Claude docs changes for August 16th, 2026 [diff]

Executive Summary

Claude Code adds a new [claude-code:unrecognized_model] stderr/debug-log diagnostic that fires whenever a request uses a model ID the installed version doesn't recognize, on every provider, with guidance on fixing it via modelOverrides.
The task-tracking tools (todos, TaskCreate/TaskGet/TaskList/TaskUpdate) are now opt-in: a session only gets them if one is explicitly named in allowedTools/tools, which also affects when Claude creates todos and when agent teams get a shared task list.
Docs now consistently refer to the "Agent tool" (previously "Task tool") for spawning subagents, matching a recent rename.
The Compliance API's "List organization users" reference page is now fully documented (it previously rendered only placeholder "Loading" text), covering the GET /v1/compliance/organizations/{org_uuid}/users endpoint, its parameters, and response shape.

Naming a task-tracking tool in allowedTools now also opts the session into those tools. [line 313] [Source]

The TaskCompleted hook's trigger is clarified as "a task is marked completed" rather than "background task completes," and now links to its own docs. [line 171] [Source]

Tracing docs now say subagents are spawned through the "Agent tool" rather than the "Task tool." [line 140] [Source]

Naming a task-tracking tool in allowed_tools now also opts the session into those tools. [line 61] [Source]

The allowed_tools reference now notes that naming a task-tracking tool there also opts the session into those tools. [line 766] [Source]

Clarifies that Claude only creates todos in a session that has the task-tracking tools. [line 34] [Source]

The allowedTools reference now notes that naming a task-tracking tool there also opts the session into those tools. [line 394] [Source]

The shared task list and "all agents can see task status" language is now scoped to agents that have the task-tracking tools. [lines 33, 67, 226, 250]

Teammates sharing a task list is now conditioned on the team having the task-tracking tools. [line 36] [Source]

Subagents are now spawned with the "Agent tool" rather than the "Task tool." [line 190] [Source]

--allowedTools/--allowed-tools and --tools now note that naming a task-tracking tool opts the session into those tools. [line 56, line 122]

Added a new "Unrecognized model ID on a request" section documenting the [claude-code:unrecognized_model] diagnostic line: when it's written, its JSON payload (model, query_source), where it goes (stderr under -p, debug log otherwise), and how to silence it with modelOverrides. [lines 2024-2058] [Source]
Notes the diagnostic line is written on every provider, even when the "not a recognized model id" check itself only runs on the Anthropic API. [line 1181] [Source]
The TaskCreate background-subagent tool-drop example was removed, leaving LSP as the example of a tool background subagents drop. [line 1678] [Source]

Notes that Claude Code can still write the new unrecognized-model diagnostic line even on providers where the model-ID check itself doesn't run. [line 102] [Source]
Documents that adding a modelOverrides entry for a gateway alias stops the [claude-code:unrecognized_model] diagnostic line for that ID. [line 660] [Source]

A completed background subagent now stays listed in /tasks for the same window as the footer hint, rather than "until the session cleans up its task list." [line 728] [Source]
Clarifies you can only resume a self-stopped subagent by typing into its transcript while its row is still in the subagent panel. [line 878] [Source]

The page previously rendered only placeholder "Loading" text; it now fully documents GET /v1/compliance/organizations/{org_uuid}/users, including the limit/page query parameters, the data array fields (id, created_at, email, full_name, organization_role), pagination fields (has_more, next_page), and example request/response. [lines 1-138] [Source]