Source Intelligence
Follow Discord

DisclaimerUnofficial, and not affiliated with Anthropic. Nearly all of this is read straight out of what ships: npm bundles, captured prompts, published docs. Anthropic's own notes go in verbatim, marked as theirs. The rest is my reading, and every entry carries the strings behind it. If one looks wrong, vote it down and say why.

← All releases Home ←olderv2.1.205 v2.1.207newer→

Claude Code v2.1.206

17 entries read diff v2.1.205 → v2.1.206 Markdown

This release delivers a significantly revamped Claude in Chrome setup experience with an interactive wizard and multi-browser detection, introduces a new EndConversation tool (feature-flagged, rolling out to eligible models), adds /plan share artifact publishing, a new dashboard artifact skill, and Chart.js runtime support for richer artifact charts. Plugin binary provisioning gains mismatch detection, and push permission rules now dynamically pick up non-origin remotes.

Find
Pick an entry · j / k steps through
5 entries

New Featuresopen

Claude in Chrome — Revamped Setup Wizard#

Feature flag
tengu_chrome_auto_enable On for this account, and not off by default

The flag server returned on for the one account this site reads, and nothing in this release compiles it off by default. The compiled default is shown below, and says which it is when we cannot read one: a fifth of gates compile in a string or a number rather than on or off, and most published releases have no gate table behind them at all. No client can see what the server returns for your account.

This account: on · anonymous baseline: off · compiled default in v2.1.206: on

These values were read against a different version of Claude Code, so treat them as the nearest reading available instead of one taken on this release.

Read once, for one account on one subscription tier, against v2.1.206. It isn't a statement about your account. What a flag value here can and cannot tell you

What

The Claude in Chrome installation flow is now an interactive step-by-step wizard inside the terminal, replacing the previous one-shot skill registration.

Details
  • When Claude needs browser tools and the extension isn't installed, a new "Claude wants to use your browser" upsell prompt appears with three options: install the extension, skip for now, or "Don't ask again."
  • Choosing to install opens a "Setting up Claude in Chrome" dialog that follows the connection progress in real time — tracking "Install the extension" and "Connect to Chrome" as live steps with loading/success/error states.
  • If setup stalls, the dialog shows: "Taking longer than expected. Check that the extension is added and that you're signed in to it in Chrome."
  • On success: "Browser tools are ready. Continuing keeps them enabled for future sessions too — manage anytime with /chrome."
  • Browser detection now covers seven Chromium-family browsers: Chrome, Brave, Arc, Edge, Chromium, Vivaldi, and Opera across macOS, Linux, and Windows.
  • Auto-enable is gated by the tengu_chrome_auto_enable feature flag: when enabled, Claude activates browser tools automatically if the extension is already installed/paired.
  • The extension install page URL and /chrome management command are unchanged.
Evidence

Setup wizard UI (search for "Setting up Claude in Chrome"); upsell dialog (search for "Claude wants to use your browser"); browser path map (search for "BraveSoftware")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Related

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

Plan Publishing as Artifact#

What

/plan share can now publish the current plan as a shareable web artifact on claude.ai, letting reviewers read the plan in a browser before the session exits.

Details
  • When a plan artifact is available, a new "Review plan as artifact" option appears alongside the existing skip option.
  • While publishing, the session shows: "Publishing plan for review…"
  • On success, the URL is displayed: "Review your plan: {url}"
  • If the plan changes after publishing, the status updates to show "(plan changed since publish — after exiting, /plan share publishes the updated plan)"
  • If publishing fails: "Couldn't publish plan — run /plan share to retry, or --debug for details."
  • If the feature is unavailable: "Publishing plans isn't available right now — the plan was not published."
Evidence

Status strings (search for "Publishing plan for review"); stale state (search for "plan changed since publish")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Dashboard Artifact Skill (artifact-dashboard)#

What

A new built-in artifact-dashboard skill provides a production-ready template for creating operational dashboard artifacts with KPI tiles, a primary chart, and a breakdown table.

Usage
Load the artifact-dashboard skill and build a dashboard from its template.
Details
  • The skill includes a template.html fragment with slot markers (<!-- SLOT: ... -->) for the title, KPI tiles, chart spec, breakdown rows, and footer note.
  • KPI tiles support directional delta styling with up/down classes plus good/bad overrides for cases where a decrease is an improvement (e.g., latency, error rate).
  • The primary chart slot takes a compact JSON spec ("type": "line" | "bar" | "donut", "series", "x", "y") and is rendered by a built-in SVG fallback plus the injected Chart.js runtime (see below).
  • A CONTRIBUTING.md-style notes section warns against fabricating data: "Replace every placeholder number — and never invent one."
  • Theming uses CDS token values with full dark mode support via @media (prefers-color-scheme: dark) and data-theme overrides.
Evidence

Skill FRONTMATTER (search for "artifact-dashboard"); slot guidance (search for "SLOT: KPI_TILES")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Chart.js Runtime in Published Artifacts#

What

Artifacts that include a <script type="application/json" data-chart-runtime> chart spec now receive an injected Chart.js runtime at publish time, enabling bar and donut charts in addition to SVG line charts.

Details
  • The publish pipeline detects the data-chart-runtime attribute and injects a bundled Chart.js into the artifact HTML between <!-- chart-runtime --> / <!-- /chart-runtime --> sentinel comments.
  • Safety checks reject bundles containing </script, HTML comment injection risks, or the sentinel strings themselves.
  • The runtime reads the spec from the data-chart-runtime element and renders into #primary-chart via Chart.js, fully replacing the SVG fallback.
  • Theme support: resolves CSS custom properties from getComputedStyle at render time and re-renders on system dark-mode changes or data-theme flips — same data, same pixels per theme.
  • Multi-series line/bar charts include a legend; donut charts accept per-slice color overrides.
  • The SVG fallback remains for environments without the runtime, covering the standard time-series line case.
Evidence

Sentinel string (search for "bundle contains the chart-runtime sentinel"); runtime injection (search for "chart-runtime -->"); Chart.js integration (search for "new Chart(canvas")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Dynamic Git Remote Detection for Push Permissions#

What

Permission rules for git push now automatically include the repo's configured push remote, not just origin.

Details
  • Before adding a Bash(git push origin *) rule, Claude now queries git config remote.pushDefault to find the actual configured push target.
  • If only one remote is configured, it is used regardless of name.
  • The resolved remote is added alongside origin in the allowed-patterns list, so both git push origin and git push {remote} are permitted.
  • This prevents unnecessary permission prompts for repos that push to upstream, a fork remote, or any non-origin name.
Evidence

Remote resolution (search for "remote.pushDefault"); dynamic rules (search for "git push -u" alongside "rwn")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

8 entries

Improvementsopen

Plugin Binary Provisioning — Mismatch Detection and Bare Name Linking#

What

Plugin binary installation now removes and re-fetches binaries whose on-disk digest doesn't match the manifest, and creates hard links from arch-specific binary names (e.g., tool-aarch64-apple-darwin) to bare names (e.g., tool) for easier execution.

Details
  • If a bin/ entry exists but its SHA-256 doesn't match the manifest, Claude logs "removed bin/{name} — it is not the artifact pinned by the manifest" and re-downloads it.
  • After placement, a hard link from the arch-specific filename to a bare name is created on macOS/Linux when exactly one declared entry derives that bare name (collision-safe).
  • Four target triples are recognized for bare-name derivation: aarch64-apple-darwin, x86_64-apple-darwin, aarch64-unknown-linux-musl, x86_64-unknown-linux-musl.
  • New telemetry fields track mismatch_removed_count and bare_placed_count separately from download and cache-hit counts.
  • A new validation pass checks .mcp.json and nested MCP server configs for binary references that aren't in the binaries map, warning: "bin/{name} is not a shipped file, a declared binaries entry, or a name derivable from the declared entries — the server will fail to start."
Evidence

Mismatch removal (search for "it is not the artifact pinned by the manifest"); validation (search for "Check for a typo against the \"binaries\" map in plugin.json")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Model Pricing Catalog#

What

Model pricing is now baked into the model catalog at build time rather than fetched at runtime, with validation that every catalog entry has complete pricing data.

Details
  • A new pricing-tier resolver (BVo) maps model entries to their pricing shapes, supporting both direct pricing objects and shared tier references.
  • Pricing requires input, output, cache_write_5m, cache_read, and web_search fields; missing fields throw at startup: "model catalog entry '{id}' has incomplete pricing — baked entries need the full ModelCosts shape."
  • Model IDs must be registered in CATALOG_MODEL_IDS; unregistered IDs throw: "model catalog id '{id}' missing from CATALOG_MODEL_IDS — regenerate with 'bun run generate:model-catalog'."
  • The first-party billing UI uses the catalog pricing to display per-token costs in the model picker.
  • Claude Sonnet 5 can show promotional pricing text when a promo expiry date is set (search for "promo through").
Evidence

Validation error (search for "baked entries need the full ModelCosts shape"); catalog check (search for "model catalog id missing from CATALOG_MODEL_IDS")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Refusal Category Fallback Routing#

What

When the API returns a refusal with a known category, Claude can now route the retry to a specific fallback model rather than always using the armed default fallback.

Details
  • The cyber refusal category routes to claude-opus-4-8 on Opus plan configs.
  • The bio category additionally routes to claude-opus-4-8 on Opus plan configs.
  • A new env var CLAUDE_CODE_REFUSAL_FALLBACK_CATCH_ALL (if set) enables a catch-all fallback for unmapped categories.
  • Per-category routing is overridable by config; unresolvable mapped targets return matched: "none" with reason "mapped_target_unresolvable".
Evidence

Category map (search for "cyber": "claude-opus-4-8"); env var (search for "CLAUDE_CODE_REFUSAL_FALLBACK_CATCH_ALL")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Observer Sidecar Persistence and Reattach#

What

Agent observers (sidecars that monitor running tasks) can now persist across session restarts and reattach to their observed task rather than always spawning fresh.

Details
  • Observer state is written to the conversation journal as "observer-ref" entries containing the observer task ID, agent type, and arming mode.
  • On a new turn, if a prior observer task ID is found in the journal and the task transcript is still readable, Claude reattaches to the existing observer rather than spawning a new one.
  • If the sidecar transcript is unreadable: "[agentObserver] reattach: observer sidecar unreadable — fresh under a new id"
  • If the prior observer's agent type doesn't match the declared type, a fresh observer is started.
  • Observer teardown now tombstones the sidecar via a stopped: true journal entry to prevent reattach loops.
Evidence

Journal type (search for "observer-ref"); reattach log (search for "observer sidecar unreadable")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Screen Reader Mode Reports Activation Source#

Feature flag
tengu_ax_screen_reader Gate removed from the code

This release deleted the gate from the code while it was still reading on for the account this site reads, so the code path no longer asks a flag before running.

This account: on · anonymous baseline: on · compiled default in v2.1.206: not a boolean we can read

These values were read against a different version of Claude Code, so treat them as the nearest reading available instead of one taken on this release.

Read once, for one account on one subscription tier, against v2.1.206. It isn't a statement about your account. What a flag value here can and cannot tell you

What

The screen reader mode status indicator now shows how the mode was activated (e.g., via environment variable or CLI flag) rather than just showing it's on.

Details
  • Old: [Accessible screen reader mode: on]
  • New: [Screen Reader Mode: on via {source}] when an activation source is known, or [Screen Reader Mode: on] otherwise.
  • The three possible {source} values are: "flag" (when the --ax-screen-reader CLI flag is passed), "env" (when the CLAUDE_AX_SCREEN_READER environment variable is set to a truthy value), and "settings" (when axScreenReader: true appears in the settings file). The priority order is flag → env → settings.
  • Activation is additionally gated by the GrowthBook feature flag tengu_ax_screen_reader. If the flag disables the feature, screen reader mode does not activate regardless of the CLI flag, env var, or settings value, and activationSource() returns undefined — so the on via {source} suffix will not appear even if an activation mechanism is present.
  • When the GrowthBook resolver is not yet available (e.g., before first-party initialization), the check defaults to true, allowing the mode to activate locally without a network round-trip.
Evidence

New format (search for "[Screen Reader Mode: on via"); activation sources (search for "activationSource()"); GrowthBook flag (search for "tengu_ax_screen_reader")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

OAuth Token Expiry Error#

What

Expired OAuth refresh tokens now throw a distinct OAuthRefreshDeadError with a clear message, rather than a generic error.

Details
  • Message: "OAuth refresh token is no longer valid; run /login to re-authenticate"
  • The error class name "OAuthRefreshDeadError" is surfaced in logs and can be caught specifically by error handlers.
  • The error is thrown at API client construction time when all three conditions hold: no API key is available, no current OAuth access token exists, and OGt() is true — meaning the stored OAuth refresh token is either an empty string "" or is found in an in-memory dead-token blacklist set (populated whenever a token refresh request fails).
  • When this error propagates to the request error handler, the displayed UI message branches by session type: interactive REPL sessions show "Login expired · Please run /login"; agent/non-interactive sessions show "Failed to authenticate: OAuth session expired and could not be refreshed".
  • The error condition is also checked in the first-party HTTP client's teleport-org auth path: after a forced token refresh attempt, if OGt() is still true, the request returns { ok: false, reason: "no-auth" } with the same detail message rather than throwing.
  • Previously, token expiry produced ambiguous error messages that did not distinguish between a missing token and a definitively-expired one.
Evidence

Error class (search for "OAuthRefreshDeadError"); dead-token check (search for "OGt()"); UI message branching (search for "Login expired" alongside "OAuthRefreshDeadError")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Login Expired Message#

What

The "login expired" notification now includes the action to take directly in the message.

Details
  • New: "Login expired · Please run /login"
  • Previously the message and the suggested action were separate.
Evidence

String (search for "Login expired \xB7 Please run /login")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Health Check Description Update#

What

The /doctor health check command description now mentions bloated memory files in addition to duplicated ones.

Details
  • Old: "Health-check your setup and fix issues: installation, unused extensions, duplicated memory files, slow hooks, updates, permissions"
  • New: "Health-check your setup and fix issues: installation, unused extensions, duplicated or bloated memory files, slow hooks, updates, permissions"
Evidence

Description string (search for "bloated memory files")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Related

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

1 entry

Bug Fixesopen

#

  • Push permission default rules (Bash(git push origin *)) now correctly add both Bash(...) and PowerShell(...) variants for dynamically resolved remotes, matching the behavior for the static origin entry. (search for "git push -u" in ANd)
  • The "scroll to bottom" button in the terminal now truncates its label to fit the available column width rather than overflowing. (search for "wrap: \"truncate-end\"")
3 entries

In Developmentopen

#

Features with infrastructure added but not yet enabled for all users.

Related

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

EndConversation Tool [Feature-Flagged]#

Feature flag
tengu_umber_kestrel Gate removed from the code

This release deleted the gate from the code while it was still reading on for the account this site reads, so the code path no longer asks a flag before running.

This account: on · anonymous baseline: off · compiled default in v2.1.206: not a boolean we can read

These values were read against a different version of Claude Code, so treat them as the nearest reading available instead of one taken on this release.

Read once, for one account on one subscription tier, against v2.1.206. It isn't a statement about your account. What a flag value here can and cannot tell you

What

A new deferred tool (EndConversation) allows the AI to end a conversation in cases of sustained, serious user abuse — after prior warnings and with strict eligibility checks.

Status

Feature-flagged (requires server-side feature flag to enable) and model-gated (opus 4-8, sonnet 5, fable 5, mythos 5).

Details
  • The tool is shouldDefer: true and only enters a model's tool list after an explicit ToolSearch("select:EndConversation") call; its searchHint is "end the conversation — only for sustained user abuse, or when the user explicitly asks to see it demonstrated." The GrowthBook feature flag tengu_umber_kestrel must be enabled server-side; if the flag is off, isEndConversationToolEnabled() returns false, the tool is disabled, and the deferred hint is suppressed entirely.
  • Model floor check (modelMeetsEndConversationFloor): parses the model ID as claude-{family}-{version} and compares version numerically against [["opus", [4,8]], ["sonnet", [5]], ["fable", [5]], ["mythos", [5]]]. Models below these thresholds (e.g., opus-4-7) never receive the deferred hint and cannot enable the tool.
  • In fork/background-task contexts (agentId is set on the tool context), the call immediately returns ended: false with a message stating forks cannot end the main conversation and instructing the fork to return welfare concerns as output instead. No session state is changed.
  • Two-call confirmation via lastAssistantTurnCalledEndConversation: scans messages backward from newest; if any assistant turn contains an EndConversation tool-use block and no real user message has appeared since that turn, the check returns true and the tool proceeds to end. If no such prior call exists (first invocation), the tool returns ended: false with the full reflection prompt plus guidance, asking the model to call again to confirm.
  • Before terminating, a journal marker is written to disk (oHs(Ct())); then abortController.abort("end_conversation") fires with the reason string "end_conversation".
  • In interactive (REPL) sessions: sets endedByModel: true in app state; the UI renders a styled result box displaying the tool result string: "Claude has ended this chat."
  • In non-interactive (print/pipe) sessions: calls gracefulShutdown(1, "other", { finalMessage: "Claude ended the conversation. To continue, please start a new session." }) — exits with process exit code 1.
  • Strict "must NOT use" list: stuck loops, task completion, harmful content requests (redirect instead), general frustration, any self-harm or imminent harm scenario. In crisis/self-harm cases the model must not even mention the possibility of ending the conversation.
  • A deferred-tool hint is surfaced only in interactive sessions ($B() check) to eligible models: "use only for sustained user abuse directed at the assistant, or when the user explicitly asks to see it demonstrated."
Evidence

Feature flag name ("tengu_umber_kestrel" in JUi); deferred hint (search for "Load the full guidance via ToolSearch"); end message (search for "Claude has ended this chat"); two-call check (lastAssistantTurnCalledEndConversation); model floor list (search for "Re-read the"); non-interactive exit (search for "gracefulShutdown"); model eligibility list (search for "mythos").

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

1 entry

Notesopen

#

The built-in claude_design first-party MCP server (https://api.anthropic.com/v1/design/mcp) has been removed from the auto-provisioned server list. Users relying on Claude Design should ensure it is configured via their MCP settings or the /design login flow, which routes through the updated hub command rather than a built-in server entry.

Verbatim
Official · Anthropic

Anthropic’s official release notes

Published verbatim by Anthropic for v2.1.206. Text is unmodified from the upstream changelog. Everything else on this page came out of the bundle instead, which is why the two lists don't match.

  • Added directory path suggestions to /cd, matching /add-dir behavior
  • Added a /doctor check that proposes trimming checked-in CLAUDE.md files by cutting content Claude could derive from the codebase
  • /commit-push-pr now auto-allows git push to the repo's configured push remote (remote.pushDefault, or the sole remote when only one is configured) in addition to origin
  • Gateway: /login now supports Anthropic-operated public gateway endpoints
  • EnterWorktree now asks for confirmation before entering a git worktree outside the project's .claude/worktrees/ directory
  • Background agents now upgrade to a new version in the background right after a Claude Code update, instead of paying a slow stale-session upgrade when you attach
  • Fixed an expired login failing every model with a misleading "There's an issue with the selected model" error instead of prompting to run /login
  • Fixed claude --resume and --continue not responding to keyboard input on startup
  • Fixed MCP servers configured via --mcp-config or .mcp.json ignoring a per-server request_timeout_ms, which caused long-running MCP tool calls to time out at the 60s default in fresh sessions
  • Fixed CLAUDE_CODE_EXTRA_BODY being silently ignored by claude agents / --bg background workers; the shell-exported override now follows the dispatching session
  • Fixed OAuth MCP servers requiring manual re-authentication after a single failed token refresh
  • Fixed --permission-prompt-tool pointing at an MCP server crashing with "MCP tool not found" on cold start before the server finishes connecting
  • Fixed /model picker rows printing a price for a different model than the row named, and stopped quoting first-party list prices on providers that don't bill them
  • Fixed server-provided model rows being misplaced in the /model picker when an entitlement or allowlist restriction drops the row they were positioned against
  • Fixed desktop sessions getting stuck showing "running" after a slash command was sent mid-turn
  • Fixed keyboard input being ignored in the agents view when a setup prompt appeared before a bare claude --resume on Windows
  • Fixed claude rm leaving the removed job in the daemon roster, causing the row to reappear in claude agents
  • Fixed /remote-control showing "Unknown command" when logged out — it now explains how to sign in
  • Fixed left arrow not stepping back out of a phase or agent in the workflow detail view
  • Fixed /status listing the same broken-install warning twice
  • Fixed false "disused plugin" tips and skewed disuse telemetry for LSP plugins
  • Fixed /doctor's update check to compare Homebrew installs against their cask's channel instead of the settings channel
  • Fixed the fullscreen jump-to-bottom pill suggesting Ctrl+End on macOS, not showing rebound chords, and wrapping over the transcript
  • Bedrock: fixed a multi-minute startup hang when using an awsCredentialExport helper on networks with restricted egress
  • Improved /code-review findings quality on claude-opus-4-8 across all effort levels
  • Improved agents view: status column now uses full terminal width instead of truncating at 64 characters
  • Changed agents view: Ctrl+X now permanently removes a completed session, and sessions no longer render twice; deleted background jobs stay deleted
Read it on GitHub →
System prompt

No change to the system prompt since v2.1.205.

Claude Code, interactive mode

Read the prompts for this release →