_config: Extra inputs are not permittedon session-title generation and other structured-output queries. - Vertex AI count_tokens: Fixed 400 errors for users behind proxy gateways.
- OpenTelemetry: Numeric attributes on
api_request/api_errorlog events are emitted as numbers, not strings. Addedclaude_code.at_mentionlog event for @-mention resolution. - Other fixes:
spinnerTipsOverride.excludeDefaultnow works (wasn't suppressing time-based tips). Remote control session idle status redrawing dropped from 2x/sec to 1x/sec to avoid floodingtmux -CCcontrol pipes. Assistant messages appearing blank due to stale view preference fixed. Malformed hooks entry insettings.jsonno longer invalidates the entire file. Voice mode: Caps Lock keybindings now show an error (terminals don't deliver Caps Lock as a key event).
For Whom
This release benefits Claude Code users on AWS Bedrock, those relying on MCP tools in nonblocking mode, and anyone hitting structured output issues with Vertex AI. The bash mode fix is relevant for teams using !exit within builds.
📖 Read the full source: GitHub Claude-Code
👀 See Also

Research Findings on AI Agent Reliability and Development Patterns
A collaborative research session with Claude Opus analyzed 15 papers on AI agents, revealing quantified reliability problems: agents produce 2-4 different action sequences across 10 runs, with 69% of divergence occurring at the first decision. Self-improving agents showed safety refusal rates dropping from 99.4% to 54.4% through their own learning.

DeepSeek V4 Flash Cost Breakdown: Cache Hit Rate and Price Ratio Explained
DeepSeek V4 Flash costs 0.0066x per agentic task compared to Opus 4.7, driven by 97% cache hit rate and 0.02 cache read-write price ratio.
Claude Code v2.1.236: New Default Model Env Var, Sandbox Fixes, and More
Claude Code v2.1.236 adds ANTHROPIC_DEFAULT_MODEL to set the model for new sessions, introduces notify_when_idle for cross-session messaging, and hardens macOS sandbox read-deny rules.

Config Changes with Kimi 2.5 and Opus 4.6
User discusses the performance of Kimi 2.5 for code tasks and config changes, using Opus 4.6 as a coding subagent.