Jonathon's AI Wiki

What's New in Claude Code — Week 36 (August 31 – September 4, 2026)

10 min read

Source: ai-research/claude-code-docs-whats-new-w36-2026-09-18.md — official W36 digest fetched from code.claude.com/docs/en/whats-new/2026-w36 on 2026-09-18. Covers v2.1.251–v2.1.261 (August 31 – September 4, 2026).

W36’s defining announcement is Claude Fable 5.1 — the first new Fable model since Fable 5’s launch in June. The week also ships background computer use on Desktop (macOS beta), a live /diff panel in fullscreen rendering, and /skill-doctor for auditing skill context costs. Continues Week 35.

Note: several “Other wins” in the official W36 digest (e.g. PreModelSwitch/PostModelSwitch hooks, /cost prompt-cache line, Enterprise→Opus 5 default) were already documented in Week 35 from changelog data since v2.1.251 is the boundary version.

Key Takeaways

  • Claude Fable 5.1 is available in Claude Code (v2.1.257+), and the fable alias now selects it. The model carries a 1M-token context window, matching Fable 5. In Claude apps gateway sessions, fable still resolves to Fable 5 unless the gateway serves 5.1 — run /model claude-fable-5-1 explicitly in that case. Enterprise seat plans should note that fable at the gateway may lag behind fable at the Anthropic API.
  • Background computer use on macOS Desktop — Claude can now see and act in approved apps while you keep working, without taking over your screen. In beta on Pro and Max plans. This removes the “full-screen takeover” requirement that limited practical use of the computer-use feature for anything you needed to monitor.
  • Live /diff panel in fullscreen rendering (v2.1.260) — /diff now opens beside the conversation instead of launching a blocking viewer. The panel refreshes on every file edit or shell command, and lines selected with the mouse attach to your next prompt. Requires fullscreen rendering, a git repo, and a terminal at least 110 columns wide.
  • /skill-doctor reveals which skills you never use (v2.1.252+) — opens the plugin manager’s Stats tab showing each skill’s context cost and usage frequency. Every skill description loads into context on every turn regardless of use; pruning unused skills is a direct context-reduction with no capability loss. Complements the claude-api skill’s 8× context reduction by on-demand loading (documented in Week 34).
  • managedMcpServers for org-wide HTTP/SSE MCP servers — organizations can now deliver HTTP and SSE MCP servers to every user via managed settings, in addition to the stdio servers users add themselves. This is the managed equivalent of a personal .mcp.json entry, pushed at organization scope.
  • /cd now hot-reloads project settings, hooks, and skills — moving a session to a new directory with /cd now immediately applies that directory’s .claude/settings.json, .mcp.json, hooks, skills, and subagents. Previously these only took effect after a --resume. Eliminates a surprise for multi-project workflows that use /cd to switch context.
  • keybindingFlavor is deprecated — word-editing shortcuts now follow readline for everyone: Ctrl+W deletes to previous whitespace, Alt+B/Alt+F/Alt+D treat / and . as word breaks. The keybindingFlavor setting no longer has any effect. (W34 introduced it; W36 bakes readline behavior as the universal default.)
  • bashOutputMaxChars and taskOutputMaxChars — new settings (up to 128,000 characters) let Claude receive more inline output from successful commands and background tasks. Previously long output was always truncated in the tool result; now you can raise the threshold for jobs that produce verbose but useful output.
  • bypassPermissions in project settings no longer starts bypass sessions — setting defaultMode: "bypassPermissions" in a project’s .claude/settings.json or .claude/settings.local.json is now ignored; the session starts in Manual mode instead. Only user or managed settings — or --permission-mode at the CLI — can start a bypassPermissions session.

Claude Fable 5.1

Fable 5.1 is the first Fable model update since the original Fable 5 launch in early June. Its 1M-token context window matches Fable 5; the release notes do not itemise capability changes over the base Fable 5 — the version increment is typical of inference/reliability improvements rather than a capability jump.

Alias behavior split: The fable alias resolves differently depending on how your session is authenticated:

  • Anthropic API / direct plans (Pro, Max, Team, Enterprise): /model fable now selects Fable 5.1. The picker lists Fable once the server reports it available for your organization; typing /model fable checks availability directly.
  • Claude apps gateway: fable still resolves to Fable 5. Use /model claude-fable-5-1 explicitly if your gateway serves 5.1. Requires v2.1.257 or later.

For wiki readers tracking the Fable 5 entity article: 5.1 is a point release, not a new entity — the spending-share discussion (11% of Anthropic tool spend on Ramp data) and the two competing explanations (price vs. data retention) remain the current analytical frame.

# Switch to Fable 5.1 and save as default:
> /model fable
# If on a gateway and you know it serves 5.1:
> /model claude-fable-5-1

Background computer use on Desktop (macOS beta)

Computer use in the Desktop app no longer requires handing over full screen control. On macOS, Claude can see and act in the apps you’ve approved while you keep working in the main window. The permission card reads “Let Claude see and act in the apps you approve, in the background or with full control of your screen” — both modes are now available, with background being the new one.

Status: beta, available on Pro and Max plans. The feature is Desktop-only (not the CLI).

Practical implication: the primary friction for computer-use workflows was the full-screen requirement — you had to stop working while Claude used the computer. Background mode removes that constraint for tasks where Claude doesn’t need to interact with the same app you’re actively using. Multi-screen setups are the natural fit.

/diff panel (v2.1.260)

The old /diff behavior was a blocking viewer you had to close before continuing. The new behavior is a persistent panel that lives beside the conversation. It:

  • Lists changed files with added/removed line counts
  • Refreshes automatically on every Edit, Write, or Bash command that modifies files
  • Accepts mouse selection: selecting lines attaches them to your next prompt

Close with a second /diff or the ✕ in the panel header. Prerequisites: fullscreen rendering active, inside a git repo, terminal at least 110 columns wide.

This is the first diff surface in Claude Code that provides live, turn-by-turn visibility into what changed without interrupting the conversation flow.

/skill-doctor — context cost audit (v2.1.252+)

Every skill’s description loads into the system prompt on every turn, whether or not Claude ever calls it. /skill-doctor surfaces this cost directly: it opens the /plugin manager’s Stats tab showing each skill’s context token cost and usage frequency.

The actionable output: skills you haven’t used in the last N sessions are candidates for disabling. The tab also shows cost in tokens, so you can quantify the context budget you recover by disabling a skill you rarely use.

Practical note: /skill-doctor requires feature-flag fetching. Sessions started with --no-feature-flags or similar flags that skip fetching will not have this command available.

> /skill-doctor
# Opens plugin manager → Stats tab
# Columns: skill name | context tokens | usage (last 30 sessions)

For organizations shipping skills as plugins: /skill-doctor is the user-facing audit tool. Skills that consistently show zero or near-zero usage are disablement candidates. The progressive disclosure pattern (lazy-loading reference docs) documented in Agent Skills Overview reduces the per-skill context cost even for kept skills.

managedMcpServers for org-wide delivery

Previously, managed settings could only deliver stdio-protocol MCP servers. The managedMcpServers setting now accepts HTTP and SSE servers too, with the same org-wide delivery model. This enables cloud-hosted or internally-hosted MCP servers to be pushed to every user without requiring each user to update their local config.

Why this matters: stdio servers are fine for local tooling but cannot be centrally hosted. HTTP/SSE servers can run behind an org’s own infrastructure. With managedMcpServers, an IT/platform team can give every user access to internal data sources (CRM, docs, ticketing) without touching individual machines.

Other wins from the official digest

These items from the W36 “Other wins” are new and not already in Week 35:

  • /cd hot-reload — project settings, hooks, .mcp.json, skills, subagents apply immediately when /cd moves to a new directory. No --resume needed.
  • bashOutputMaxChars / taskOutputMaxChars — raise these settings up to 128,000 chars for more verbose inline output from Bash commands and background tasks.
  • readline word-editing universal — keybindingFlavor deprecated; readline is now the only mode. Ctrl+W to whitespace, Alt+B/Alt+F/Alt+D with / . as word breaks.
  • bypassPermissions project-scope restriction — project-level defaultMode: "bypassPermissions" now silently ignored; session starts in Manual mode instead. User or managed settings remain effective.
  • VS Code: click model name to open picker — the model name at the bottom of the VS Code prompt box is now clickable to open the model picker directly.
  • VS Code: Output styles in command menu — select “Output styles” in the Customize section to pick a style, including custom ones authored in skills.

Try It

  • Switch to Fable 5.1: /model fable on Pro/Max/Team/Enterprise. On a gateway, use /model claude-fable-5-1 if your gateway is serving 5.1.
  • Run /skill-doctor — identify which skills you haven’t used and disable them. Even a single unused skill with a large description recovers hundreds of tokens per turn.
  • Try /diff in fullscreen — toggle with /diff mid-session in a git repo. Select lines with the mouse to scope your next prompt to specific changes.
  • Check any automation that uses bypassPermissions in project settings — those sessions now start in Manual mode. Move the setting to user or managed settings, or pass --permission-mode bypassPermissions.
  • Audit /cd workflows — if you relied on --resume to apply new directory settings after /cd, test that the immediate hot-reload works as expected for your project setup.

Open Questions

  • What changed in Fable 5.1 over Fable 5? The release note does not enumerate capability changes — standard for a point release. Whether 5.1 has meaningful reasoning improvements, latency changes, or cost changes is not stated.
  • Which macOS versions support background computer use? The changelog says “macOS” but does not specify a minimum version or whether Apple Silicon vs Intel matters.
  • Does /skill-doctor count usage per-skill or per-plugin? A plugin with multiple skills — does it show per-skill data or a plugin-level aggregate?
  • Is bashOutputMaxChars per-command or cumulative? The changelog describes it as a “max chars” setting for a “successful command” — unclear whether it applies per individual command invocation or across all Bash output in a session.
  • What happens to skills synced from claude.ai when an organization disables Skills? (Noted in the v2.1.273 changelog — a fix was shipped: “Fixed skills synced from claude.ai staying after organization disables Skills.” So the prior behavior was that they persisted; v2.1.273 fixes the removal. Noted for the organizations that may have synced skills in the W36 window.)
  • What’s New — Week 35 — the preceding digest (v2.1.240–251); security batch, PreModelSwitch/PostModelSwitch hooks, restricted mode, /loop on Bedrock/Vertex/Foundry.
  • What’s New — Week 37 — the following digest (v2.1.263–269); plugin eval, Desktop pane pop-out.
  • Mythos 5 — entity article for the Fable model family; 5.1 is a point release within it.
  • CLI Reference — bashOutputMaxChars, taskOutputMaxChars, maxEffortLevel, managedMcpServers settings and deprecated keybindingFlavor.
  • Hooks — PreModelSwitch/PostModelSwitch events (boundary features from v2.1.251, documented in W35 but officially included in W36 digest).
  • Agent Skills Overview — /skill-doctor is the audit companion to the progressive-disclosure pattern described there.
  • Plugins and Marketplaces — managedMcpServers for org-wide MCP server delivery.
  • Essential MCP Servers — managedMcpServers enables org-delivered HTTP/SSE MCP servers without per-user config.

Graph View

Backlinks