Claude Code v2.1.300 Adds /advisor, Mod Hooks, and Split Cache Telemetry

Claude Code v2.1.300 landed, and most of the changelog is mod API surface plus telemetry fixes. If you write mods or run Claude Code through OpenTelemetry, this one matters more than the version number suggests.

New: /advisor on Claude Platform on AWS

There's a new /advisor command on Claude Platform on AWS. Nothing runs until you actually pick an advisor — you can set one via /advisor, an advisorModel config value, or the --advisor flag. Setting CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1 turns the whole thing off.

Telemetry: cache writes split by lifetime

The OpenTelemetry api_request event now carries two new fields, cache_creation_5m_tokens and cache_creation_1h_tokens. Previously you'd see a single cache-creation number; now you can separate 5-minute and 1-hour prompt-cache writes, which matters if you're costing sessions by TTL tier.

Trace spans also picked up GenAI semantic-convention attributes: gen_ai.provider.name, gen_ai.conversation.id, gen_ai.tool.name, and error.type.

Mod API changes

• virtualText added to prompt decorations: text drawn in the prompt box that never enters the draft or the submitted prompt.
• $.ui.view({ agentId }) shows an agent's transcript in the main view; $.ui.view({}) goes back.
• New dir prop on a mod's Text for right-to-left text — on Windows and in VS Code's terminal, rows read RTL and sit flush right.
• New prompt.change hook, fired when the prompt box text changes without a typed edit: recalled history, a clear, an undo, or a plugin fill.
• The agent.spawn hook can now rewrite a workflow agent's prompt and model. Before this release it could only refuse workflow agents.

Notable fixes

• Prompt rewrites via a mod's prompt.submit hook were still being recorded as typed in the OTel user_prompt event, interaction span, and session title requests.
• Agent hooks were ignoring their model: aliases like haiku now resolve, a subagent's model no longer overrides it, and an unavailable model falls back to the caller's instead of letting the action run.
• Cloud sessions blamed an organization policy for blocking a host when it was the session's own network settings.
• Subagents you'd typed messages to vanished from /tasks shortly after finishing.
• The agents view treated a burst-typed word like "up", "down", or "escape" as the actual key, which could move the peek mid-reply or clear a draft.
• ENABLE_TOOL_SEARCH=auto could switch off mid-conversation, e.g. while MCP servers reconnect after a resume.
• Plugin WorktreeCreate and WorktreeRemove hooks didn't run under --worktree (still not under --tmux).
• --worktree sessions could start from the wrong commit when a tag or branch was named like a commit ID, and failed to reopen worktrees in reftable repos.

Full changelog below — there are more fixes around /branch, claude purge, Unicode line separators (U+2028/U+2029), forceLoginOrgUUID leaving tokens behind, and the Artifact tool's verbose line reading.