Source: ai-research/claude-code-release-v2-1-{265,266,267,268,269,270}.md — anthropics/claude-code GitHub release notes for v2.1.265 through v2.1.270, published 2026-09-08 to 2026-09-12 and fetched with gh release view on 2026-09-29. Also ai-research/claude-code-docs-whats-new-w37-2026-09-18.md (the official W37 digest, fetched 2026-09-18) and ai-research/claude-code-docs-changelog-2026-09-18.md (the code.claude.com changelog for v2.1.268 onward, fetched the same day).

Product: Claude Code · URL: https://github.com/anthropics/claude-code/releases · Official digest: https://code.claude.com/docs/en/whats-new/2026-w37 · Date: 2026-09-08 → 2026-09-12

Six releases (v2.1.265 through v2.1.270), two of them single-fix releases for regressions the week introduced. The headline for plugin authors is claude plugin eval, which runs a plugin’s eval suite and produces scored, reproducible results. The headline for everyone else is quieter: the task-tracking tools (TodoWrite and the Task* tools) became an allowlist of older models, maxEffortLevel lets admins cap effort, and v2.1.267 shipped a large batch of prompt-cache stability fixes. Continues Week 36.

The official W37 digest (dated September 7–11, releases v2.1.263 → v2.1.269) has two headline features: claude plugin eval and Desktop pane pop-out. v2.1.263 shipped on 2026-09-06 and is recorded in Week 36; releases from v2.1.271 (2026-09-14) onward are in Week 38.

Key Takeaways

  • claude plugin eval runs a plugin’s eval suite against Claude Code and returns scored, reproducible results as JSON plus an HTML report (v2.1.269). By default it also runs each case without the plugin, so the score shows what the plugin contributes; claude plugin eval init drafts the suite for you. See Plugins and Marketplaces.
  • Desktop panes pop out into their own windows (official digest, Desktop app): drag the diff or terminal to a second screen and dock it back later.
  • Task-tracking tools are now offered only on older models — Claude 3.x, Opus 4.0–4.7, Sonnet 4.0–4.6 and Haiku 4.5; set CLAUDE_CODE_ENABLE_TODO_TOOLS=1 to get TaskCreate/Get/Update/List and TodoWrite elsewhere (v2.1.268). Week 33 removed them from named newer models; this turns that denylist into an allowlist, so any model not on the list — including future releases — starts without them.
  • maxEffortLevel (top-level or per model under modelSettings) caps effort on every provider including Bedrock, Vertex and Foundry; users can still pick a lower level, and any higher level runs at the cap (v2.1.267). See Model vs Effort.
  • Effort frontmatter now works on pinned-effort models. effort: on custom commands, skills and subagents was being ignored on Opus 4.7, Opus 4.8 and Fable 5, whose default effort is still pinned (v2.1.267).
  • Prompt-cache stability batch (v2.1.267): /model switches no longer re-send every tool definition, commit and PR attribution text arrives as a conversation note that updates on model change, and mid-session MCP and plugin tools arrive as deferred definitions on supported models.
  • WebFetch now fails after 300 seconds on a server that never finishes the response; CLAUDE_CODE_WEBFETCH_DEADLINE_MS overrides it, and 0 turns it off (v2.1.268).
  • Claude Tag public channels no longer share memory (v2.1.268): each public channel keeps its own notes, Claude no longer recalls notes saved in other public channels, and workspace notes stay shared. See Claude Tag.
  • /plugin changes apply when the menu closes — /reload-plugins is no longer needed after install, enable or disable (v2.1.268).
  • Auto mode denials name the rule that blocked the action, such as [Data Exfiltration], and ask Claude to try a safer method and finish unrelated work before stopping to ask you (v2.1.268). See Auto Mode.
  • Two regressions from v2.1.265 hit gateway and proxy users: every turn failing with HTTP 400 on third-party ANTHROPIC_BASE_URL endpoints (fixed v2.1.268) and CLAUDE_CODE_USE_GATEWAY forcing Cloud-gateway sign-in (fixed v2.1.266).

Release table

VersionDate (UTC)Shape
v2.1.264—No GitHub release page
v2.1.2652026-09-08--plugin-dir folders, 1 GB tool-result cap, parallel worktree checkout
v2.1.2662026-09-08Single fix: CLAUDE_CODE_USE_GATEWAY regression from 2.1.265
v2.1.2672026-09-09maxEffortLevel, prompt-cache stability batch, --system-prompt-snapshot off
v2.1.2682026-09-10Todo-tool allowlist, WebFetch deadline, plugin --json, Claude Tag memory change
v2.1.2692026-09-11claude plugin eval, /output-style, Workflow concurrency setting
v2.1.2702026-09-12Single fix: read-only git commands asking permission (regression in 2.1.269)

In this batch, v2.1.267 is the first release carrying a [Claude Tag] section and v2.1.268 the first with a [Code Review] section.

From the official W37 digest

claude plugin eval in practice (v2.1.269)

  • It runs your plugin against a suite of test cases, scores the results, and by default runs each case again without the plugin so you can see what the plugin contributes.
  • claude plugin eval init, run from the plugin’s root directory, asks what a good result looks like, then proposes test cases and the checks that score them, tries the suite once, and writes the files.
  • When Claude says the suite is ready, exit the session init opened and run claude plugin eval . to score every case. The summary table prints in the terminal; report.html under evals/results/ holds the per-run detail.
  • Cost: every run, and every check that has a second model judge the reply, is a real model call on your account. With the default without-plugin baseline, each case runs twice.
  • Docs: code.claude.com/docs/en/plugin-evals.

Desktop pane pop-out

  • In the Claude Code Desktop app you can pop any pane out into its own window — for example, drag the diff or terminal to a second screen while Claude keeps working in the main window — and dock it back when done. Reference: code.claude.com/docs/en/desktop#arrange-your-workspace. The digest gives no version (Desktop app).

Models and defaults

  • maxEffortLevel setting (v2.1.267).
  • effort: frontmatter honoured on pinned-effort models (v2.1.267).
  • Task-tracking tools allowlist and CLAUDE_CODE_ENABLE_TODO_TOOLS=1 (v2.1.268).
  • Long-context 429s on Fable models showed the usage-credits consent prompt instead of the 1M-context message on Pro and Team — fixed (v2.1.268).
  • A running session could silently switch to the organization’s default model when another process refreshed a stale model-access entry — fixed (v2.1.268).
  • /model opusplan[1m] no longer rejected with “Model not found” (v2.1.265).
  • The advisor tool decision is made once and announced in the conversation when it changes, instead of per request (v2.1.265). See Advisor Tool.
  • The claude-api skill’s error-code reference corrected: model access failures return 404 and unavailable beta headers return 400, not 403 (v2.1.265).

New commands and features

  • claude plugin eval (v2.1.269).
  • /output-style [name] lists and switches output styles, including over Remote Control and in cloud and other headless sessions (v2.1.269).
  • bashEditDiffEnabled: the Bash tool result includes a diff of the files a Bash command changed (v2.1.269).
  • A spinner tip now suggests /focus, a view with just your prompt, a one-line work summary and the response (v2.1.269).
  • --system-prompt-snapshot off renders the system prompt fresh on every request, for iterating on prompt text (v2.1.267).
  • --plugin-dir can point at a folder of plugins: each immediate subfolder with a manifest loads, and children added or removed while running are picked up (v2.1.265).
  • claude plugin install/uninstall/update/enable/disable --json prints the result as one JSON object on the last line of stdout; claude plugin list --json adds errorDetails/noteDetails (v2.1.268).
  • 1 GB cap on tool results saved to disk; the preview says when a file was truncated (v2.1.265).
  • --worktree checks out in parallel on git 2.32+ (v2.1.265). See Worktrees.
  • Browser-tab icons for published artifacts, chosen by Claude to match each page (v2.1.268 in the GitHub release notes; the code.claude.com changelog snapshot of 2026-09-18 lists it under v2.1.269).^[ambiguous] See Artifacts.
  • Slash commands typed mid-prompt show a list of matches instead of a single suggestion — it opens as you type in fullscreen, and Tab opens it outside fullscreen — and a plugin skill is found by its bare name without the plugin prefix (v2.1.265).
  • VS Code agent map: click the “N agents” pill at the bottom of the prompt box to see the session’s subagents, open a read-only transcript, or stop one (v2.1.269).
  • VS Code Hooks and Permission rules dialogs in the command menu’s Customize section add or remove hooks and permission rules in user, project and local settings; managed, plugin and session entries stay read-only (v2.1.269). See Hooks.
  • Claude Code on the web: take back a queued message in a cloud session before Claude reads it — remove it from the queue, or press Esc or Up, and the text returns to the message box (v2.1.269).
  • /ultrareview --post posts the PR comment directly when findings arrive, instead of starting a second cloud session (v2.1.269).
  • /goal runs retry with backoff or pause with a reason after API errors, network drops or usage limits, instead of stalling silently (v2.1.269).
  • Claude apps gateway: pricing: in gateway.yaml flows to signed-in clients so /cost and telemetry match the spend meter; gatewayInternalNetworks managed setting (v2.1.268).
  • claude self-hosted-runner --remove-session-state (v2.1.268); claude auth status --json includes configDirectory (v2.1.268).

Settings and environment variables

  • CLAUDE_CODE_WEBFETCH_DEADLINE_MS (v2.1.268).
  • CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS (1–256) raises the Workflow tool’s per-run concurrent agent limit (v2.1.269). See Dynamic Workflows.
  • CLAUDE_CODE_GATEWAY_MODEL_DISCOVERY_TIMEOUT_MS extends the gateway /v1/models discovery timeout (default 3 s) (v2.1.269).
  • OTEL_METRICS_INCLUDE_REPOSITORY tags OpenTelemetry metrics and events with vcs.* attributes (v2.1.269).
  • CLAUDE_CODE_BG_TASKS_REPORT_RUNNING=0 restores the old reporting: remote and headless sessions had been saying “waiting for your input” while background agents still ran (v2.1.269).
  • CLAUDE_CODE_RESUME_INTERRUPTED_TURN no longer re-runs a turn that failed more than 6 hours earlier, or longer ago than CLAUDE_CODE_RESUME_INTERRUPTED_TURN_MAX_AGE_MS when set (v2.1.269).
  • forceLoginGatewayUrl managed setting makes machines Claude apps gateway sessions from startup (v2.1.265).
  • Managed allowedHttpHookUrls, httpHookAllowedEnvVars and allowedChannelPlugins admit nothing, not everything, when unreadable (v2.1.267).
  • Deny and ask rules starting with ! apply only within their own settings source; a bare ! is ignored (v2.1.269).

Hooks, MCP, plugins, skills and subagents

  • PermissionRequest hooks fire in --print mode (v2.1.268). See Hooks.
  • CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS now extends SessionEnd hooks that have no per-hook timeout; they had been cancelled after 1.5 seconds (v2.1.268).
  • --continue / --resume show the conversation immediately instead of waiting for SessionStart hooks (v2.1.268).
  • SubagentStart hook context and preloaded skills no longer move out of the prompt prefix on later turns (v2.1.265).
  • Forked skills (context: fork) stream their kickoff prompt and text turns in stream-json (v2.1.265).
  • Skills synced from claude.ai in cloud sessions are named anthropic-skills:<name>, matching Claude Desktop (v2.1.269).
  • MCP: http servers that only speak legacy HTTP+SSE fall back to SSE as the spec describes (v2.1.265); no OAuth client registration until you actually authenticate (v2.1.265); secrets from ${VAR} placeholders no longer shown in /mcp or claude mcp list/get (v2.1.268); “No available ports for OAuth redirect” fixed (v2.1.268). See MCP.
  • Plugin LSP servers that reject shutdown params (e.g. rust-analyzer) now get exit anyway (v2.1.269).
  • Organization plugins enabled through managed settings load in headless sessions and on Claude Desktop (v2.1.269).
  • A respawned in-process teammate no longer picks up tools or a system prompt from a same-named agent file in an untrusted folder (v2.1.268).
  • Workflow agent() calls with large output schemas are checked by the classifier in auto mode instead of refused (v2.1.267).

Breaking and behaviour changes

  • Plain WebFetch deny and ask rules no longer apply to Artifact tool reads and updates; use an Artifact rule or WebFetch(domain:claude.ai) (v2.1.268).
  • Bedrock, Vertex and Foundry sessions deliver environment, model and settings details as attachments and keep the tool list byte-stable, matching first-party sessions (v2.1.268).
  • The attribution reminder no longer overrides a CLAUDE.md or memory rule against commit/PR attribution; managed-settings lines still apply (v2.1.269).
  • Claude Tag public channels keep separate memory (v2.1.268).

Notable fixes

  • v2.1.265 regressions: HTTP 400 on every turn against third-party endpoints, from a regex in the Artifact tool’s input schema (fixed v2.1.268); CLAUDE_CODE_USE_GATEWAY forcing Cloud-gateway sign-in (fixed v2.1.266).
  • Prompt-cache stability: dozens of resume, MCP-reconnect and model-switch cache breaks (v2.1.267).
  • Edit() deny rules now cover the file a Bash tee writes, and Bash(tee:*) no longer covers destinations outside the working directories (v2.1.269); deny and ask rules on symlinked directories (/etc, /tmp, /var on macOS) apply by real path (v2.1.268).
  • /compact summaries mangling text containing $ sequences (v2.1.268).
  • Sessions permanently stuck on “Prompt is too long” when auto-compaction had nothing to summarize (v2.1.269).
  • Plugin archives readable by other local users, or keeping world-writable bits (v2.1.269).
  • F1/F2/F4 broken in kitty-protocol terminals and other key issues, a regression from v2.1.247 (v2.1.269).
  • v2.1.270: read-only git commands asking for permission after a session had run a while (regression in v2.1.269).

Try It

  • Plugin authors: add an eval suite and run claude plugin eval; see claude plugin eval --help for the format. Or start with claude plugin eval init and let Claude draft the cases; the without-plugin baseline shows whether the plugin helps.
  • On a multi-monitor setup, pop the Desktop diff or terminal pane onto the second screen.
  • If an unattended pipeline fetches slow or very large pages, set CLAUDE_CODE_WEBFETCH_DEADLINE_MS to change the 300-second deadline, or 0 to remove it.
  • If a newer model stopped keeping a task list, set CLAUDE_CODE_ENABLE_TODO_TOOLS=1.
  • Admins who want to cap spend per session: set maxEffortLevel, globally or per model under modelSettings.
  • Behind a gateway or proxy, skip v2.1.265 and take v2.1.268 or later.
  • If you gated Artifact reads with a WebFetch rule, add an Artifact rule instead.

Open Questions

  • v2.1.264 has no GitHub release page.
  • What format does a plugin eval suite take? The release note points only at claude plugin eval --help; the digest links a docs page (code.claude.com/docs/en/plugin-evals) that was not fetched.
  • Which model judges plugin eval checks? The digest says checks can have “a second model judge the reply” but doesn’t name the model.
  • Which release shipped artifact browser-tab icons — v2.1.268 (GitHub notes) or v2.1.269 (code.claude.com changelog snapshot)?
  • Why is the todo-tool list now an allowlist? Neither W33 nor v2.1.268 gives a rationale.