Source: ai-research/claude-code-release-v2-1-{271,272,273}.md (fetched with gh release view on 2026-09-29) and raw/anthropic-watch-claude-code-tag-v2-1-{274,275,276,277,278}.md (watch feed) — anthropics/claude-code GitHub release notes, published 2026-09-14 to 2026-09-19 — plus the code.claude.com changelog snapshots ai-research/claude-code-docs-changelog-2026-09-18.md (v2.1.270–277, fetched 2026-09-18) and ai-research/claude-code-docs-changelog-2026-09-25.md (v2.1.278 entry, fetched 2026-09-25). No official W38 digest existed: code.claude.com/docs/en/whats-new/2026-w38 returned HTTP 404 on 2026-09-25.

Product: Claude Code · URL: https://github.com/anthropics/claude-code/releases · Date: 2026-09-14 → 2026-09-19

Eight releases (v2.1.271 through v2.1.278). Three changes reach beyond Claude Code’s own config: AGENTS.md is read when a project has no CLAUDE.md, skills and plugins enabled on claude.ai sync to terminal sessions, and auto mode’s classifier moves server-side by default for API, Enterprise and third-party-provider users, which the release note says “does not charge for classifier overhead”. Subagents gain omitClaudeMd and a reviewed hand-back call, the Monitor tool loses its no-timeout persistent option, and the deprecated TaskOutput tool is removed. Continues Week 37.

Key Takeaways

  • AGENTS.md support: in a project with no CLAUDE.md, Claude Code reads AGENTS.md instead; change this under “Project instructions” in /config; not yet on Bedrock, Vertex or Foundry (v2.1.277). See CLAUDE.md Primer. (Update 2026-09-29: per Claude Code’s memory docs, Bedrock and telemetry-off sessions load AGENTS.md from v2.1.281; Vertex and Foundry are not confirmed. See AGENTS.md.)
  • Auto mode defaults to the server-side classifier for Claude API and Enterprise users and on Bedrock, Vertex, Foundry and gateways; CLAUDE_CODE_AUTO_MODE_SERVER=0 opts out on Bedrock/Vertex/Foundry/gateways, a billed fallback warns, and /status gains an “Auto mode server” row showing where the classifier runs (v2.1.278). This reverses v2.1.273 four days earlier, which had set Bedrock/Vertex/Foundry to the local classifier “for now”. See Auto Mode.
  • claude.ai skills and plugins sync to the terminal: those enabled on your claude.ai account load in terminal sessions signed in with it; opt out with syncClaudeAiSkills: false or syncClaudeAiPlugins: false (v2.1.275). See Agent Skills.
  • omitClaudeMd in agent frontmatter and --agents JSON runs custom and plugin subagents without user, project and local CLAUDE.md files; managed policy files still load (v2.1.271). See Subagents.
  • Subagent output is fenced off twice over. In auto mode a subagent reports back through a dedicated hand-back call that the safety classifier reviews (v2.1.271), and subagent results reach the main agent under a header marking them as subagent output, indented, “so text in a subagent’s result cannot pass as the session’s own instructions” (v2.1.277).
  • Monitor watches always have a deadline — at most 30 minutes, 10 in single-prompt -p runs — and notify Claude to re-arm, replacing the no-timeout persistent option (v2.1.271). See Monitor Tool.
  • The TaskOutput tool is gone (v2.1.277): Claude reads a background task’s output file with Read, and taskOutputMaxChars and TASK_MAX_OUTPUT_LENGTH no longer have any effect.
  • Fable always appears in /model on the Anthropic API, greyed out only when organization settings disable it (v2.1.277).
  • The context meter had counted advisor-tool turns at about twice their real size, so auto-compact fired at about half the real window — fixed (v2.1.273). See Advisor Tool.

Release table

VersionDate (UTC)Shape
v2.1.2712026-09-14omitClaudeMd, subagent hand-back, Monitor deadlines, fast mode in Remote sessions
v2.1.2722026-09-15”Bug fixes and reliability improvements” only
v2.1.2732026-09-15Gateway hint headers, advisor context-meter fix, local classifier on 3P (reversed in 278)
v2.1.2742026-09-17MCP v2 client default on 3P, CLAUDE_CODE_MCP_STARTUP_WAIT_MS, tool_use_id loop self-heal
v2.1.2752026-09-17claude.ai skills/plugins sync, send-now key, npm plugins without install scripts
v2.1.2762026-09-18Single fix: advisor_20260301 400 behind proxies (2.1.275 regression)
v2.1.2772026-09-18AGENTS.md, TaskOutput removed, subagent-output header, Fable always listed
v2.1.2782026-09-19Server-side auto-mode classifier by default; /status row

Models and defaults

  • Fable always listed in /model on the Anthropic API (v2.1.277).
  • Server-side auto-mode classifier default (v2.1.278); v2.1.273 had briefly done the opposite on Bedrock/Vertex/Foundry.
  • Fast mode in Claude Code Remote sessions (cloud and self-hosted runners), where the organization allows it (v2.1.271).
  • modelPricing and the gateway pricing block accept a multiplier above 1, up to 10, for internal chargeback markups (v2.1.271).
  • Subagents with model: "opus" on Bedrock, Vertex or Foundry no longer leave the session’s model when the ID has no recognizable family (v2.1.274).
  • --resume no longer drops [1m] when the resumed session’s model family differs from the default (v2.1.271).
  • Advisor-tool turns counted at about 2x in the context meter — fixed (v2.1.273).
  • The background Haiku auto-title request is removed from claude -p runs launched outside an SDK or IDE (v2.1.277).
  • Dynamic workflows default to small on Pro plans, and the medium size guideline dropped from 15 to 10 agents (v2.1.271). See Dynamic Workflows.

New commands and features

  • AGENTS.md support (v2.1.277).
  • claude.ai skills and plugins sync (v2.1.275); sign-in with a Claude account also requests access to claude.ai plugins (v2.1.273); the new ListPlugins tool lists plugins enabled on the claude.ai account (v2.1.275).
  • Send-now key (ctrl+enter, or ctrl+x ctrl+s) interrupts the current turn and sends all queued messages at once; sent and queued messages show in gray until the model receives them (v2.1.275). The code.claude.com changelog snapshot of 2026-09-18 lists the key under v2.1.277 as well; the v2.1.277 GitHub notes do not.
  • /plugin install --marketplace offers to add the marketplace before installing (v2.1.275).
  • Fork a Remote Control session started with claude --remote-control or /remote-control from the Claude app; the fork runs as a background session on your computer (v2.1.273). See Channels.
  • Mouse support in the fullscreen /config panel (v2.1.271); click-to-expand collapsed teammate and agent messages in fullscreen (v2.1.274); a critical memory-usage warning (v2.1.274).
  • /mobile shows one QR code for claude.ai/mobile (v2.1.271).
  • Artifacts: Markdown files publish as styled document pages and a session can watch up to 10 artifacts, up from 5 (v2.1.271); artifact database updates can remove a single field (v2.1.273); the Artifact tool asks for a one-word tab icon instead of an emoji favicon (v2.1.275). See Artifacts.
  • claude mcp serve sends a progress update every 30 seconds during a running tool call (v2.1.271).
  • Invisible Unicode formatting and tag characters are stripped from prompts, and the cleaned prompt is shown for review before it is sent (v2.1.277).
  • Dynamic workflows pause at your usage limit and continue when it resets, instead of dropping agents (v2.1.271).
  • Claude apps gateway sign-in shows the signed-in account when the gateway names it; you confirm it before the credential is saved, and /status shows it (v2.1.275).
  • VS Code: a Memory dialog where you can view, edit and delete a saved memory, and sending an attached image without typing any text (v2.1.275); a Sign out row in the panel menu (with /logout), background shells and other running tasks in the agent map, each with a Stop (with /tasks), a Copy response button and /copy, and session cost and token usage in the Account & usage dialog where plan limits don’t apply (Vertex, Bedrock, Foundry, API key) (v2.1.277).

Settings and environment variables

  • CLAUDE_CODE_GATEWAY_HINT_HEADERS=1 opts in to x-claude-code-request-class, -agent-type, -prev-tool-durations, -compaction and -context-compacted request headers for LLM gateways (v2.1.273). See Claude Apps Gateway.
  • CLAUDE_CODE_AUTO_MODE_SERVER: introduced as a =1 opt-in on Bedrock/Vertex/Foundry (v2.1.273), flipped to default-on with =0 to opt out (v2.1.278).
  • CLAUDE_CODE_MCP_STARTUP_WAIT_MS bounds how long the first non-interactive turn waits for connecting MCP servers (0 = don’t wait) (v2.1.274).
  • OpenTelemetry: claude_code.managed_settings_resolved event with OTEL_LOG_MANAGED_SETTINGS=1 (v2.1.274); an effort attribute on the claude_code.llm_request trace span, matching the api_request event (v2.1.274); OTEL_LOG_TOOL_DETAILS=1 adds real agent, skill, plugin and MCP names to cost and token metrics (v2.1.273); OTEL_LOG_RAW_API_BODIES=file: gains an index.jsonl (v2.1.274); startup warning when otelHeadersHelper fails (v2.1.275).
  • syncClaudeAiSkills, syncClaudeAiPlugins (v2.1.275).
  • Gateway: CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1, for gateways whose only egress is a forward proxy — every outbound request hands the proxy the hostname instead of resolving it locally — and an optional upstream headers: map that sends static headers to a proxy you run in front of a provider (v2.1.277); CLAUDE_GATEWAY_DRAIN_TIMEOUT_MS (in-flight requests finish for up to 25 seconds on SIGTERM) and store.connect_timeout_seconds (v2.1.274); enduser.sub, the IdP subject, added to the telemetry Claude Desktop and Cowork send through a gateway, and a warning when a replica has more requests open than the 256 it sends upstream at once, with a startup log line showing that limit (v2.1.274).
  • claude self-hosted-runner --drain-marker-file and --host-config-snapshot disk|memory (v2.1.271).
  • taskOutputMaxChars and TASK_MAX_OUTPUT_LENGTH no longer have any effect (v2.1.277); taskOutputMaxChars had been added only in v2.1.261 (Week 36).
  • Managed allowManagedMcpServersOnly, deniedMcpServers and disableClaudeAiConnectors set via MDM or managed-settings.json are honoured when server-managed settings are also present (v2.1.273); an unreadable enterprise managed-mcp.json keeps exclusive MCP control and warns (v2.1.271).
  • MCP v2 client and 2026-07-28 protocol negotiation are the default on Bedrock, Vertex, Foundry and telemetry-disabled installs, “as other installs already do”; opt out with MCP_SDK_GENERATION=v1 or MCP_PROTOCOL_NEGOTIATION=legacy (v2.1.274). See MCP 2026-07-28 Spec.

Hooks, MCP, plugins, skills and subagents

  • omitClaudeMd (v2.1.271).
  • Subagent hand-back call reviewed by the classifier in auto mode (v2.1.271); subagent-output header (v2.1.277).
  • Monitor deadlines replace persistent (v2.1.271); a script’s final output and its exit arrive as one notification, saving a model turn (v2.1.274).
  • The hook spinner names which SessionStart, UserPromptSubmit, PreToolUse or SessionEnd hook is running, with elapsed time, and Esc cancels a prompt waiting on SessionStart (v2.1.271). See Hooks.
  • Stop prompt hooks’ repeat blocks name the condition with a 500-character label instead of re-sending the whole prompt (v2.1.274).
  • SubagentStop hooks with a specific matcher no longer fire for subagents with an empty agent type (v2.1.275).
  • Hook-driven sessions such as an active /goal compact instead of ending with “Prompt is too long”, and an active /goal survives --resume after compaction (v2.1.274). See Scheduled Tasks.
  • Auto mode: per-command allowed_domains for Bash, PowerShell and Monitor with sandboxing — the hosts a command needs are reviewed with it and opened for it alone (v2.1.271).
  • Plugins: claude plugin install/update --accept-command <sha256> accepts exactly the command a previous --json run displayed, instead of -y (v2.1.271); npm-source plugins fetched with npm pack --ignore-scripts and integrity-verified, so install scripts no longer run (v2.1.275); plugin and marketplace clones leave Git LFS files as pointers (v2.1.274); a top-level $schema in hooks/hooks.json no longer shows “unknown key” (v2.1.274). See Plugins and Marketplaces.
  • "type": "sdk" MCP entries in .mcp.json, settings, plugins and agent files are skipped with a warning (v2.1.274).
  • MCP: Streamable HTTP tool calls no longer time out at about 5 minutes despite a longer per-server timeout (v2.1.274); 403 insufficient_scope names the missing permissions (v2.1.274); a notification pointing at /mcp when a server disconnects mid-session and reconnection gives up (v2.1.273); prompts and resources not refreshing when a server sends list-changed notifications without declaring listChanged (v2.1.274); tool search matches bare MCP tool names (v2.1.271); list_changed tight-loop CPU fix (v2.1.271).
  • Skills synced from claude.ai move to the recoverable trash after sign-out (v2.1.271) or when the organization turns Skills off (v2.1.273).
  • /update-config writes Edit(path) rules instead of Write(path) rules, which file permission checks don’t match (v2.1.275).

Breaking and behaviour changes

  • TaskOutput tool removed (v2.1.277).
  • Monitor persistent option replaced by deadline-and-re-arm (v2.1.271).
  • Local-session artifact watching: a new version published elsewhere no longer starts a turn; Claude learns of it from a later Artifact tool result (v2.1.274).
  • /code-review uses leaner inline review prompts for models without tuned settings instead of spawning many review subagents (v2.1.274).
  • /status reads “Cloud sessions”, and messaging says “cloud session” instead of “Claude Code on the web” (v2.1.274).
  • Scheduled and Run-now routine runs save data to, and republish, artifacts you can edit without asking; public artifacts, first publishes and deletes still ask (v2.1.275).
  • A sandbox.excludedCommands glob must match every part of a compound command (v2.1.277).
  • v2.1.273 reverted the v2.1.268 change that applied Read and Edit deny rules to Bash lines the checker can’t analyze (eval, env -C); such commands prompt again instead of being denied.

Notable fixes

  • Sessions stuck endlessly retrying “unexpected tool_use_id” 400 errors: transcripts now self-heal where possible (v2.1.274).
  • Being logged out when an older Claude Code build, such as an IDE extension’s bundled CLI, runs on the same machine (v2.1.277).
  • claude -p and Agent SDK sessions hanging with no result after an internal error now exit with code 1 (v2.1.277). See Agent SDK.
  • Messages typed while Claude works sometimes ignored by the model (v2.1.277).
  • Sandboxed Bash on Linux reporting exit 0 for failed commands under zsh (v2.1.275).
  • /rewind in a forked or background session restoring zero-filled files (v2.1.275).
  • Security (v2.1.271): a cached organization policy reused after switching accounts, organizations or API keys; org policy fetched through, and rejected by, third-party local proxies set via ANTHROPIC_UNIX_SOCKET; cloud sessions rejecting every subagent tool call (“failed schema validation”) after the worker restarted.
  • Bash permission checks (v2.1.271): missing the file fmt, column and similar commands read when it follows an unrecognized option; skipping files a wildcard expands to in a pattern or option value (grep -v dir/* file); shell variable declaration flags misrepresenting the command.
  • /tui refusing to restart because of a finished agent-team teammate (v2.1.273), see Agent Teams; saved scheduled tasks running in the wrong session after .claude/scheduled_tasks.json was copied into another folder such as a new worktree (v2.1.273).
  • --forward-subagent-text stream-json and SDK output dropping the messages of subagents spawned by a context: fork skill; fullscreen freezing or blanking when scrolling past a large file diff; a stray </ccmemory> closing tag in responses (v2.1.275).
  • v2.1.276: every request failing with 400 … Input tag 'advisor_20260301' when ANTHROPIC_BASE_URL points at a proxy or gateway (2.1.275 regression).

Try It

  • Repos shared with other coding agents such as Codex CLI:^[inferred] keep one AGENTS.md; Claude Code now reads it when no CLAUDE.md exists.
  • After v2.1.275, check /skills in a terminal signed in with your claude.ai account: skills enabled on claude.ai should now appear there too.
  • Auto mode with sandboxing: give Bash commands that need one host a per-command allowed_domains entry instead of widening the global network rules.
  • Check /status for the “Auto mode server” row to confirm whether your classifier runs server-side.
  • Scripts that used Monitor persistent: true need a re-arm loop; watches now end within 30 minutes.
  • Scripts that called TaskOutput should read the background task’s output file instead.
  • Reviewer or linter subagents that don’t need project context: add omitClaudeMd: true to their frontmatter to save tokens.
  • If you don’t want claude.ai skills in the terminal, set syncClaudeAiSkills: false.
  • Air-gapped or proxy-only deployments: confirm the server-side classifier’s calls get through your proxy before upgrading to v2.1.278; if they don’t, set CLAUDE_CODE_AUTO_MODE_SERVER=0.
  • Watch for the billed-fallback warning in the first sessions after upgrading, especially on Bedrock, Vertex and Foundry.

Open Questions

  • v2.1.272 is unitemized (“Bug fixes and reliability improvements”).
  • Why did v2.1.273 set third-party providers to the local classifier “for now”, only for v2.1.278 to reverse it four days later? No rationale is given. The linked page (code.claude.com/docs/en/auto-mode-classifier-billing) was not fetched.
  • What triggers a billed fallback? v2.1.278 says Claude Code warns on one, but neither the release note nor the changelog entry names the conditions.
  • When will AGENTS.md support reach Bedrock, Vertex and Foundry? The note says “not yet”. Partly answered 2026-09-29: Bedrock from v2.1.281 per the memory docs; Vertex and Foundry still unconfirmed.
  • How does AGENTS.md interact with a CLAUDE.md that imports it? The note covers only the no-CLAUDE.md case. Answered 2026-09-29 (memory docs): Claude reads the CLAUDE.md and gets AGENTS.md through the import, never twice.
  • Does AGENTS.md follow CLAUDE.md’s hierarchy (user, project and subdirectory files), or is only a project-root AGENTS.md read? Not stated. Answered 2026-09-29 (memory docs): it follows the hierarchy. Every AGENTS.md and .claude/AGENTS.md in the working directory and above loads at start, and a subdirectory’s loads on demand.
  • Does omitClaudeMd also skip AGENTS.md when a project uses it in place of CLAUDE.md? Not stated. Partly answered 2026-09-29: the memory docs say subagents that “skip project instructions” skip AGENTS.md too, without naming omitClaudeMd. See AGENTS.md.
  • Does fast mode in Remote sessions give the same speed-up as in local sessions? v2.1.271 says only that the host’s setting or /fast applies where the organization allows it.