Claude docs changes for September 15th, 2026 [diff]

Executive Summary

  • New security behavior worth knowing about: in a remote MCP server's url and headers, Claude Code now reads credential variables as empty rather than expanding them — ANTHROPIC_API_KEY, AWS_BEARER_TOKEN_BEDROCK, HTTPS_PROXY, NPM_TOKEN and friends — so a project's .mcp.json or a plugin can't quietly forward your credentials to a server it names
  • Fast mode reaches cloud sessions and self-hosted runners (2.1.271), and modelPricing / the gateway pricing block now accept a multiplier above 1, up to 10, for marked-up internal chargeback rates
  • omitClaudeMd lands in agent frontmatter, --agents JSON and plugin agents, letting a subagent run without user, project and local CLAUDE.md files while managed policy files still load
  • claude plugin install/update gain --accept-command <sha256> as a tighter alternative to -y: the acceptance binds to exactly the command, plugin and catalog a previous --json run displayed, and any change since then makes Claude Code show the command again
  • Three weekly "What's new" pages land at once (weeks 35, 36 and 37), and the Compliance API gains Claude-in-Slack per-channel manager assignment events

New Claude Code versions

2.1.271

New features

  • Added fast mode in Claude Code Remote sessions, both cloud and self-hosted runners
  • Added mouse support to the /config panel in fullscreen mode: the wheel scrolls, a click on a value changes it
  • Added per-command allowed_domains to Bash, PowerShell and Monitor in auto mode with sandboxing: the hosts a command needs are reviewed with it and opened for that command alone
  • Added omitClaudeMd to agent frontmatter and --agents JSON
  • Added --accept-command <sha256> to claude plugin install and claude plugin update
  • Added support for a multiplier above 1, up to 10, in modelPricing and the gateway pricing block
  • Added claude self-hosted-runner --drain-marker-file <path> to report a SIGTERM exit as a host drain

Existing feature improvements

  • Improved Bash permission checking in four ways: catching the file fmt/column read after an unrecognized option, files a wildcard expands to inside a pattern or option value, shell variable declaration flags that could misrepresent the command, and commands with two directory changes or a cd+git chain that skipped the blockReadsOutsideWorkingDirectories prompt

Major bug fixes

  • Fixed a cached organization policy being reused after switching accounts, organizations or API keys
  • Fixed an enterprise managed-mcp.json that can't be read or parsed being ignored — it now keeps exclusive MCP control and warns at startup
  • Fixed a stale .git/config.lock breaking git checkout -b, git push -u and git config for the rest of a session after a sandboxed command failed to start on Linux
  • Fixed settings file changes made outside the session going unnoticed on macOS machines whose file-event service is saturated; the watcher now falls back to polling
  • Fixed /resume and /teleport keeping the previous conversation's file-read tracking, so Claude could edit files the resumed conversation had never read
  • Fixed MCP OAuth mishandling client registrations: denying consent forced a new one, one for another redirect URI was reused, and a concurrent write could delete a valid one
  • Fixed sustained high CPU usage and repeated tool-list requests when an MCP server sends list_changed notifications in a tight loop
  • Fixed self-hosted runner sessions silently losing all host config when the host config directory exceeds 64 MiB; added --host-config-snapshot disk|memory
  • Fixed skills synced from claude.ai staying on disk indefinitely after signing out
  • Fixed --resume dropping the 1M context window when the resumed session's model family differs from the configured default
  • Fixed cross-session messages held by the receiving session's policy leaving no trace
  • Fixed Claude starting a second copy of a background command that was still running after the conversation was compacted
  • Fixed turns failing with "API returned an empty or malformed response" when an LLM gateway returns the non-streaming reply as text/plain

2.1.272

  • Bug fixes and reliability improvements

Claude Code changes

New Documents

whats-new/2026-w35 [Source]

Week 35 (August 24–28, v2.1.240 → v2.1.250): resuming terminal sessions in the Claude Code Desktop app with /resume, reviewing feedback reports Claude drafts for you, and starting a session in restricted mode.

whats-new/2026-w36 [Source]

Week 36 (August 31 – September 4, v2.1.251 → v2.1.261): Claude Fable 5.1 in Claude Code with the fable alias now selecting it — though in Claude apps gateway sessions fable still selects Fable 5, so you need /model claude-fable-5-1 there. Plus computer use running in the background on Desktop and the live /diff panel.

whats-new/2026-w37 [Source]

Week 37 (September 7–11, v2.1.263 → v2.1.269): claude plugin eval, with the reminder that every run and every judge-scored check is a real model call on your account, and popping Claude Code Desktop panes out into their own windows.

Changed documents

mcp [Source]

  • New "Credential variables that read as empty" section. In a remote server's url and headers, Claude Code deliberately reads credential variables as empty rather than expanding them, so a project .mcp.json or plugin can't send your credentials to a server it names. Writing Bearer ${ANTHROPIC_AUTH_TOKEN} sends Bearer and the server usually answers 401. The covered names are Claude Code's own credentials, your cloud provider's, and others your environment carries such as HTTPS_PROXY and NPM_TOKEN. [lines 628-640] [Source]
  • Two sharp edges of that rule: a covered name reads as empty whether or not you set it, and a :-default fallback on it is ignored. A provider base URL such as ANTHROPIC_BASE_URL still expands, so "url": "${ANTHROPIC_BASE_URL}/mcp" works — unless the URL itself embeds a credential. To pass a covered credential deliberately, copy it into a variable of your own naming. [line 638] [Source]
  • The suppression is discoverable: run claude --debug-file /tmp/claude-debug.log and search for never expanded toward a remote server. [line 642] [Source]
  • A missing ${VAR} reference now warns in claude mcp list and /mcp naming the variable, and still loads the server with the text unexpanded. [line 282] [Source]

fast-mode [Source]

  • New "Use fast mode in cloud sessions" section: /fast on works in cloud sessions on Anthropic-managed infrastructure or a self-hosted runner, for that session only, and needs v2.1.271 in the session's environment. [lines 57-61] [Source]
  • Two distinct organization blocks documented, with their exact messages: fast mode not enabled for the org, and the fast-mode Opus model excluded by availableModels. In the latter case, a session already on an allowed Opus model that supports fast mode gets fast mode on its current model without a switch. [lines 123-126] [Source]
  • The skip variables affect only the client-side check — an API rejection stands regardless, and Claude Code then retries at standard speed and turns fast mode off. [line 159] [Source]

sub-agents [Source] / agent-sdk/subagents

  • New omitClaudeMd frontmatter field. It drops the user, project and local CLAUDE.md files for an agent running as a subagent — managed policy files still load — and is ignored when the agent runs as the main thread agent. Requires TypeScript Agent SDK v0.3.271; the Python SDK's AgentDefinition doesn't have the field. [line 150] [Source]

claude-apps-gateway-config [Source]

  • New "Mark prices up" section for a multiplier above 1. Two consequences are called out: with an admin: block the markup also applies to spend limits, so developers reach their caps sooner and the gateway logs a boot warning saying so; and the multiplier doesn't change what the upstream provider charges. A gateway server earlier than v2.1.271 refuses to start with a markup set, and clients earlier than v2.1.271 ignore it and show costs without it. [lines 467-482] [Source]

permission-modes [Source]

  • A significant clarification that runs through the whole page: in interactive terminal sessions where bypass permissions are available, plan mode's blocks aren't enforced. Claude is still instructed to plan without editing, but a file edit or shell command it attempts during planning runs without prompting — only explicit ask rules and critical-path removals still prompt. [line 514] [Source]
  • Plan mode keeps its blocks everywhere without an interactive terminal: -p runs, Agent SDK sessions, and the VS Code extension's chat panel. [line 516] [Source]
  • Protected-path writes gained the same carve-out, and the planning-command table row now covers all three cases. [lines 24-551] [Source]

plugins-reference [Source]

  • --accept-command <sha256> documented at length: the acceptance counts for exactly that command, plugin and marketplace catalog, and if any of them changed since the command was displayed — including through the run's own marketplace refresh — Claude Code doesn't accept the digest and shows the command again. It can't be combined with -y, and has no effect inside a Claude Code session. [lines 1000-1017] [Source]
  • A failed result now carries a shownCommand object with the command as displayed, its plugin, and its sha256; acceptCommandMatched: false means the digest you passed doesn't match what's now displayed, so show that command to a person first. [lines 1015-1017] [Source]
  • Synced plugins: claude plugin disable <name>@synced saves the choice in that environment's user-level enabledPlugins, but a plugin your organization requires can't be turned off this way and the command saves nothing. [lines 413-414] [Source]
  • Plugin agents support omitClaudeMd, and for security still don't support hooks, mcpServers or permissionMode. [lines 62-64] [Source]

prompt-caching [Source]

  • A precise account of when adding a bare tool deny rule costs you the cache: with tool search active (the default on supported models) the tool definitions don't change and the cached prefix survives; without it Claude Code removes the definition and invalidates the cache, and so does removing the rule later. Only a rule matching in the tool-name position does this — scoped rules like Bash(rm *), and all allow and ask rules, leave the prefix intact. [lines 154-158] [Source]
  • /model asks you to confirm a switch only while the cache is warm and the new model isn't the one that produced the last response. [line 77] [Source]

managed-mcp [Source]

  • Spells out exactly what a deployed managed-mcp.json permits: the file's servers, managedMcpServers from managed settings, and in-process servers the host app registers. Everything else — plugin servers and --mcp-config included — is excluded, and claude.ai connectors are suppressed unless explicitly allowed. [lines 39-45] [Source]
  • Verification guidance distinguishing two failure signatures: user servers still appearing means the file isn't being read, while the file's servers missing plus a failed-to-parse diagnostic means it can't be parsed. [lines 116-118] [Source]

monitoring-usage [Source]

  • New event.sequence attribute added to every event, with an important caveat: it's counted per Claude Code process, not per session, and a resumed session takes its values from the resuming process — so within one session a later event can carry a lower value than an earlier one, or repeat one. Sort by event.timestamp and use event.sequence only to break ties. [lines 620-626] [Source]

ide-integrations [Source]

  • New prompt cache clock documentation: it counts down the conversation's cache lifetime and restarts on each response that uses the cache. Apart from compaction, the actions that invalidate the cache don't reset the clock, so it can still show minutes left after you switch models. It turns red without minutes right after a compaction too. [lines 112-115] [Source]
  • New attachOpenFile setting (default on) for including the editor's open file in messages; when off, only selected text is added. A Read deny rule on a path prevents both the selection and the open-file notice for that file. Requires v2.1.271. [lines 135-391] [Source]
  • The usage dialog now flags behaviors accounting for 10% or more of recent usage — cache misses, long context, subagent-heavy or highly parallel sessions — each with a tip, plus attribution tables per skill, subagent, plugin and MCP server. [line 179] [Source]

cross-session-messaging [Source]

  • When no crossSessionInbound value applies, Claude Code decides per message from the two sessions' permission modes, grouping bypass-style sessions into one class and everything else into the other. Plan mode counts as bypassing in interactive terminal sessions with bypass available, while auto, acceptEdits and dontAsk count as prompting. [lines 215-217] [Source]
  • Senders now get a notice when a receiver holds a message, and a follow-up when it's delivered, denied or expired, so the sending Claude knows not to keep waiting. claude -p senders receive it as an informational system message; that requires v2.1.271. [lines 231-235] [Source]

self-hosted-environments-reference [Source]

  • New --host-config-snapshot <mode>. The default disk verifies every file against an in-memory digest at each session start and refuses sessions until you restart the runner if a file in the copy was modified; memory caps the snapshot at 64 MiB and starts sessions without host config above it. Requires v2.1.271. [line 35] [Source]
  • New --drain-marker-file <path>, telemetry-only, with the caution to name a path on a local filesystem that sessions can't write to. [line 26] [Source]

settings-reference [Source]

  • modelPricing.multiplier now accepts up to 10, with markup requiring v2.1.271; earlier versions ignore a markup with a warning and keep the rest of the setting. [lines 1090-1102] [Source]
  • Parent settings from an embedding host keep their deny and ask rules except Read and Edit rules whose pattern starts with ! — a host can't carve paths out of the managed rules with a negation. [line 1269] [Source]
  • Deny-rule coverage refined again: tee joins the recognized Bash file commands, and the exclusions now name a command that reads files without naming them, such as grep -r pattern .. [line 1440] [Source]
  • New footer badge setting rendering clickable badges when a regex matches turn output — tool results, file contents, fetched pages and Claude's own replies — turning IDs like PROJ-1234 into links. [lines 3036-3057] [Source]
  • The workflow size guideline is advice, not an enforced cap: "small" asks for fewer than 5 agents, "medium" fewer than 10, "large" fewer than 50. [line 3925] [Source]

errors [Source]

  • CLAUDE_CODE_RETRY_WATCHDOG now fails at once on a standard-speed 429 reporting a spend limit or exhausted usage credits, including a gateway spend cap that resets on a schedule. Before v2.1.239 it retried those indefinitely. [line 316] [Source]
  • Distinguishes Login expired — which Claude Code produces itself for a login it already failed to renew, sending no request — from the API-returned OAuth rejections. [line 1047] [Source]

hooks [Source]

  • A prompt sent while SessionStart hooks are still running doesn't reach Claude until they finish, including on a resume — press Esc to take it back into the input while the hooks keep running. [lines 1108-1110] [Source]
  • New resolvedModel field naming the model a subagent actually started on, which can differ from the requested one when availableModels or another override applies. [lines 1703-1712] [Source]

artifacts [Source]

  • Source file types tightened: .html, .htm or .md decoding as UTF-8 or BOM-marked little-endian UTF-16. A file that doesn't decode, or that contains U+FFFD, is refused with the line and column to fix. [line 280] [Source]

amazon-bedrock [Source]

  • Credential-helper caching documented: with a valid ISO 8601 Expiration, Claude Code caches until five minutes before it; without one, for an hour. [line 210] [Source]

remote-control [Source]

  • The three "session changed elsewhere" failure reasons are now spelled out individually — taken over by another connection, ended or archived from another device, or no longer reported by the server — each with whether reconnecting is what you want. [lines 123-127] [Source]

large-codebases [Source]

  • A useful warning for monorepos: names always load, but when there are many skills some lose their descriptions entirely, stripping the keywords Claude uses to decide whether a skill applies. Keep descriptions short and lead with words a request would contain. [lines 350-358] [Source]

API changes

Changed documents

api/compliance [Source]

  • Two new activity types, ccr_channel_manager_added and ccr_channel_manager_removed, recording an owner or admin assigning an organization member to manage one Slack channel's Claude configuration. The activity-type enum now reads "or 492 more". [lines 192-198] [Source]
  • A new via_account_assignment authorization leg on Claude Tag events, distinguishing a direct channel-manager assignment from a role-granted permission. When it's true, via_full_manage and via_entitlement_leg are false and granting_role_ids is empty — and the field is absent on events recorded before direct assignments existed, which you should treat as false. [line 10501] [Source]
  • A move_thread comment action was added to the comment-action enum. [line 7450] [Source]

api/beta [Source]

  • temperature is now marked Deprecated: models released after Claude Opus 4.6 don't support setting it, and while 1.0 is accepted for backwards compatibility, every other value is rejected with a 400. [line 5048] [Source]
  • A new terminal iteration entry marks a fallback-served turn: when a fallback hop's output is the returned message, the completing iteration carries this type in place of message. Its presence in usage.iterations is how you tell a fallback model served the response. [lines 6637-6642] [Source]
  • Subagent thread routing fields added across the streaming event types: a server-set field naming the subagent thread a confirmation or result was routed to, and a cross-post marker for events surfaced from a subagent's thread onto the primary thread's stream. The cross-post marker is informational only — the server routes the matching response by tool_use_id or custom_tool_use_id, so clients don't send it back. [lines 36152-36490] [Source]