Claude Code Week 26 Release Digest (June 12–25, 2026)

Source: ai-research/claude-code-docs-changelog-2026-06-13.md (live changelog snapshot 2026-06-13, covering CC v2.1.174–v2.1.176, June 12 2026); raw/anthropic-watch-claude-code-tag-v2-1-{174,175,176,177,178}.md (anthropic-ecosystem-watch stubs); ai-research/claude-code-changelog-v2-1-178-2026-06-15.md (de-truncated v2.1.178 GitHub release notes, June 15 2026); raw/anthropic-watch-claude-code-tag-v2-1-{186,187,190,191,193}.md (anthropic-ecosystem-watch stubs, June 22–25 2026 — see the Update section below)

Week 26 is a Remote Control reliability, managed model governance, and background-session hardening week. Three releases landed on June 12 in quick succession. The headline changes: v2.1.175 ships enforceAvailableModels — the first managed setting that locks the Default model selection (not just which models are offered), closing the gap where a user could still slip to a disallowed model via the Default fallback; v2.1.176 localises session titles to the conversation language, adds footerLinksRegexes for custom footer link badges, fixes a long-standing bug where hook if conditions on file-path tools (Read/Edit/Write) silently failed to match documented patterns, and closes a cluster of Remote Control edge cases; v2.1.174 adds wheelScrollAccelerationEnabled for fullscreen ergonomics, brings the /model picker inline with actual plan defaults (Opus now shows as its own row on Max/Enterprise), and adds a VSCode Account & usage breakdown showing per-skill/plugin/subagent/MCP token costs over 24h or 7d. A fourth release, v2.1.178 (June 15), extends the permission grammar with Tool(param:value) parameter matching (e.g. Agent(model:opus) to block Opus subagents), loads skills from nested .claude/skills directories, and runs auto mode’s classifier on subagent spawns before launch.

Headline changes (Claude Code)

v2.1.174 (June 12) — /model picker overhaul, VSCode usage attribution, session-env fix

/model picker

• Picker now shows the model that Default actually resolves to. Previously, the Default row did not reveal which model it mapped to for the current plan. Now Opus shows as its own row on Max/Team Premium/Enterprise and pay-as-you-go API accounts; Sonnet on Pro/Team plans. No more guessing what Default will give you.
• Hardcoded Sonnet version label fixed. When ANTHROPIC_DEFAULT_SONNET_MODEL was set to a specific Sonnet version, the /model picker still showed the hardcoded label instead of the pinned version. Fixed.
• /advisor no longer pre-selects a blocked model. If a saved advisor model is excluded by the availableModels allowlist, the dialog now opens with no pre-selection instead of pre-selecting a disallowed model.

VSCode usage attribution (new)

• [VSCode] Account & usage dialog (/usage) now shows a full attribution breakdown: cache misses, long context, subagents, and per-skill / per-agent / per-plugin / per-MCP server token costs over the last 24h or 7d. This is the VSCode companion to the CLI /usage per-component breakdown that shipped in W21. The breakdown answers “which skill is eating my plan credits?”

Settings

• wheelScrollAccelerationEnabled setting. Disable mouse-wheel scroll acceleration in fullscreen mode (user setting). Useful on trackpads where aggressive acceleration makes code review scroll past the target.

Fixes (selection)

• Fixed background sessions inheriting another session’s ANTHROPIC_* provider environment (gateway URL, custom headers, /model aliases) from the shell that launched the background daemon — sessions were picking up a different session’s gateway config.
• Fixed wheelScrollAccelerationEnabled setting (see above).
• Fixed “Fable 5 is now consuming usage credits” billing banner incorrectly showing for enterprise accounts on usage-based billing.
• Fixed Bedrock GovCloud regions (us-gov-*) deriving the wrong inference profile prefix (global instead of us-gov), causing 400 errors on inferred model IDs.
• Fixed git commit co-author attribution showing the wrong model name for some models.
• Fixed skill hot-reload re-sending the entire skill listing when only a single skill changed; now re-announces only changed skills, reducing noise in long sessions.
• Fixed Workflow tool agent() subagents missing per-agent attribution headers.
• Fixed pre-warmed background workers failing with “Could not resolve authentication method” when claimed after sitting idle.
• Fixed 1–2 second pause when exiting Claude Code shortly after a shell command was interrupted or killed on macOS and Linux.

v2.1.175 (June 12) — enforceAvailableModels managed setting

• enforceAvailableModels managed setting (enterprise/org admins). When enabled:
  1. The availableModels allowlist also constrains the Default model — if the Default would resolve to a disallowed model, it falls back to the first allowed model instead of silently using the disallowed one.
  2. User or project settings cannot widen a managed availableModels list — a user cannot add their own preferred models on top of the org’s allowlist. Previously, availableModels only restricted which models appeared in the picker and subagent dispatches; the Default model and user-side settings could still escape it. This setting closes both gaps. See W24’s requiredMinimumVersion/requiredMaximumVersion for the parallel binary-version governance surface.

v2.1.176 (June 12) — Session title language, footerLinksRegexes, hook if-condition fix, Remote Control cluster

Localisation

• Session titles are now generated in the conversation language. If you work in Spanish, session titles appear in Spanish. Set the language setting to pin a specific language if you want titles in a language other than the one detected from your messages.

New settings

• footerLinksRegexes setting (user or managed settings). Specify regex patterns that match link text or targets; matching links appear as compact badges in the footer row. Useful for surfacing ticket numbers, PR URLs, or custom deep links without them competing with session content. Configurable per user or pinned by an org policy.

Bedrock

• Bedrock credential caching improvement. Credentials from awsCredentialExport are now cached until their actual Expiration field instead of a hardcoded 1-hour TTL. Prevents unnecessary re-fetches for short-lived role credentials and avoids stale-credential errors when the 1-hour window didn’t match the actual expiry. See W24 for the Bedrock region autodetect context.

availableModels enforcement hardening

• Fixed alias model picks being redirected to a blocked model via ANTHROPIC_DEFAULT_*_MODEL environment variables — alias overrides now respect the allowlist.
• /fast now refuses to toggle when the toggle would switch to a model outside the allowlist.

Auto mode on Fable 5

• Fixed auto mode failing for organisations that have Fable 5 but not Opus 4.8 enabled — the auto-mode classifier now falls back to the best available Opus model instead of erroring. See Fable 5 and W22 where auto mode’s Opus-4.8 dependency was first introduced.

Hook if condition fix (important)

• Fixed hook if conditions for Read/Edit/Write tool paths. Documented patterns — Edit(src/**), Read(~/.ssh/**), Read(.env) — now match correctly. Previously these patterns silently failed on those three tools, meaning hooks gated on file paths were not firing as documented. If you wrote hooks with path conditions on Read/Edit/Write and found them unreliable, this release is the fix. See Hooks for the full if condition syntax.

Linux sandbox

• Fixed Linux sandbox failing to start when .claude/settings.json is a symlink with an absolute target.

Remote Control cluster

• Fixed Remote Control connecting from web or mobile silently switching the session’s model — model selection is now preserved when Remote Control attaches.
• Fixed Remote Control disconnect notifications showing a bare numeric reason code instead of a human-readable message; connection failures no longer add a duplicate line to the conversation transcript.
• Fixed Remote Control sessions not disconnecting when you sign in to a different account.

Clipboard and tmux

• Fixed /copy and mouse-selection copy not reaching the system clipboard inside tmux over SSH.
• Fixed tmux paste buffer not loading on tmux versions older than 3.2.

Background session fixes (cluster)

• Fixed backgrounded sessions showing “Working” forever when /bg was called mid-turn but there was nothing left to continue.
• Fixed /cd and worktree moves leaving the session reporting the previous directory’s git branch after moving.
• Fixed background agent PR URL search: PRs opened during scheduled wakeups or while a job was blocked now appear correctly in claude agents search.
• Fixed agents view input showing no text cursor on Windows.
• Fixed claude --bg -cn <name> not seeding the session name from the -cn flag.
• Fixed background sessions neutralizing Windows network paths in persisted state before respawn — paths like \\server\share were causing respawn failures.
• Fixed background-session respawn rejecting malformed resume IDs from corrupted state files (defensive parse instead of hard reject).
• Fixed Windows background-service daemon not starting when ~/.claude/daemon has the ReadOnly attribute set.
• Fixed cloud sessions failing with “Could not resolve authentication method” when idle too long before being claimed.
• Background sessions now show clearer guidance when a window left open across an auto-update can’t submit a reply; claude daemon status explains version-skew behavior.
• Fixed claude agents pressing back in one window detaching other windows attached to the same session.

v2.1.177 (June 13) — shipped, no release notes

• v2.1.177 was tagged 2026-06-13 (the day after the W26 trio), but the GitHub release body is empty — no published changelog as of this writing. Recorded here for completeness; the substantive Week 26 releases are v2.1.174–176 plus v2.1.178 (June 15, below). The next changelog snapshot should backfill any 2.1.177 changes.

v2.1.178 (June 15) — Tool(param:value) permission syntax, nested .claude/ skills + precedence, auto-mode classifier on subagent spawns

Permission grammar (headline)

• Tool(param:value) permission-rule syntax. Permission rules can now match a tool’s input parameters, with * wildcard support — e.g. Agent(model:opus) blocks Opus subagents. This pushes the permission grammar past tool-name matching (and past W25’s glob deny-rules) down to per-parameter matching, a finer-grained governance surface. See CLI Reference.
• MCP server-level specs now honoured in subagent disallowedTools (fix). mcp__server, mcp__server__*, and mcp__* specs were being silently ignored in a subagent’s disallowedTools; server-scoped MCP denials now actually apply to subagents.

Nested .claude/ directories

• Skills in nested .claude/skills directories now load when you work on files in that subtree; on a name clash with a higher-level skill, the nested skill is surfaced as <dir>:<name> so both stay available. (Companion fix: directory-qualified nested-skill names were being blocked by permission prompts in non-interactive runs.)
• Closest-directory-wins precedence. When agent, workflow, and output-style names collide across nested .claude/ directories, the one closest to the working directory wins; project-scope workflow saves now target the closest existing .claude/workflows/. Makes monorepo / multi-package .claude/ layouts behave predictably.

Auto mode + subagents

• Subagent spawns are now evaluated by the classifier before launch — closes a gap where a subagent could request a blocked action without review (continues W25’s cross-session-authority hardening).
• Subagent UX fixes: transcripts now show tool results + live progress; messages sent while a subagent finishes its turn are no longer dropped; backgrounding a running subagent (ctrl+b) no longer restarts it from scratch.

Other

• /doctor redesigned — consistent flat tree layout across all sections, clearer status icons, highlighted command names.
• Workflow prompt keyword now uses a purple shimmer highlight and triggers only on explicit phrases (“run a workflow”, “workflow:”), not on any mention of the word.
• /bug now requires a description before submitting and no longer uses model-refusal text as the GitHub issue title.
• Notable fixes: compaction now honours --fallback-model on overload / model-availability errors; an OOM crash from a stale websocket/OAuth file-descriptor env var; claude agents workers 401 Invalid bearer token under a custom ANTHROPIC_BASE_URL/ANTHROPIC_AUTH_TOKEN gateway; stale cached request config after an external credential refresh; /bg sessions stuck “Working” forever after a finished turn; Claude in Chrome silently failing when the OAuth token belonged to a different account.

v2.1.179 (June 16) — bugfix release

Streaming and connectivity

• Fixed mid-stream connection drops: partial responses are now preserved instead of showing a raw error, and the spinner no longer gets stuck at “running tool.”
• Fixed mouse-wheel scrolling in WSL2 under Windows Terminal and VS Code (regression introduced in v2.1.172).

Sandbox

• Fixed denyRead/allowRead glob over a large directory tree making the Bash tool description enormous and the session unusable on Linux.

UI fixes

• Fixed feedback survey capturing a single-digit reply as a session rating immediately after a turn completes.
• Fixed welcome screen stacking multiple promotional banners — at most one promo now shows per session.
• Fixed Ctrl+O not showing the subagent’s transcript when viewing a subagent.
• Fixed clicking the prompt input not returning focus from the subagent/footer panel.
• Fixed remote session background tasks appearing stuck as “still running” between turns.
• Improved plugin loading performance in remote sessions.

v2.1.181 (June 17) — /config key=value, CLAUDE_CLIENT_PRESENCE_FILE, sandbox.allowAppleEvents

New commands and settings

• /config key=value syntax — set any setting from the prompt without opening the interactive config UI (e.g. /config thinking=false). Works in interactive, -p, and Remote Control modes. /config --help lists all available shorthand keys (also added in v2.1.183 — see below).
• sandbox.allowAppleEvents setting — opt-in to let sandboxed commands send Apple Events on macOS (the sandbox previously blocked them, breaking open, osascript, and browser-based auth flows with error -600).
• CLAUDE_CLIENT_PRESENCE_FILE env var — point it at a marker file on disk; while the file exists, mobile push notifications for that session are suppressed. Useful when you’re at the machine and don’t want your phone buzzing for every agent turn.

Runtime and performance

• Upgraded the bundled Bun runtime to 1.4.
• Improved streaming of long paragraphs: text now appears line-by-line instead of buffering until the first line break.
• Improved auto-retry: API connection drops mid-thinking now automatically retry instead of surfacing “Connection closed while thinking.”

Subagent panel

• Idle subagents auto-hide after 30 seconds.
• List caps at 5 rows with scroll hints.
• Keyboard hints now show in the footer.

MCP OAuth

• OAuth browser page now matches Claude Code’s visual style and auto-closes on success.

Fixes (cluster)

• Fixed prompt caching not reading on custom ANTHROPIC_BASE_URL and on Foundry due to a per-request attestation token changing every turn.
• Fixed Write/Edit producing 0-byte or truncated files on network drives and cloud-synced folders.
• Fixed a startup regression (~120ms per launch in fresh environments, introduced in v2.1.169).
• Fixed startup blocking with a blank terminal for up to 15 seconds when the account settings fetch is slow on a degraded network.
• Fixed startup crash (TypeError: Cannot read properties of null) when .claude.json contains corrupted null project entries.
• Fixed macOS TUI freezing at session start (Ctrl+C unresponsive) when Spotlight is busy reindexing.
• Fixed long-running idle sessions losing their history when another Claude Code process ran the 30-day transcript cleanup.
• Fixed foreground subagents spawning unbounded nested chains; they now respect the same 5-level depth limit as background subagents.
• Fixed AWS awsCredentialExport credentials with a short remaining lifetime causing credential refreshes every minute.
• Fixed fullscreen mode URL opening to require Cmd+click (macOS) / Ctrl+click, matching native terminal behavior.

v2.1.183 (June 18) — auto-mode destructive-command blocks, deprecated-model warnings, attribution.sessionUrl

Auto mode safety (headline)

• Destructive commands are now blocked in auto mode unless you asked for them. Destructive git commands (git reset --hard, git checkout -- ., git clean -fd, git stash drop) are blocked when you didn’t ask to discard local work; git commit --amend is blocked when the commit wasn’t made by the agent this session; and terraform destroy / pulumi destroy / cdk destroy are blocked unless you asked for that specific stack. This extends the auto-mode classifier from vetting new actions (W25/W26) to guarding irreversible ones.

Model governance

• Deprecated / auto-updated model warning now shows on stderr in print mode (-p) and also covers models set in agent frontmatter — continuing W26’s managed-model theme (enforceAvailableModels, v2.1.175).

Settings + UX

• attribution.sessionUrl setting omits the claude.ai session link from commits and PRs in web and Remote Control sessions.
• /config --help lists all shorthand keys for /config key=value; the /config toggle now uses Enter/Space to change and Esc to save-and-close; the startup “setup issues” line under the logo was removed (run /doctor instead).

Fixes

• Fixed thinking.disabled.display: Extra inputs are not permitted 400 errors on subagent spawns and session-title generation; fixed WebSearch returning empty results in subagents; fixed the terminal cursor being stranded after vim-mode history navigation; fixed fullscreen TUI corruption (statusline mid-screen, duplicated spinners) in Windows Terminal under heavy nested-subagent load.

SDK releases (June 18)

• code_execution_20260120 tool added to the Python SDK (v0.110.0) and TypeScript SDK (sdk-v0.105.0) — a new dated server-side code-execution tool version. The Python SDK also added (v0.111.0) a refusal-fallback middleware tag (fallback-refusal-middleware) — the SDK-side counterpart to the Fable 5→Opus visible-fallback behavior. Remaining TypeScript package bumps (aws/bedrock/foundry/vertex) are routine (“lazily parse partial tool json input” + deprecated-model test chores).

Update — Claude Code v2.1.186–195 (June 22–26)

Five more releases landed June 22–25, extending Week 26’s managed-model governance and auto-mode-classifier themes. New surface grouped by area (pure-bugfix releases noted as one-liners):

Security and governance

• sandbox.credentials setting (v2.1.187). Blocks sandboxed commands from reading credential files and secret environment variables — a sandbox-scoped secret-exfiltration guard alongside the existing sandbox network controls.
• Org-configured model restrictions are now enforced everywhere (v2.1.187). Restrictions apply across the model picker, --model, /model, and ANTHROPIC_MODEL; selecting a restricted model shows a “restricted by your organization’s settings” message. Extends v2.1.175’s enforceAvailableModels down to the env-var and flag entry points.
• autoMode.classifyAllShell setting (v2.1.193). Routes all Bash/PowerShell commands through the auto-mode classifier instead of only arbitrary-code-execution patterns. Same release surfaces auto-mode denial reasons in the transcript, the denial toast, and /permissions recent denials.

MCP and auth

• claude mcp login <name> / claude mcp logout <name> (v2.1.186). Authenticate MCP servers from the CLI without opening the interactive /mcp menu; --no-browser stdin-redirect support completes the OAuth flow over SSH. See CLI Reference.
• Remote MCP tool-call hang timeout (v2.1.187). Remote MCP tool calls that hang with no response now abort with an error after 5 minutes instead of blocking the turn indefinitely. Override the window with CLAUDE_CODE_MCP_TOOL_IDLE_TIMEOUT.
• Startup notice when MCP servers need authentication (v2.1.193) — points you at /mcp instead of failing silently mid-session.

Workflow and reliability

• /rewind from before /clear (v2.1.191). /rewind can now resume a conversation from a checkpoint before /clear was run — /clear is no longer a hard context wall.
• Stopped background agents stay stopped (v2.1.191). Stopping an agent from the tasks panel is now permanent; previously stopped agents could resurrect.
• Comma-separated hook matchers fixed (v2.1.191). Hooks with comma-separated matchers (e.g. "Bash,PowerShell") were silently never firing — now fixed. If you used multi-tool matchers and found a hook dead, update and re-test. See Hooks.
• /permissions denial approval persists (v2.1.191). Approving an entry in the Recently-denied tab now persists on close instead of being silently discarded.
• ! bash auto-response (v2.1.186). ! bash commands now trigger Claude to respond to the output automatically; set "respondToBashCommands": false to keep the previous context-only behavior. Same release adds a Skills section to the /plugin Installed tab and a teammateMode: "iterm2" setting.
• Live file-path autocomplete in bash (!) mode (v2.1.193).
• v2.1.190 (June 24) — bug fixes and reliability improvements only.

Observability gotcha (v2.1.193)

• New claude_code.assistant_response OpenTelemetry log event carries the model’s response text. It is redacted unless OTEL_LOG_ASSISTANT_RESPONSES=1 — but when that variable is unset it falls back to OTEL_LOG_USER_PROMPTS. So any deployment already logging prompt content will start logging response text on upgrade. Set OTEL_LOG_ASSISTANT_RESPONSES=0 to keep prompts-only logging.

v2.1.194–195 (June 26)

• Hook matchers with hyphenated identifiers now exact-match (v2.1.195). Matchers like code-reviewer and mcp__brave-search were accidentally substring-matching; they now exact-match. To match all tools from a hyphenated MCP server, use mcp__brave-search__.*. Continues W26’s run of hook-matcher correctness fixes (cf. the comma-separated-matcher fix in v2.1.191). See Hooks.
• CLAUDE_CODE_DISABLE_MOUSE_CLICKS (v2.1.195) — disables mouse click/drag/hover in fullscreen mode while keeping wheel scroll (pairs with v2.1.174’s wheelScrollAccelerationEnabled).
• Background-agent reliability cluster (v2.1.195): fixed background jobs disappearing / losing data when written by a newer Claude Code version; a crashed background task showing a blank screen for up to 5s instead of its restart; and background-agent daemons running unreachable when the control socket fails to start (which had blocked restarts). Remote session startup now shows a provisioning checklist.
• Voice + plugins (v2.1.195): voice-dictation fixes (macOS capturing silence after the default input device changes; auto-submit never firing for space-less languages — Japanese/Chinese/Thai); Linux voice mode now distinguishes “no microphone” from “SoX not installed.” Plugin fixes: external plugins enabled only by project .claude/settings.json now require explicit install consent on every loader path; /plugin Enable/Disable now works when a plugin’s plugin.json name differs from its marketplace entry name.

Key Takeaways

• /config key=value (v2.1.181) brings inline settings changes — set any configuration option from the prompt in interactive, -p, and Remote Control modes; /config --help (v2.1.183) lists all shorthand keys. Combined with sandbox.allowAppleEvents (fixes error -600 on macOS osascript/browser auth flows) and CLAUDE_CLIENT_PRESENCE_FILE (suppress mobile push notifications while at the machine), v2.1.181 is a quality-of-life cluster comparable in density to v2.1.176’s bugfix batch.
• v2.1.181 also fixes Write/Edit producing 0-byte files on network/cloud-sync drives and unbounded nested foreground subagent chains (now capped at the same 5-level depth limit as background subagents). Prompt caching now reads correctly on ANTHROPIC_BASE_URL and Foundry. Startup regression (~120ms/launch since v2.1.169) is fixed.
• Auto mode now blocks destructive commands (v2.1.183). Destructive git (reset --hard, checkout -- ., clean -fd, stash drop), git commit --amend on commits the agent didn’t make this session, and terraform/pulumi/cdk destroy are blocked in auto mode unless explicitly requested — the auto-mode classifier moves from vetting new actions to guarding irreversible ones. Same release fixes WebSearch returning empty in subagents and thinking.disabled 400 errors on subagent spawns. (Corroborates the Fable-5-card finding that overeager/destructive actions run slightly elevated — these are the product-side guardrails.)
• Tool(param:value) permission rules (v2.1.178) bring per-parameter governance to Claude Code. Agent(model:opus) blocks Opus subagents; combine with v2.1.175’s enforceAvailableModels for model control across the picker, Default resolution, and subagent dispatch. The same release fixes server-level MCP specs (mcp__server__*) being silently ignored in subagent disallowedTools.
• Nested .claude/skills + closest-directory-wins precedence (v2.1.178) make monorepo / multi-package skill layouts work: subtree-local skills load when you’re in that subtree and surface as <dir>:<name> on a clash, and the nearest agent / workflow / output-style wins on a name collision.
• enforceAvailableModels (v2.1.175) closes the managed-model governance gap. Previously, availableModels only filtered the picker — Default could still resolve to a disallowed model, and users could add models on top of the org list. Now both loopholes are closed. Pair with W24’s version gates for a complete compliance surface.
• Hook if conditions for Read/Edit/Write now work correctly (v2.1.176). If you had path-gated hooks on those three tools and found them unreliable, update and re-test. This is a silent regression fix, not a new feature.
• VSCode Account & usage now has per-skill/plugin/MCP attribution (v2.1.174). The CLI has had /usage per-component breakdown since W21; the VSCode panel now matches. Use it to audit which plugin or MCP server is driving plan consumption.
• Session title localisation (v2.1.176) is cosmetic but useful in multilingual teams — session lists in shared dashboards now surface titles in the language the operator used, not always English.
• Remote Control model-switching bug is fixed (v2.1.176). Web/mobile Remote Control was silently overriding the model on attach — now model selection is preserved.
• Background session Windows robustness (v2.1.176): three Windows-specific fixes (ReadOnly daemon dir, network paths in state, no text cursor in agents view) in one release — a pattern of Windows hardening continuing from W24/W25.
• Bedrock credential caching now honours actual Expiration (v2.1.176) — relevant for roles with TTLs other than 1 hour (GovCloud, short-lived OIDC-provisioned credentials).

Related

• Week 25 Digest — Managed Agents deployments/vault credentials, fallbackModel, --safe-mode, post-session hook, Fable 5 access gate (v2.1.170).
• Week 24 Digest — requiredMinimumVersion/requiredMaximumVersion version gates, Stop/SubagentStop additionalContext hooks, OTEL metric dimensions — the governance context for v2.1.175’s enforceAvailableModels.
• Claude Code Hooks — hook if condition fix for Read/Edit/Write tool paths (v2.1.176); verify path-gated hooks after updating.
• CLI Reference — footerLinksRegexes, wheelScrollAccelerationEnabled, language, enforceAvailableModels settings (v2.1.174–v2.1.176); Tool(param:value) permission syntax + per-parameter rules (v2.1.178).
• Claude Fable 5 and Mythos 5 — auto-mode Opus fallback fix (v2.1.176) for orgs without Opus 4.8.
• Agent Teams — background session PR URL search, /bg Working-forever fix.
• Claude Code Subagents — Workflow tool agent() attribution headers fix; background session respawn hardening; auto-mode classifier now vets subagent spawns before launch and MCP server-level disallowedTools apply to subagents (v2.1.178).