Claude Code 2.1.193 adds live path autocomplete and OTEL response logs

• ClaudeCodeLog's 2.1.193 changelog says Claude Code can now route all Bash and PowerShell commands through the auto mode classifier with autoMode.classifyAllShell, extending the safety path beyond just obviously dangerous patterns.
• The 2.1.193 changelog also adds denial reasons to transcripts, toast notifications, and /permissions, which turns blocked actions into something teams can actually audit instead of just notice in the moment.
• ClaudeCodeLog's release summary highlights live file path autocomplete in bash mode, while the full changelog adds automatic reaping for idle background shell commands under memory pressure.
• ClaudeCodeLog's changelog introduces a claude_code.assistant_response OpenTelemetry event, and the default behavior is the real gotcha: deployments already logging prompt content may start receiving response content too unless OTEL_LOG_ASSISTANT_RESPONSES=0 is set.
• ClaudeCodeLog's further-changes post shows the release also touched hidden surface area, including new env vars, model list changes, and a big prompt-mix swing from tool tokens toward system reminders.

You can jump from the official source link in the changelog thread to prompt stats and the fuller diff roundup. The odd little detail in ClaudeCodeLog's further-changes post is the prompt token mix shift, where system-reminder nearly doubled its share even though total prompt tokens barely moved.

The headline change is not autocomplete. It is the decision to send all Bash and PowerShell commands through the same classifier path with autoMode.classifyAllShell.

Before this release, the classifier only handled arbitrary-code-execution patterns, according to the changelog. In 2.1.193, Anthropic also started writing denial reasons into three places:

• the transcript
• the denial toast
• /permissions recent denials

That combination makes blocked actions legible after the fact. The release note in ClaudeCodeLog's summary thread frames it as uniform shell safety checks, while the full itemized changelog makes clear the change is also about auditability.

The other operationally sharp edge sits in telemetry. According to the changelog entry for OpenTelemetry, claude_code.assistant_response is redacted unless OTEL_LOG_ASSISTANT_RESPONSES=1, but when that variable is unset it inherits from OTEL_LOG_USER_PROMPTS. Teams already exporting prompt text can therefore pick up response text on upgrade unless they explicitly force prompts only.

The most visible UX upgrade is live file path autocomplete in bash mode, which ClaudeCodeLog's release summary put in the top three highlights.

The release pairs that with two smaller shell changes that matter more in long running sessions:

• startup notices now point users to /mcp when MCP servers need authentication, per the changelog
• idle background shell commands can now be reaped automatically under memory pressure, also per the changelog

2.1.193 also continues Anthropic's recent cleanup of background-agent behavior. The changelog says launch results no longer tell Claude to end its response while an agent runs, and it fixes a phantom general-purpose (resumed) subagent that could re-run the main conversation after backgrounding.

That follows the background-agent fixes in the 2.1.191 changelog, where stopped agents were made permanently stopped and streaming updates were coalesced to cut CPU use by about 37%.

The follow-up diff post surfaces changes that do not show up in the highlight bullets:

• added env vars: CLAUDE_CODE_FORCE_TIP_ID, CLAUDE_CODE_SHOJI_ENGINE, JAVA_TOOL_OPTIONS
• added model slug: opus-fast-mode-deprecation
• removed CLI option: --mcp-debug
• removed model slugs: o5-qn, o5-u:

The prompt files only grew by two and total prompt tokens by 198, but the mix changed a lot more than the totals suggest. According to the prompt stats summary, tools dropped from 68.2% to 49.0% of prompt tokens while system-reminder jumped from 18.8% to 38.1%.

The same post notes two prompt-string edits: preserving exact indentation when editing files from the Read tool, and documenting grep explicitly as a ripgrep-based search tool that should always be used for search tasks. That is small release-note stuff, but it is new behavior shaping the harness, not just the UI.