Source: raw/anthropic-watch-claude-code-tag-v2-1-317.md, raw/anthropic-watch-claude-code-tag-v2-1-318.md and raw/anthropic-watch-claude-code-tag-v2-1-319.md (GitHub releases of anthropics/claude-code, fetched 2026-11-05 and 2026-11-09). The v2.1.319 watch stub was cut at 20,000 characters, so the complete notes come from ai-research/claude-code-release-v2-1-319.md (gh release view, 2026-11-09). Continues Week 44, which ended with v2.1.316 (October 31 UTC). Merged 2026-11-11 with the scheduled weekly sweep’s digest (2026-11-07), which added the changelog snapshot ai-research/claude-code-docs-changelog-2026-11-07.md, the message-threads post and two @ClaudeDevs announcements; their unique items are in “Announced alongside the releases” below.

Product: Claude Code · URL: https://github.com/anthropics/claude-code/releases · Date: 2026-11-03 to 2026-11-06 (UTC)

Three releases, each with 60 to 150 entries. The changes most likely to affect a working setup: auto mode no longer falls back to the built-in classifier when the server-side check can’t run (telemetry or feature flags off, or behind an LLM gateway), so those calls are now asked about or denied (v2.1.319); self-paced /loop now ends when a turn finishes without scheduling the next wakeup (v2.1.319); and subagents on a different model now use that model’s own effort level, not the session’s (v2.1.317). v2.1.319 also closes four permission or sandbox holes, including a Bash permission bypass and a sandbox escape when creating a worktree.

Key Takeaways

  • Auto mode without the server check now asks or denies (v2.1.319). With telemetry or feature flags off, including behind LLM gateways, calls the server can’t check are “asked about or denied, with no fallback to the built-in classifier”. Set CLAUDE_CODE_AUTO_MODE_SERVER=0 to keep the built-in classifier. Bedrock, Vertex and Foundry are unchanged. CLAUDE_CODE_AUTO_MODE_SERVER=1 turns the server check on behind a gateway even with CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS set.

  • Self-paced /loop stops when a turn ends without rescheduling (v2.1.319), instead of scheduling a fallback wakeup. CLAUDE_CODE_LOOP_KEEPALIVE=1 restores the old behavior. In cloud sessions, /loop now warns that a timer stops with its container.

  • Four security fixes in v2.1.319: a Bash permission bypass where “a crafted environment prefix could hide part of a command from permission checks”; a sed command that could skip the permission check in accept-edits and auto mode; “a sandbox escape when creating a worktree”; and /rewind restoring or deleting a file through a directory symlink. Bash deny and ask rules also missed commands written with quoted words.

  • permissions.denyInsteadOfPrompting (v2.1.319): a tool call that would wait on a permission prompt is denied and the turn continues, in terminal sessions, subagents and teammates. Useful for unattended runs that should never block on a prompt.

  • MCP tasks run in the background (v2.1.319): when an MCP server runs a tool call as a task, Claude Code backgrounds it immediately and tells Claude the result when it finishes. CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND makes it wait inline; CLAUDE_CODE_DISABLE_MCP_TASKS=1 turns tasks off.

  • Subagent effort follows the subagent’s model (v2.1.317): subagents and teammates on a different model from the main conversation use that model’s saved effort level or its default, not the session’s. Week 44’s CLAUDE_CODE_SUBAGENT_EFFORT still sets an explicit level.

  • Flagged subagents stop instead of switching models when “Switch models when a message is flagged” is set to “Ask each time” (v2.1.318). After safeguards stop a subagent, Claude is now told to retry with a narrower prompt or tell you, “instead of doing the work itself” (v2.1.319).

  • The Read tool’s “file unchanged” stub is gone (v2.1.317): reading an unchanged file again returns its content.

  • Sign-in expiry is visible (v2.1.318): claude auth status shows loginExpiresAt, claude -p JSON results carry credential_expires_at, and claude -p text output warns on stderr starting 10 days before a stored claude.ai sign-in ends. v2.1.319 adds a warning when Claude Code falls back to an ant auth login sign-in, which bills your Claude Console organization.

  • /claude-api latency-optimize (v2.1.319) finds where your Claude API code makes people wait and tries faster settings against your eval. The cookbook gained a matching latency optimization notebook the same week (see SDK releases, November 2026).

  • Claude apps gateway cost data (v2.1.317): GET /v1/organizations/analytics/cost reports past spend over a date range, in total, per developer, per day or per IdP group. v2.1.319 adds pooled-budget priority rules and today/week/month spend in /usage for developers with no spend limit.

  • Background subagent status updates cost about 99% less (@ClaudeDevs, 2026-11-04): each 30-second status-line update used to re-read the whole conversation; “now Haiku 5.5 writes it from short notes instead”. Sessions with background subagents “used about 5% less of the limit, so the same limit gets you about 6% more prompts.”

Release table

VersionDate (UTC)Headline changes
v2.1.3172026-11-03Subagent effort follows its model; Read “unchanged” stub removed; email-address network check; Agent SDK hook onFailure: "block"; gateway cost report API; OTel plan/seat/provider attributes; claude -p exits 1 on lost writes
v2.1.3182026-11-04Sign-in expiry in auth status and -p JSON; flagged subagents stop on “Ask each time”; background subagents background shell commands after 3 min; no IDE-extension auto-install; deniedMcpServers URL entries cover connectors; --continue regression fix; network home directory robustness
v2.1.3192026-11-06Auto mode no-fallback change; self-paced /loop ends; four permission/sandbox fixes; permissions.denyInsteadOfPrompting; MCP tasks; /claude-api latency-optimize; CLAUDE_CODE_DISABLE_FEATURE_FLAGS; /tasks “Earlier in this session”; gateway pooled-budget rules

Auto mode and permissions

  • No classifier fallback (v2.1.319), as above. /status no longer shows “Auto mode server: Enabled” when the server-side review isn’t reaching the session, for example behind a gateway that drops it.
  • Project files can’t turn the server check back on: once you switch it off outside the project, a project’s settings files or a plugin can no longer switch it on again (v2.1.319).
  • claude -p in auto mode exits 1 when its actions were refused because no safety check was available (v2.1.319); it used to exit 0.
  • Large rule sets: auto mode used to refuse actions without saying why when permission rules or autoMode settings were very large. It now explains and, unless autoMode deny rules are set, asks you in the terminal (v2.1.319).
  • Rule precedence tightened (v2.1.319): a session’s exact Bash or PowerShell ask rule is refused, not asked about, when a broader deny rule covers the command. Bash now asks before running which with flags outside a lookup/display set.
  • Fewer needless prompts: read-only commands wrapped in a plain timeout DURATION (for example timeout 30 rg foo) no longer prompt (v2.1.319).
  • Email address check (v2.1.317): Claude Code asks before a web or shell tool call sends your account email address to the network.
  • Connector names in Desktop rules (v2.1.319): deny and ask rules, hook matchers and disallowedTools entries that use a claude.ai connector’s terminal name (mcp__claude_ai_Slack) were ignored in Claude Desktop; fixed once Desktop includes this version.
  • deniedMcpServers URL entries (v2.1.318) now also block Anthropic-provided servers whose upstream URL matches, and connectors in cloud sessions and on self-hosted runners. Otherwise deny a connector by name.
  • Monitor WebSocket watch follows Bash rules (v2.1.319): a deny refuses it, and under an ask rule an allow on Monitor, bypass mode or auto mode no longer skips the prompt.
  • Trust boundaries (v2.1.319): --add-dir directories register plugin marketplaces only after that directory’s trust dialog is accepted; with the Bash sandbox on, skills in .claude/skills folders below the starting folder load only if you start Claude Code there or /cd there; CLAUDE_CODE_PROXY_RESOLVES_HOSTS can no longer be set by settings files where the host manages the network.

Subagents, background work and /loop

  • SendMessage to a running background agent arrives earlier (v2.1.317): its running command moves to the background without being stopped.
  • Background subagents background a shell command after 3 minutes on all providers, so the prompt cache doesn’t expire while it runs (v2.1.318).
  • /tasks lists finished subagents under “Earlier in this session”, so you can reopen one and message it after it leaves the list or you restart (v2.1.319).
  • Fixes: Ctrl+C at the prompt no longer stops all running background agents (v2.1.317); workflows resumed after a usage limit no longer re-run finished agents (v2.1.318); /clear no longer kills running commands of background agents, teammates and workflows that keep running (v2.1.319); a background agent’s refused late report is now reported instead of lost silently (v2.1.319).
  • /loop: v2.1.317 fixed self-paced wakeups sometimes not being scheduled when Claude left a field out of its request; v2.1.319 changes self-paced loops to end when a turn finishes without rescheduling (CLAUDE_CODE_LOOP_KEEPALIVE=1 to opt out).

MCP

  • MCP tasks (v2.1.319) as above; they work without feature-flag fetching (Bedrock, Vertex, Foundry, LLM gateways, DISABLE_TELEMETRY).
  • Older-handshake servers: v2.1.319 fixed servers on the older handshake failing to connect or returning empty results, because Claude Code declared an extension it only uses with newer servers. v2.1.318 stopped re-probing stdio servers that exit on the version probe on every start; a server that dies in under 3 seconds for its own reason is pinned to the legacy handshake for 7 days.
  • Tool calls before load now return the tool’s input schema in the error (v2.1.318).
  • claude mcp add and remove now wait for the config lock or exit 1 (v2.1.318).

Headless, SDK and sign-in

  • claude -p exit codes: exits 1 when a full disk or quota lost its transcript or ~/.claude.json writes (v2.1.317), and in the auto-mode case above (v2.1.319).
  • Agent SDK hook callbacks take onFailure: "block", so a failing callback blocks the action instead of letting it run (v2.1.317).
  • Nested sessions: claude -p run inside a cloud or self-hosted runner session no longer shares that session’s ID and transcript (v2.1.317); CLAUDE_CODE_DISABLE_NESTED_SESSION_ID=1 restores the old behavior on a self-hosted runner.
  • ~/.claude.json safety (v2.1.318–319): non-interactive launches no longer replace an invalid ~/.claude.json with defaults (the start fails and says how to restore it, #87551); saves on shared network home directories and full disks are less likely to empty or truncate the file; unattended runs stop rewriting it when nothing changed; a home holding two config files now gets a stderr notice.
  • Fixes: --continue losing the earlier conversation after a malformed tool call’s retries ran out (regression in 2.1.314, fixed in v2.1.318); idle compaction replacing a conversation with a reply that was not a summary (v2.1.318); “Prompt is too long” after automatic compaction when a very large CLAUDE.md changed (v2.1.317).
  • v2.1.318 removed the 2.1.288 fix that kept the model’s earlier thinking when resuming a conversation started on 2.1.286 or earlier; that thinking can now be dropped.

Mods

  • v2.1.318 records your answer when Claude Code asks to turn on the mods it writes in a session, adds an effort option to a mod’s $.model.fork (“so a quick side job can run at low whatever effort the session uses”), and keeps what a mod attaches at prompt.submit in the user’s turn instead of a system message.
  • The built-in plugin-authoring skill now offers once, after a mod you built works, to help share it (v2.1.317). See Claude Mods.

Context and token use

  • /context no longer re-sends its visual grid to Claude on every later turn; Claude reads only the table (v2.1.319). At the terminal each list keeps its 10 largest rows, and on-demand MCP tools share one row per server until called.
  • Grep and Glob time limits: Claude no longer treats partial Grep and Glob results as complete when a search hit its time limit (v2.1.317).
  • Prompt caching switches: DISABLE_PROMPT_CACHING_OPUS and DISABLE_PROMPT_CACHING_SONNET now apply to [1m], other-region and undated spellings of the default model (v2.1.317).
  • Feature flags: CLAUDE_CODE_DISABLE_FEATURE_FLAGS stops feature-flag fetching without turning off telemetry; every flag uses its built-in default (v2.1.319).

Claude in Chrome

  • javascript_tool results up to 60,000 characters stay in the conversation instead of going to a file (v2.1.319).
  • After a brief disconnect it waits up to 6 seconds for the session’s own browser before picking another, and checks the page before repeating an action that may already have run (v2.1.319).
  • shortcuts_list and shortcuts_execute were removed from the built-in Chrome tools (v2.1.319). See Claude in Chrome.

Claude apps gateway

  • Cost report API GET /v1/organizations/analytics/cost and a daily record of each developer’s IdP groups, kept for admin.identity_retention_days (v2.1.317).
  • assume_role on mantle upstreams for a role in another AWS account and per-developer AWS cost attribution (v2.1.318).
  • v2.1.319: admin.pooled_conflict_rule and admin.pooled_priority_order (which pooled group budget pays first); an optional login_notice; CLAUDE_GATEWAY_LOG_FORMAT=json; spend for today, this week and this month in /usage; and start-up warnings for policies with no desktopHome that rely on coworkTabEnabled: false or chatSessionRetentionDays, “ahead of Claude Desktop’s November 10 change”. The spend-limit error now reads “spend limit reached · resets Dec 1”. See Claude apps gateway.

Cloud sessions and self-hosted runners

  • New in cloud sessions: a link card for links an MCP server asks you to visit (v2.1.317); a No repository option on Anthropic-hosted environments, the usual Chrome names (google-chrome, CHROME_PATH) for the built-in Chromium, and working gh run view --log (v2.1.318).
  • Scheduled tasks: /schedule warns when a new cloud scheduled task gets connectors you did not ask for (v2.1.317); new tasks start on the organization’s default environment (v2.1.317); a connector named memory is no longer dropped (v2.1.319). See Scheduled tasks.
  • Self-hosted runners: --trust-workspace false now also stops hooks in the repository’s settings files and skills from running (v2.1.319); the end-of-session debug-log upload was removed, so a failed session’s log stays on the runner host (v2.1.319); CLAUDE_RUNNER_SPAWN_REASON reaches the orchestrator’s spawn-runner hook (v2.1.318). Orchestrator log lines changed format in v2.1.318: update scripts that match retryable=true: or retryable=false:.

Other changes worth knowing

  • OpenTelemetry: anthropic.plan, anthropic.seat_type, anthropic.provider and anthropic.custom_base_url on metrics, events and spans (v2.1.317); exports behind a proxy time out after 10 seconds, or OTEL_EXPORTER_OTLP_TIMEOUT up to 30 (v2.1.317).
  • IDE extension: Claude Code no longer installs or updates the IDE extension when started in an IDE terminal; use /ide from an outside terminal or the editor’s extensions view (v2.1.318).
  • /model picker groups older models under “More models” (v2.1.319); a mistyped model name such as opus-5-5 now gets a suggestion instead of an update prompt (v2.1.319).
  • /claude-api managed-agents-onboard gives a new agent’s built-in tools the auto permission policy (“Judge risky actions”) by default (v2.1.318), and its templates now match the Claude Console (v2.1.319).
  • WebFetch on a claude.ai shared-chat link now says shared chats can’t be read and to paste the conversation (v2.1.319).
  • Claude Tag: pasted artifact links in Slack show a simpler card with an Open button (v2.1.317); a Claude Tag row on the Models admin page for Enterprise orgs (v2.1.319). See Claude Tag.
  • VS Code: prompts over 50,000 characters are no longer cut off silently (v2.1.319).

Announced alongside the releases

  • Cheaper subagent status updates (@ClaudeDevs, 2026-11-04): as in Key Takeaways. “The subagent itself keeps its own model.” Live “for most Pro, Max and Ultra users and for everyone on Team and Enterprise. Update to v2.1.316+.”
  • Cloud-session repo setup (@ClaudeDevs, 2026-11-06): on Pro, Max and Ultra, “Claude figures out what dependencies, network access, and API keys it needs, then saves the environment for next time. API keys are saved securely, so Claude never sees them.” The announcement gives no version number. See Claude Code in the Cloud: Field Guide.
  • Message threads in public beta (Messages API, 2026-11-02): agents send only the new messages each turn instead of the full history. “With message threads on by default for most Claude Code users, requests with 500k to 1M tokens see median TTFT drop by about 35%,” and “the median request went from about 800 KB to 1.8 KB.” Token usage is unchanged, because “Claude still reads the full conversation on every turn.” The SDK side shipped in October (see SDK releases, October 2026).

Breaking and behaviour changes

ChangeVersionWhat to do
Auto mode without server check asks or denies, no classifier fallbackv2.1.319Behind a gateway or with telemetry off, set CLAUDE_CODE_AUTO_MODE_SERVER=0 to keep the classifier, or =1 to force the server check
Self-paced /loop ends when a turn doesn’t reschedulev2.1.319Set CLAUDE_CODE_LOOP_KEEPALIVE=1 if you relied on fallback wakeups
Subagent effort follows the subagent’s modelv2.1.317Set CLAUDE_CODE_SUBAGENT_EFFORT (Week 44) for a fixed level
Flagged subagents stop on “Ask each time”v2.1.318Expect a stop, not a silent model switch
Read returns content for unchanged filesv2.1.317Re-reads cost tokens again
No IDE-extension auto-install from IDE terminalsv2.1.318Install via /ide or the editor
Exact ask rule refused under a broader denyv2.1.319Review rule overlap
Self-hosted orchestrator log formatv2.1.318Update log-matching scripts
Debug log no longer uploaded at session end (self-hosted)v2.1.319Collect logs on the runner host

Try It

  1. If you run Claude Code behind an LLM gateway or with telemetry off, run /status after updating to v2.1.319 and decide between CLAUDE_CODE_AUTO_MODE_SERVER=0 and =1 before an unattended job hits a denial.
  2. For unattended claude -p jobs, add permissions.denyInsteadOfPrompting so a stray prompt denies and continues instead of blocking, and check credential_expires_at in the JSON result to catch an expiring sign-in.
  3. If a /loop job depended on fallback wakeups, set CLAUDE_CODE_LOOP_KEEPALIVE=1 or make sure each turn schedules its next wakeup.
  4. Update to v2.1.319 for the Bash permission bypass, sed permission skip and worktree sandbox-escape fixes.

Open Questions

  • Does CLAUDE_CODE_DISABLE_FEATURE_FLAGS also opt a session out of server-side experiments? A 2026-11-08 Reddit analysis (raw/reddit-1xr66pd.md, see Claude Opus 5.5) traced behavior changes to experiment IDs sent as an x-cc-atis header. The release note says only that every flag “uses its built-in default”; whether experiment assignment is a feature flag is not stated.
  • What is “Claude Desktop’s November 10 change” to coworkTabEnabled and chatSessionRetentionDays? v2.1.319 warns ahead of it but does not describe it.
  • The sandbox escape and permission bypass fixes carry no CVE or advisory link in the notes.