Claude Code 2.1.300: Separate prompt decorations, GenAI observability attributes

1. Separating UI decorations from actual input

Released on October 14, 2026, this version adds virtualText for input decorations to mods that extend the UI or behavior. This text appears in the input field but is not part of the draft or submitted prompt. It is more like a sign laid over a sheet of paper than text written on it.

prompt.change is a mod hook triggered when input content changes without direct typing. A hook is a connection point that calls an extension when a specific event occurs. This includes loading history, clearing, undoing, and plugins filling in content.

The distinction matters. Guidance needed only on screen can appear as decoration, while changes to the actual draft from sources other than typing can use the change hook. These separate roles keep displayed text distinct from editable text.

2. Adding common labels to activity logs

OpenTelemetry is a framework for observing workflows, usage, and more. This update adds GenAI semantic convention attributes to spans that trace individual operations. The change gives providers, conversations, tools, and errors common labels.

For example, gen_ai.provider.name holds the provider name, and gen_ai.conversation.id holds the conversation identifier. Adding standard attributes does not automatically change existing collection tool settings. The release notes give no collector configuration or dashboard migration instructions.

The api_request event also gains metrics that break down cache writes by retention period. This means cache writes can be examined separately for the 5-minute and 1-hour retention periods.

The two metrics are named cache_creation_5m_tokens and cache_creation_1h_tokens. They distinguish cache write totals, not response speed improvements or pricing changes. The source provides no measurements supporting such effects.

3. The advisor requires opt-in

Claude Platform on AWS now has an /advisor command. The advisor does not run until a model is selected for that role through this command, model settings, or a launch option. Updating alone does not activate it automatically.

The environment variable setting for disabling the advisor tool is shown below, exactly as written in the release notes.

CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1

The advisor tool's specific tasks, available models, and costs are not listed. There is thus no basis for presenting it as an expert advisor in a particular field or assuming the same command works in other provider environments.

4. Checking log accuracy and sharing scope

A bug logged prompts modified or discarded by a mod's prompt.submit hook as though they were the user's original input. This is fixed for observability events, interaction spans, and session title requests. Alongside separate UI decorations, this suggests logs now better reflect actual input processing results.

Local session logs now also retain each API request's system prompt, tools, and parameters. The source says this enables exact request reproduction from logs users choose to share. Reproducibility improves, but shared logs also contain more information, making it prudent to review them before sending.

This update clarifies the boundaries between UI extensions and activity logs, but opting into the advisor and reviewing shared logs still require separate decisions.