Claude Code v2.1.268: Fixes HTTP 400 on Third-Party Endpoints, WebFetch Hangs, and Secret Leaks
Claude Code v2.1.268 is out, and it's mostly a bug-fix release with a few gateway and plugin additions. The headline item for anyone not pointing at Anthropic's own API: every request on third-party Anthropic-compatible endpoints was failing with HTTP 400 since v2.1.265.
The HTTP 400 regression
If you're running Claude Code against a proxy or self-hosted endpoint via ANTHROPIC_BASE_URL, every turn has been failing since 2.1.265. The cause: a regex in the Artifact tool's input schema that those endpoints reject. If you skipped 2.1.265–2.1.267, upgrading here is the fix.
WebFetch no longer hangs forever
WebFetch would hang indefinitely on a server that kept the response open without finishing. Fetches now fail after 300 seconds. Override with an env var:
CLAUDE_CODE_WEBFETCH_DEADLINE_MS=<ms> # 0 turns the deadline off
Gateway additions
pricing:ingateway.yamlnow propagates to signed-in Claude Code clients via managed settings, so/costand telemetry line up with the spend meter.- Startup warning when
access_control.allow_cidrsis empty, plus a one-time warning the first time a request arrives from a public address. - New
gatewayInternalNetworksmanaged setting lets admins allow/loginto a Claude apps gateway on their org's own public IPv4 block.
Secrets no longer leak in errors
- Plugin and marketplace errors no longer print a token or password from a git source URL.
/mcp,/pluginserver details,claude mcp list/get, and MCP login errors no longer display secrets resolved from${VAR}placeholders in MCP configs.
Permission and tooling fixes
- Deny/ask rules on symlinked directories (
/etc,/tmp,/varon macOS;/binon Linux) were not applying when a path was given by its real location. Bash commands also ignored deny rules written on a symlinked path spelling. - Read/Edit deny rules were skipped when
env -C,eval, or a similar unanalyzable command sat on the same line. - A respawned in-process teammate could pick up tools or a system prompt from a same-named agent file in an untrusted folder.
Session and CPU fixes
- A busy loop in long-running idle sessions no longer pins a CPU core; rapid terminal focus reports during a session recap no longer keep CPU high.
- SDK sessions using
excludeDynamicSectionsno longer break prompt caching and extended thinking mid-session — the first message isn't re-rendered each request. /compactand auto-compact summaries no longer mangle text containing$sequences.- Restored-file notes from a
/compact-ended conversation now load in the same order on every resume.
JSON output and runner flags
--jsonadded toclaude plugin install,uninstall,update,enable,disable, witherrorDetails/noteDetailsper row ofclaude plugin list --json.configDirectoryadded toclaude auth status --json.claude self-hosted-runner --remove-session-state(default off) deletes per-session directories under<base-dir>/_sessions/when a session ends.
Also fixed: MCP OAuth sign-in failing with "No available ports for OAuth redirect" when the local callback port range can't be bound, workload identity federation via a profile failing mid-run with 401 … jti reused, and stale model-access denials telling entitled users a model is restricted.
📖 Read the full source: GitHub Claude-Code
👀 See Also

Claude for Word Add-in Evidence Found in Analytics API
Anthropic's analytics API now returns metrics for Claude for Word alongside existing Excel and PowerPoint add-ins, indicating the Word integration is in development. The API shows zero usage counts for Word, suggesting it's not yet publicly available.

Claude Opus 4.5 and Sonnet 4.5 removed from /model selection, require launch flag
Claude Opus 4.5 and Sonnet 4.5 are no longer available in the /model selection menu during sessions. Users must now start sessions with the --model flag specifying the full model ID to access these older versions.

Polsia Platform Shows Repetitive SaaS Patterns in Live Founder Launches
Polsia is an autonomous business platform where users describe their business, pay money, and it executes autonomously. A behavioral scientist observed 72 hours of live founder launches, identifying repetitive patterns like AI SDR automation solutions and underserved international markets.

Sarvam AI releases 30B and 105B open-source LLMs with Indian training infrastructure
Sarvam AI has open-sourced Sarvam 30B and Sarvam 105B, two reasoning models trained from scratch in India on compute provided under the IndiaAI mission. Both models use Mixture-of-Experts architecture with sparse expert routing and are optimized for efficient deployment across hardware from GPUs to laptops.