Source: ai-research/claude-code-changelog-v2-1-310-to-312-2026-10-28.md and ai-research/claude-code-docs-changelog-2026-10-30.md — the Claude Code changelog, versions v2.1.310–v2.1.312 and v2.1.314, fetched 2026-10-28 and 2026-10-30 respectively. v2.1.313 does not appear in the changelog and may be an internal/skipped build. v2.1.316 (GitHub release, 2026-10-31 01:01 UTC, which is still October 30 in Pacific time) was added on 2026-11-01 from raw/anthropic-watch-claude-code-tag-v2-1-316.md; v2.1.315 has no release. Continues Week 43, which covered through v2.1.309 (October 22).

Product: Claude Code · Date: 2026-10-23 to 2026-10-31 (UTC)

Four releases (v2.1.310–312, v2.1.314), with significant behaviour changes across the week. The ones to know: ultrathink now raises a turn to xhigh effort on the 5.5 models and Fable 5.1 (v2.1.311); Claude Code stopped sending the claude_code.safety_event OpenTelemetry event that Week 43 introduced (v2.1.312); standard-context sessions that run past 200K tokens now compact instead of stopping (v2.1.311); cloud routines get an effort picker (v2.1.312); and a new permissions.blockWritesOutsideWorkingDirectories setting hard-blocks all writes outside the working tree (v2.1.314). v2.1.310 shipped on October 23, the day after the Week 43 digest was written, so it is covered here. The week closes with v2.1.316, which deprecates fast mode on Opus 4.8 (removal November 30) and adds CLAUDE_CODE_SUBAGENT_EFFORT.

Key Takeaways

  • ultrathink = xhigh effort (v2.1.311): in the interactive terminal, the keyword runs its turn at xhigh on Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 through Anthropic’s API. Typed while Claude is working, it raises the rest of that turn. The rainbow highlight now appears only where the keyword can actually raise effort.
  • claude_code.safety_event is no longer sent (v2.1.312). If you built an audit pipeline on the Week 43 event, it has gone quiet.
  • No more “Usage credits required for 1M context” stops (v2.1.311): a standard-context session that crosses 200K tokens now compacts and continues.
  • Cloud routines get an effort level in the model picker when you create or edit a routine that runs in the cloud (v2.1.312).
  • Enterprise custom roles now cover cloud sessions and Remote Control (v2.1.310), so owners can grant each by role; turning either on now asks for confirmation and lists the roles that get it.
  • The Monitor tool and plugin monitors now work on Bedrock, Vertex and Foundry, and when telemetry or nonessential traffic is off (v2.1.310).
  • Effort-change cache warnings are more accurate (v2.1.311–312): /effort now warns about a cache miss in plan mode when the plan model differs, and a mid-task change may trigger the re-read before your next message.
  • permissions.blockWritesOutsideWorkingDirectories (v2.1.314): Write, Edit, and NotebookEdit refuse any path outside the working directories in every permission mode — a hard filesystem boundary.
  • Idle compaction in VS Code and Claude Desktop Code tab (v2.1.314): conversations over 200K tokens that sit idle are now compacted there too, matching terminal behaviour. Also runs with telemetry off and on Bedrock/Vertex/Foundry once the one-hour prompt cache is on.
  • Fast mode on Opus 4.8 is deprecated (v2.1.316): deprecated as of October 30, removed November 30, 2026. After that, fast-mode requests on Opus 4.8 return an error instead of falling back to standard Opus 4.8. Migrate to Opus 5.5 to keep the speedup.
  • CLAUDE_CODE_SUBAGENT_EFFORT (v2.1.316) sets the effort level subagents run at when nothing else names one, instead of inheriting the session’s. CLAUDE_CODE_DISABLE_ULTRATHINK turns the ultrathink keyword off.
  • Hook if conditions now match the Monitor tool (v2.1.314): Monitor(tail *) in a hook condition matches Monitor shell commands the same way Bash(tail *) matches Bash. Hooks matching Bash also run on commands started by the Monitor tool.

Release table

VersionDate (UTC)Headline changes
v2.1.3102026-10-23Monitor tool on Bedrock/Vertex/Foundry; custom roles for cloud sessions and Remote Control; gateway retry without effort-2025-11-24; cloud environment dialog redesign
v2.1.3112026-10-26ultrathink → xhigh; 200K-token sessions compact instead of stopping; self-hosted-runner idle-hold flags; Bedrock/Vertex/Foundry model settings from a signed catalog; ~70 fixes
v2.1.3122026-10-27safety_event OTel event dropped; routine effort picker; hook concurrency caps; oidc.user_id_claim for the Claude apps gateway; “5-hour limit” wording
v2.1.313—Not in changelog; may be an internal/skipped build.
v2.1.315—No release on GitHub.
v2.1.3162026-10-31Opus 4.8 fast mode deprecated; CLAUDE_CODE_SUBAGENT_EFFORT; CLAUDE_CODE_DISABLE_ULTRATHINK; gateway pooled group budgets and layered policies; --restricted upload fix; ~40 fixes
v2.1.3142026-10-30permissions.blockWritesOutsideWorkingDirectories; idle compaction in VS Code + Desktop Code tab; hook if for Monitor tool; groups for gateway upstreams; guardrail_trace for Bedrock guardrails; 81 bug fixes

Effort and ultrathink (v2.1.311, v2.1.312)

  • ultrathink in the interactive terminal runs that turn at xhigh on Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 (Anthropic API only). Typed mid-turn, it raises the rest of the turn.
  • The rainbow appears only where the keyword can raise effort, and not when the word is quoted or part of a file path.
  • The prompt’s effort indicator now says a mid-task change may wait for your next message; the spinner, OpenTelemetry effort values and transcripts show the level actually in force rather than the one just picked (v2.1.312).
  • /effort warns about a prompt-cache miss in plan mode when the plan model differs from the session model (v2.1.311).
  • /code-review now tells Claude to put the effort level first and to use max only when asked (v2.1.311).

How effort levels behave: Effort levels: when to use each.

claude_code.safety_event dropped (v2.1.312)

Week 43 (v2.1.307) added an OpenTelemetry event for every tool call stopped, prompted or sent back by a permission rule, auto mode, a safety check, a hook or the sandbox. v2.1.312 says only: “Changed Claude Code to stop sending the claude_code.safety_event OpenTelemetry event.” The changelog gives no reason and names no replacement. See Agent Observability.

Sessions and compaction (v2.1.311)

  • Standard-context sessions no longer stop with “Usage credits required for 1M context” when a request runs over 200K tokens; Claude Code compacts and continues.
  • When the model’s compaction reply contains no summary, Claude Code usually asks once more before replacing the conversation.
  • Claude no longer sees stale MEMORY.md and CLAUDE.md content right after automatic compaction.
  • A subagent’s autoCompactWindow now lowers the window set by CLAUDE_CODE_AUTO_COMPACT_WINDOW (it had no effect before); it still never raises it (v2.1.312).
  • DISABLE_COMPACT in settings files and in a mod’s $.env.set can only turn compaction off, not back on.

Cloud sessions, routines and Remote Control

  • v2.1.310: cloud sessions and Remote Control permissions added to Enterprise custom roles; a Choose file button loads a cloud environment’s setup script from a file; the cloud environment dialog opens network access, setup script and variables as rows.
  • v2.1.311: self-hosted-runner gains --hold-idle-file and --max-hold-min to keep an idle session while the host reports processes it left running; /teleport warns when a compacted cloud session loads from its last compaction.
  • v2.1.312: effort level in the routine model picker; a Bash editor (colours, line numbers, Copy) for setup scripts; cloud sessions started from the CLI get names and branch names from the server; WebFetch errors in cloud sessions now say an unresolved host may be outside the allowed network domains.

A practical guide to cloud sessions: Claude Code in the Cloud: Field Guide.

Hooks (v2.1.312)

  • Prompt and agent hooks now run at most 24 at a time per session, sub-agents included, plus at most 12 that have waited longer than their event’s timeout.
  • A timeout under 1 second now counts as 1 second.
  • v2.1.311 fixed SubagentStop hooks running about twice a minute per running subagent instead of once when it finishes.

Third-party platforms

  • Monitor tool and plugin monitors now on Bedrock, Vertex and Foundry, and when telemetry or nonessential traffic is disabled; long foreground sleeps are blocked there too (v2.1.310).
  • Bedrock, Mantle, Vertex and Foundry now load model settings from a signed file on downloads.claude.ai; CLAUDE_CODE_MODEL_CATALOG=off turns it off (v2.1.311).
  • Unrecognised model names (custom deployment names, inference-profile ARNs) now get the shorter system prompt even when they front Sonnet or Haiku; CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0 keeps the full prompt. Automatic model switches after a flagged message now also cover those names; CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK=1 turns them off (v2.1.312).
  • Requests behind gateways that reject the effort-2025-11-24 beta value or mid-conversation system messages are retried without them (v2.1.310).

Claude Tag and Code Review

  • Claude Tag’s Slack retry notice names the model in use or says rate limits were hit (v2.1.312).
  • Fixed: Claude in Slack not replying in channels whose environment has network access set to None; files posted in the wrong thread; channel settings falling back to workspace defaults when a private channel goes public (v2.1.312).
  • Code Review now reviews the pull request’s current commit when GitHub delivers a push event late (v2.1.312).

Other additions and changes

  • oidc.user_id_claim for the Claude apps gateway, to identify people by a claim other than sub (for example Entra ID’s oid) (v2.1.312).
  • Click to send a queued message in fullscreen (v2.1.312).
  • Usage limit messages say “5-hour limit” instead of “session limit”; VS Code shows reset times as a clock time (v2.1.312).
  • The [1m] suffix is dropped from model names where it has no effect; it is still accepted (v2.1.311).
  • --output-format json and stream-json escape DEL, C1 control and Unicode format characters as \uXXXX (v2.1.311).
  • Write now asks Claude to read a very large CLAUDE.md, rules file or @-included file in full, or edit in place, instead of overwriting it after seeing only its start (v2.1.312).
  • The Artifact tool now sends explainers and visual or interactive work to artifact pages, and documents only for writing people keep editing (v2.1.311).

Permissions and write safety (v2.1.314)

  • permissions.blockWritesOutsideWorkingDirectories: when set, Write, Edit, and NotebookEdit refuse any path outside the working directories in every permission mode — including Bypass. This is a hard filesystem boundary, not a permission-prompt decision.
  • /permissions Recently blocked now refuses to approve a call you cannot read in full.
  • File tools ask before editing the local copy of your organization’s server-managed settings.
  • /ultrareview now checks workspace trust before running git, matching the CLI claude ultrareview behaviour (see also the v2.1.309 note in Week 43).

Hooks and Monitor tool (v2.1.314)

  • Hook if conditions such as Monitor(tail *) now match the Monitor tool’s shell commands the same way Bash(tail *) matches Bash commands. Previously, Monitor invocations were invisible to hook pattern matching.
  • Hooks that match Bash now also run on commands started by the Monitor tool.
  • In Cowork cloud sessions, a hook’s allow decision on an MCP tool call now counts as no answer — the classifier or a permission prompt decides instead.

See Hooks for the full hook reference.

Idle compaction expands (v2.1.314)

  • Idle compaction (introduced v2.1.299) now runs in VS Code and the Claude Desktop Code tab as well as the terminal.
  • Idle compaction also runs for subscribers with telemetry off, and on Bedrock, Vertex, Foundry, and LLM gateways once the one-hour prompt cache is enabled.
  • The trigger remains: a conversation over 200K tokens that sits idle gets summarized; the summary is used on the next message if the prompt cache has expired.

See the Sessions and compaction section above for the earlier compaction changes in this week.

Enterprise gateway and security (v2.1.314)

  • groups list for Claude apps gateway upstreams: limit which IdP groups can reach a given upstream, including on failover.
  • guardrail_trace for Bedrock upstreams: logs which Bedrock guardrail filter fired without recording the matched text — audit-friendly.
  • oidc.user_id_claim (carried from v2.1.312): identify users by a custom JWT claim (oid for Entra ID instead of sub).
  • Flagged requests retry on another model instead of returning a refusal when ANTHROPIC_DEFAULT_OPUS_MODEL, availableModels, or deniedModels leaves only an older Opus or no model available.

Other v2.1.314 additions and changes

  • The bundled dataviz skill prints y-axis ticks in full for values under 10,000, and converts donut/pie chart requests to ranked bars or a segmented bar that prints values.
  • FORCE_HYPERLINK=false (also no or off) now correctly turns hyperlinks off instead of on.
  • The Bash tool clears any inherited which alias or function, so which resolves to the system command.
  • macOS: git grep in most Claude commands uses the C locale, avoiding a reported memory leak in macOS’s regex library.
  • Auto mode’s blocked-call messages now name the command and point to /permissions.
  • Auto-titles now derive from a skill or prompt-based command such as /security-review.
  • Ctrl+R in prompt history matches text inside collapsed [Pasted text] pills.
  • claude -p exits with code 1 and prints the error when startup fails (was: silent exit 0).
  • [Cloud sessions] Scheduled tasks load plugins enabled on your claude.ai account; sessions start even when a setup script exits with an error; an “Include completed” toggle on the Scheduled tasks page.
  • [VS Code] CLAUDE_CODE_DISABLE_CFC_PROMPT also stops the extension adding Claude in Chrome instructions.
  • [Claude Tag] Dismissible link to the Claude Tag API docs in admin settings for Enterprise orgs with Slack connected.

v2.1.316 (October 31 UTC)

Model and effort

  • Fast mode for Opus 4.8 deprecated as of October 30, 2026, and removed on November 30, 2026. After removal, fast-mode requests on Opus 4.8 return an error and do not fall back to standard Opus 4.8. The release note’s advice: migrate to Opus 5.5 to keep the speedup.
  • CLAUDE_CODE_SUBAGENT_EFFORT sets the effort subagents run at when nothing else names one, instead of inheriting the session’s level.
  • CLAUDE_CODE_DISABLE_ULTRATHINK turns the ultrathink keyword off.
  • [VS Code] ultrathink now runs its turn at xhigh effort with the rainbow highlight, matching the terminal (v2.1.311).
  • Fixed: sub-agents that inherited a parent sub-agent’s model moving to the main model when a background job or message woke them.
  • Fixed: /model not warning when it doesn’t recognise a model and auto-compact will assume a 200K context window.

Claude apps gateway (Enterprise)

  • pooled_budget on the spend-limits admin API: one monthly budget a group’s members share, on top of each member’s own limit.
  • managed.policy_mode: layered: every policy that matches a developer applies, each merged over the ones below it, instead of only the first match.
  • guardrail.input_tagging on Bedrock upstreams wraps prompts in the input tags a Bedrock guardrail’s prompt-attacks filter needs; all_messages_and_system also tags the system prompt.
  • Fixed: the gateway not checking every model a request names against availableModels, deniedModels and models. A request naming a blocked model now gets a 400 that names the field.
  • [VS Code] Sign in to a Claude apps gateway from the chat panel, no terminal needed.

Security and permissions

  • Fixed: --restricted sessions uploading files from outside the working directories when a Read allow rule covered them.
  • Fixed: a managed autoMode.allow list without "$defaults" getting the built-in exceptions back from a host app’s --settings, a first /permissions rule, or /auto-mode-setup. Such a list also no longer removes the built-in rule on scheduled-task prompts.
  • Server-managed settings fetched from Anthropic’s API for Team and Enterprise logins now apply without an approval prompt, and a stored copy that doesn’t verify is refused.
  • CLAUDE_CODE_DISABLE_LOOPED_CROSS_CERT_DROP keeps system cross-certificates that certify each other in a loop, for servers whose private CA needs one. SSL errors now name likely causes, such as a server omitting an intermediate certificate.

Shell and tools

  • Fixed: shell commands left running on macOS and Linux when Claude Code exits by itself (for example, when credentials are refused or a cloud session is archived).
  • Fixed: multi-line bash aliases and dash-named functions being garbled in the Bash tool; scripts stopping after their first command when SHELLOPTS is exported; PowerShell hooks not blocking on exit code 2 when the command starts with a type cast, param, using or an attribute.
  • Fixed: Grep and Glob never returning when the search process is killed at the timeout but doesn’t exit (seen on Windows).
  • Fixed: sonnet, opus and haiku on Vertex AI and Bedrock erroring mid-session instead of switching to an older model when the newest isn’t available to the account or region.

Cloud sessions, Claude Tag, Code Review

  • [Cloud sessions] A form card lets you fill in or decline a form an MCP server asks for; previously the request was cancelled unseen. A “Catching up…” line shows while a cloud or Remote Control session syncs. Enterprise members can now add and delete network secrets on their own cloud environments.
  • Fixed: cloud sessions and cloud scheduled tasks not being marked failed when auto mode stopped a turn because the server returned no safety verdict.
  • [Claude Tag] A channel’s “Respond automatically” switch became a “How Claude participates” choice on the channel’s Configure page and in admin settings. Fixes for canvas edits after cross-channel sharing and for Slack Connect threads where both organizations installed Claude.
  • [Code Review] A findings-by-severity card on the Code Review analytics page. /code-review started from the Desktop Code tab or the web now shows the findings card in cloud and Remote Control sessions.

Breaking and behaviour changes

  • claude_code.safety_event no longer emitted (v2.1.312).
  • ultrathink now changes effort on the 5.5 models and Fable 5.1, so a habit of typing it costs more than it used to
  • Hook concurrency caps (24 per session) and the 1-second minimum timeout (v2.1.312).
  • DISABLE_COMPACT can no longer be used to turn compaction back on (v2.1.311).
  • Model catalog download on Bedrock, Mantle, Vertex and Foundry (v2.1.311).
  • JSON output escaping of control and format characters (v2.1.311); parsed values are unchanged.
  • FORCE_HYPERLINK=false now turns hyperlinks off; the previous behaviour (turning them on) was a bug (v2.1.314).
  • Hook if conditions now match Monitor tool commands, so existing Bash(*) hook conditions fire on Monitor-invoked commands too (v2.1.314).
  • Opus 4.8 fast mode removed November 30, 2026 (deprecated in v2.1.316): requests error instead of falling back.
  • Server-managed settings apply without an approval prompt for Team and Enterprise logins (v2.1.316).
  • Hooks matching Bash run on Monitor-started commands — hooks scoped to Bash fire even when the command originates from the Monitor tool (v2.1.314).

Try It

  • Check any OpenTelemetry dashboard or alert built on claude_code.safety_event; it stopped receiving data with v2.1.312.
  • If you use fast mode on Opus 4.8 (scripts, settings.json, routines), move it to Opus 5.5 before November 30, 2026.
  • Set CLAUDE_CODE_SUBAGENT_EFFORT (for example low) if your subagents do routine lookups and shouldn’t inherit a high session effort
  • Use ultrathink for one hard turn instead of raising /effort for the whole session.
  • Set an explicit effort level on each cloud routine; a nightly report rarely needs more than the default
  • If you run Claude Code through a gateway with custom model names, decide whether you want CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0.

Open Questions

  • Why claude_code.safety_event was dropped one week after it shipped, and whether another event replaces it, is not stated in the changelog.
  • What the signed model catalog on downloads.claude.ai contains beyond “model settings” is not documented in the changelog.