Claude Code 2.1.261 prunes skills that eat context

Claude Code 2.1.261 adds /skill-doctor, a command that reports which loaded skills go unused and how much context they cost. The release, published in the changelog on Github, also raises the inline output ceiling for commands and background tasks to 128K characters through two new settings.

At a glance

• bashOutputMaxChars and taskOutputMaxChars control how much command and background-task output reaches Claude inline before it is written to a file, with a ceiling of 128K characters.
• /status and claude doctor now carry an Organization policy line explaining why the policy could not be loaded, such as a proxy that does not pass the endpoint through.
• A new flag, --append-subagent-system-prompt-file, reads the subagent system prompt from a file for prompts too large to pass on the command line; the VS Code model picker becomes one flat list.

Skills, subagents and background tasks all draw on the same context window, and until now the cost of a loaded but unused skill was invisible. /skill-doctor and the two output-size settings read as an attempt to make that budget legible and adjustable rather than fixed. The rest of the release is mostly repair work: a large share of the fixes concern corporate proxies, managed settings and Remote Control, which suggests where the load now sits.

/skill-doctor lists loaded skills that go unused and what they cost in context

/skill-doctor shows which loaded skills are never used in a session and what each costs in context, so the set can be pruned. bashOutputMaxChars and taskOutputMaxChars raise how much command and background-task output arrives inline before it is saved to a file, up to 128K characters.

--append-subagent-system-prompt-file reads the subagent system prompt from a file, for prompts too large to pass on the command line. /status and claude doctor gained an Organization policy line that states why an organization's policy could not be loaded, for example when a proxy does not pass the endpoint through.

The /model picker and the VS Code model pill now show a recognized model's name rather than its raw Bedrock, Vertex AI or LLM gateway ID. /context falls back to a local token estimate when the token-counting API is unavailable, instead of issuing extra small-model requests.

The Bedrock setup wizard now times out instead of hanging on an unresponsive credential helper

The Bedrock setup wizard hung when AWS or an AWS credential helper never responded; it now times out with a clear error, and its model checks work behind a TLS-inspecting proxy. A gateway 403 on the managed settings load now says Claude Code may not be enabled for the organization, instead of advising a new sign-in.

Machines whose managed settings pin forceLoginMethod to gateway ignore a leftover API key or claude.ai login and require /login; Bedrock, Vertex AI and Foundry sessions are unaffected. On Vertex AI with GOOGLE_APPLICATION_CREDENTIALS set, API client creation no longer repeats Google Cloud project discovery or spawns extra gcloud processes.

claude.ai connectors could stay absent for a whole session when the startup fetch timed out; the CLI now retries in the background. gcpAuthRefresh no longer opens a browser at startup when a slow Google credential check still returns a valid credential. Remote Control's inbound event stream failed behind TLS-inspecting proxies on native Windows.

Remote Control reports a message to an offline session as queued rather than delivered

Remote Control fixes cover a stale permission mode when a phone, browser or claude.ai app attaches to a terminal session, and sessions that kept showing a spinner and Stop button after a turn was stopped from a phone. A message sent to an offline session now reads as queued until that machine reconnects.

In VS Code, the MCP servers dialog gains an Add server form and a Remove action, so servers can be added and removed without leaving the IDE. An Output styles walkthrough writes a custom output style file and lists it right away, and Archive session joins the session list's right-click menu.

The session list marks sessions open in a terminal, another VS Code window or Claude Desktop with a hollow ring, permission and question prompts gain a fold button that keeps the conversation readable, and the model picker becomes one flat list with older model spellings last.

Where keybindingFlavor stops applying. The prompt's word-editing keys now follow Bash: Ctrl+W deletes back to whitespace, Alt+F and Alt+D stop at word end, and punctuation separates words. The keybindingFlavor setting has no effect after this release, and the changelog names no replacement for it. Auto mode also treats a link that packs content into a public diagram renderer's URL as an upload, so it is not auto-approved unless requested.