Released: 2026-08-27
A long, complex Claude session could lose its extended-thinking context when tool definitions were rendered again after an OAuth token refresh.
The roughly hourly prompt-cache miss is fixed. ScheduleWakeup also keeps the same definition on the first --resume turn after an account enters usage overage, avoiding a full cache miss.
Run claude --restricted, or CLAUDE_CODE_RESTRICTED=1 claude, when Claude Code should work within a narrower boundary. It removes built-in tools that execute commands or code and removes WebFetch, unless you explicitly name a tool in --tools; this is not a promise of no network access. File tools stay inside the working directory, bypassPermissions is refused, and user, project, and local settings files are ignored.
An agent can now carry its own prompt-cache lifetime in its frontmatter.
Set experimental.cacheTtl to "5m" or "1h". This per-agent value is used only when no subagent TTL setting has been configured. Copy this frontmatter example:
---
name: reviewer
experimental:
cacheTtl: "1h"
---A row of self-hosted runners labeled only by hostname can be hard to identify. Override the registration label with --client-label; hostname remains the default. Run:
claude self-hosted-runner --client-label ci-box-3Or set the environment variable:
export SELF_HOSTED_RUNNER_CLIENT_LABEL=ci-box-3Both forms override the registration label.
A centrally managed setting could fail to load without giving users a useful starting point.
Startup shows a warning, while /doctor and /status add a line explaining the load failure or why settings were not fetched. The latter covers Bedrock, Vertex, third-party providers, and a custom ANTHROPIC_BASE_URL. It is diagnosis, not an automatic repair.
Changing the managed environment variables for client-side timeout, MCP startup-mode, or stream-watchdog no longer triggers a settings-approval prompt. The changelog does not provide their exact variable names, so there is no new name to copy from this card. The exemption is limited to those categories.
During /web-setup, Claude Code now warns if the GitHub CLI token lacks the workflow scope. Without that scope, a push to a very large repository can be rejected. This is a warning about a possible failure, not a claim that every push will fail, and /web-setup does not grant the scope for you.
Members of eligible Enterprise organizations can now run /usage-credits to ask an administrator for a higher usage limit. Eligibility covers organizations billed through AWS Marketplace, self-serve Enterprise, and Enterprise trials. Rate-limit, usage, and fast-mode messages also stop recommending the command when the organization does not offer it, including when DISABLE_EXTRA_USAGE_COMMAND hides it.
Sessions on the same machine can now use SendMessage and ListAgents on Bedrock, Vertex, Foundry, and when telemetry is disabled.
If the default /tmp cannot be used, messaging falls back to a private per-user /tmp directory; the notice and /status name it. Mentions also match non-Latin session names, including Korean entered through an IME.
An invalid crossSessionInbound value is no longer ignored. The key lives in ~/.claude/settings.json; the changelog lists no valid values, so choose one from the documentation. User settings warn and hold cross-session messages, while managed settings warn and refuse them, in both cases until fixed. In Linux user namespaces, root-equivalent trust for unmapped owners is limited to canonical system directories. A reply to a subagent's SendMessage goes to the parent session conversation.
After a machine wakes, you need to know where background work really stopped.
A new session no longer opens an older conversation or drops the prompt you typed. Opening a stopped session already resumed in another terminal does not start a second process; its claude agents row says it is open in a terminal.
On Windows, the claude agents list responds to the keyboard again after detaching from a session or when launched in a terminal tab left in win32-input-mode. The dispatch input also gives its keys distinct jobs: shift+enter inserts a newline, while ctrl+enter dispatches and attaches.
When claude agents opens, setting the CI environment variable no longer skips the workspace trust prompt. A malformed entry in the PR-status cache also no longer crashes launch.
A background session no longer waits silently when a PermissionRequest or PreToolUse hook prints an invalid answer. Its claude agents row names the hook and schema error. If hook stdout prints a {…} object that is not valid JSON, it is no longer accepted as plain text; Claude Code reports a hook error with the parse message.
If a worktree branch is already merged into the checked-out default branch, such as local main, claude agents and claude rm no longer refuse deletion merely because that merge has not been pushed. Meanwhile, a running background session holds its worktree lock. Cleanup and git worktree remove leave it alone instead of taking away its checkout.
If another Claude Code process holds the token-refresh lock just as the current session token expires, you are no longer sent to the login screen.
The request fails with a retryable error. It is still a failed request, not a promise of automatic success; the temporary lock conflict is simply no longer mistaken for a lost login.
Some machines cannot use the recommended Console sign-in because ANTHROPIC_API_KEY or an API-key helper is set. /login used to stop with an OAuth error before even showing the sign-in URL. It now falls back to API-key sign-in and uses the credential path already available. This is a login-method fallback, separate from retrying token-refresh contention.
/login no longer hangs when connecting to a Claude apps gateway requires the managed-settings security-approval dialog. Gateway model discovery also runs when apiKeyHelper is the only credential source. The environment variable named in the changelog is:
CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERYMCP entries now appear under their real scope. A project .mcp.json entry declaring the claude.ai connector type no longer appears under the trusted “claude.ai” heading. That type declaration does not change its classification.
When an MCP server’s headersHelper supplies the Authorization header and gets a 401, Claude Code now reruns the helper and retries the call as documented instead of entering OAuth discovery.
A session you opened in Claude Desktop or Cowork keeps its transcript as long as the app still holds it; routine cleanup no longer deletes it after 30 days. An organization retention rule, if you have one, still wins. To put a limit on that keep-alive, add one line to ~/.claude/settings.json and change the number to the days you want (30 is the example):
{"desktopSessionCleanupPeriodDays": 30}After claude logs exits, its terminal should no longer behave as if the log viewer were still open. Claude Code now switches off mouse tracking, bracketed paste, and the alternate screen in the terminal where the command ran. The fix concerns modes enabled by claude logs; it does not claim to replace other terminal preferences.
Several terminal glitches had one theme: text should remain literal and columns should align.
Model names in /model and fast-mode notices render as code, so a suffix such as [1m] stays literal instead of becoming a link. Long repository-permission rules no longer show a garbled character when truncated inside an emoji. Startup warnings also align with the rest of the transcript instead of shifting one column right.
Pressing shift+tab immediately after ctrl+c used to leave the permission-mode indicator hidden behind “Press Ctrl-C again to exit.” The indicator now remains visible, so you can see the selected mode while the exit hint is present. The fix changes this display overlap, not what ctrl+c or shift+tab does.
/ultrareview and locally seeded cloud sessions no longer upload uncommitted edits hidden in swap, temporary, or backup copies of sensitive files. This covers prod.env-style files, *.tfvars, and credential-file copies such as key.pem.tmp and id_rsa.swo; they stay on your machine. The statement is specific to these files, not every uncommitted edit.
Before /ultrareview <PR#> launches, Claude Code checks whether the GitHub account connected to your Claude account can access the repository. If not, it explains how to fix the connection before creating the cloud session instead of failing afterward. The check concerns the linked account's repository access.
After the CLI silently reconnected a Remote Control session, a connected device could remain on stale state: no permission prompt and no latest messages.
That occasional catch-up failure is fixed. The scope is the missing prompt and messages after silent CLI reconnection, not every possible Remote Control connection problem.
A wrapper may inject an option before the claude remote-control subcommand, and users may place a global flag there too. That prefix used to make remote-control reject its own flags, including --spawn and --name. The parser now accepts those flags despite the preceding options.
A cloud session could occasionally fail at startup because its container's session credentials were not readable yet. Startup now handles that transition instead of treating “not readable yet” as a final failure. The changelog gives no wait duration or retry count, and does not say invalid credentials are accepted.
A VS Code chat tab could become stuck on “No conversation found” when its session had never been saved, meaning there was no old conversation to load. It now starts a new conversation so you can type. This does not recover missing history; it handles the specific never-saved state.
The Workflow tool carries its description in the prompt whenever it is available.
The reference material was moved, not removed. This change reduces prompt footprint; the changelog makes no numerical promise about execution speed or output quality.
The prompt-footer PR badge now checks GitHub less often while its pull request remains unchanged. A push or any gh pr command still refreshes it immediately. Automatic checking has not disappeared; explicit actions that can change the PR remain immediate triggers.
/loop no longer loses an operating style because of the model provider.
Self-paced dynamic mode and the no-prompt autonomous default are always available, including on Bedrock, Vertex, and Foundry. This expands existing modes; it does not add another mode or declare one faster.
[3P telemetry] OTEL diag error could make operators suspect their own OTel collector even when Anthropic's telemetry export had failed. These failures now log at debug level as [Anthropic telemetry], identifying the source and lowering the apparent severity. The export failure itself is not claimed to be fixed; the message no longer blames the user's collector.
Fixed a prompt-cache miss (and lost extended-thinking context) roughly once an hour in long sessions, caused by tool definitions being re-rendered after an OAuth token refresh
Fixed the `ScheduleWakeup` tool definition changing between a session and its `--resume` when the account had entered usage overage, causing a full prompt-cache miss on the resumed session's first turnAdded `--restricted` (or `CLAUDE_CODE_RESTRICTED=1`): removes the built-in tools that run commands or code and `WebFetch` (unless named in `--tools`), keeps file tools inside the working directory, refuses `bypassPermissions`, and ignores user, project and local settings filesAdded `experimental.cacheTtl` (`"5m"` or `"1h"`) to agent frontmatter: a per-agent prompt cache TTL used when no subagent TTL setting is configuredAdded `claude self-hosted-runner --client-label <label>` (or `SELF_HOSTED_RUNNER_CLIENT_LABEL`) to override the label the runner registers with (default: hostname)Added server-managed settings diagnostics: a startup warning when the settings fail to load, and a `/doctor` and `/status` line explaining a load failure or why they weren't fetched (Bedrock/Vertex/third-party provider, custom `ANTHROPIC_BASE_URL`)Improved managed settings: client-side timeout, MCP startup-mode, and stream-watchdog env vars no longer trigger the settings-approval promptAdded a warning in `/web-setup` when the GitHub CLI token lacks the `workflow` scope, since pushes to very large repositories can be rejected without itAdded `/usage-credits` for Enterprise organizations billed through AWS Marketplace, self-serve Enterprise, and Enterprise trials, so members can request a higher usage limit from their admin
Fixed rate-limit, usage, and fast-mode messages telling you to run `/usage-credits` when that command isn't available for your organization (e.g. hidden with `DISABLE_EXTRA_USAGE_COMMAND`)Added cross-session messaging (`SendMessage` / `ListAgents`) between sessions on the same machine on Bedrock, Vertex, and Foundry, and when telemetry is disabled
Improved cross-session messaging: falls back to a private per-user `/tmp` directory when the default one can't be used, and the notice and `/status` name the directory to fix
Fixed @-mentions of other sessions not matching names typed with non-Latin characters (for example Korean entered through an IME)Fixed an invalid `crossSessionInbound` value being silently ignored: it now warns and holds cross-session messages (user settings) or refuses them (managed settings) until fixed
Changed cross-session messaging in Linux user namespaces: root-equivalent trust for unmapped owners is limited to canonical system directories
Changed `SendMessage` from a subagent to another session: the result now notes that any reply is delivered to the parent session's conversation, not to the subagentFixed agent view resurrecting a weeks-old background session after the machine was off: such a session now shows as stopped at its real end, and opening it asks before resuming its saved conversation
Fixed agent view sometimes opening an older conversation, and dropping the typed prompt, when starting a new session
Fixed `claude agents`: opening a stopped session that you already resumed in another terminal no longer starts a second process on that conversation; the row now says it is open in a terminalWindows: Fixed the `claude agents` list not responding to the keyboard after detaching from a session, or when launched in a terminal tab left in win32-input-mode
Changed shift+enter in the agent view dispatch input to insert a newline (matching the prompt); ctrl+enter now dispatches and attachesFixed `claude agents` skipping the workspace trust prompt when the `CI` environment variable is set
Fixed `claude agents` crashing on launch when the PR-status cache held a malformed entryFixed background sessions waiting silently when a `PermissionRequest` or `PreToolUse` hook prints an invalid answer: the `claude agents` row now names the hook and the schema error
Fixed hooks silently treating a stdout `{…}` object that isn't valid JSON as plain text; it's now reported as a hook error with the parse messageFixed `claude agents` and `claude rm` refusing to delete a session ("has commits that are not pushed anywhere") when its worktree branch was already merged into your checked-out default branch (e.g. local `main`) but not yet pushed
Fixed a backgrounded worktree session losing its checkout: the background session now holds the worktree's lock while it runs, so cleanup and `git worktree remove` leave it aloneFixed being sent to the login screen when another Claude Code process held the token refresh lock while the session token had expired; the request now fails with a retryable error insteadFixed the recommended Console sign-in in `/login` failing with an OAuth error before showing a sign-in URL on machines where it can't be used (for example when `ANTHROPIC_API_KEY` or an API key helper is set); it now falls back to the API-key sign-inFixed `/login` to a Claude apps gateway hanging when the managed-settings security approval dialog was required
Fixed gateway model discovery (`CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY`) never running when `apiKeyHelper` is the only credentialFixed `/mcp` listing a project `.mcp.json` entry that declares the claude.ai connector type under the trusted "claude.ai" heading; it now appears under its real scopeFixed MCP servers whose `headersHelper` supplies the `Authorization` header falling into OAuth discovery on a 401 instead of re-running the helper and retrying the call as documentedFixed Claude Desktop and Cowork sessions disappearing after 30 days: the transcript cleanup now keeps desktop-written sessions while they are in the app (unless org policy manages retention); the new `desktopSessionCleanupPeriodDays` setting caps the exemptionFixed `claude logs` leaving mouse tracking, bracketed paste and the alternate screen switched on in the terminal it was run fromFixed model names in `/model` and fast-mode switch notices to render as code, so suffixes like `[1m]` display literally instead of as a link
Fixed the trust dialog's list of repo permission rules showing a garbled character when a long rule was cut off in the middle of an emoji
Fixed startup warnings (e.g. "N MCP servers need authentication") rendering one column right of the rest of the transcriptFixed the permission mode indicator staying hidden behind the "Press Ctrl-C again to exit" hint when you press shift+tab right after ctrl+cFixed `/ultrareview` and locally seeded cloud sessions uploading uncommitted edits to `prod.env`-style and `*.tfvars` files, or to editor swap, temp, and backup copies of credential files (e.g. `key.pem.tmp`, `id_rsa.swo`); they now stay on your machineImproved `/ultrareview <PR#>` to check before launch that the GitHub account connected to your Claude account can access the repository, and to explain how to fix it, instead of failing after the cloud session startsFixed Remote Control sessions occasionally never showing a permission prompt or the latest messages on the connected device after the CLI silently reconnectedFixed `claude remote-control` rejecting its own flags (e.g. `--spawn`, `--name`) when a global flag or a wrapper-injected option precedes the subcommandFixed cloud sessions occasionally failing at startup when the container's session credentials were not yet readable[VSCode] Fixed a chat tab getting stuck on "No conversation found" when its session was never saved; it now starts a new conversation insteadImproved the Workflow tool's prompt footprint: its description is now about 1k tokens instead of 5.7k, with the script-writing reference moved into a bundled `workflow-authoring` skillImproved the prompt-footer PR badge to check GitHub less often while the pull request is unchanged; a push or a `gh pr` command still refreshes it right awayChanged `/loop`: self-paced dynamic mode and the no-prompt autonomous default are now always available, including on Bedrock/Vertex/FoundryChanged Anthropic telemetry export failures to log at debug level as `[Anthropic telemetry]` instead of `[3P telemetry] OTEL diag error`, so they are not mistaken for your OTel collector failing