Home › Changelog › 2026-11-16
Last updated: 2026-11-16
Changelog — November 16, 2026
Migration steps and an experimental way of calling tools. Claude Code 2.1.323 adds Code Mode, an opt-in where Claude writes JavaScript to call its tools, and 2.1.325 fixes a security issue in how compressed responses are decoded. OpenClaw 2026.11.4 changes defaults you may have to act on (IRC sender lists, PostgreSQL commit durability, a renamed file-safety variable). Hermes 0.23.2 rolls up about 2,000 merged pull requests and defers its curated notes to 0.24.0. NemoClaw 0.0.144 changes how managed Hermes routing is bound, and Google's Gemini 4 Hydrogen is generally available at $1.50 input and $7.50 output per million tokens with no introductory end date.
2026-11-16
Claude Code
2.1.323 → 2.1.326 — Code Mode, plugin version ranges, a zstd security fix
2.1.323 (2026-11-12) adds Code Mode, an experimental opt-in in which Claude writes JavaScript to call its tools. Anthropic's Code Mode page covers turning it on, or turning it off for an organization; 2.1.326 then lets you turn it on while agent teams are enabled, and sessions with teams on start in Code Mode if you had already enabled it. Also in 2.1.323: engines["claude-code"] in plugin.json (plugins outside their stated Claude Code version range are not installed, updated or loaded, and install and update pick the newest version that fits), type checking of a plugin's TypeScript hooks module in claude plugin validate (each finding is a warning; --strict fails on it), CLAUDE_CODE_AUTO_MODE_MODEL in managed settings for Claude Desktop sessions on third-party providers, a 30-second progress heartbeat while hooks run long in claude -p and SDK sessions, and a startup warning for Bash allow rules with a * in a variable's value such as Bash(PORT=* npm run dev).
2.1.325 (2026-11-13) fixes a security issue in how zstd-compressed responses are decoded, so update. It also fixes subagents with their own model being moved to the plan-mode model when the model setting is opusplan or haiku, subagents on a different model sizing file reads and MCP results with the main session's model, sessions on many third-party gateways and Foundry stopping with a context-length error on every turn instead of compacting, and a saved sonnet[1m] setting being rewritten on a new install to a model id the API rejects.
2.1.326 (2026-11-16) adds a startup warning when an --allowedTools, --disallowedTools or --tools value has an entry that doesn't look like a complete rule, a message explaining why MCP servers stop loading when an administrator adds managed-mcp.json during a session, a /status row saying when your provider or an environment variable such as DISABLE_TELEMETRY turns features off, and OTEL_RESOURCE_INCLUDE_HOST (off by default) to add host.name and host.id to OpenTelemetry resource attributes. Fixes include cp being refused when only its source was covered by an Edit deny rule, and Bash commands that delete or overwrite files (find -delete, cat a > b, sort -o) being summed up in the transcript as a search or a read. There is no 2.1.324 in the release feed.
Releases →
Affects: /claude-cowork/setup/, /claude-cowork/faq/claude-code-undocumented-configuration/, /commands/
2026-11-16
OpenClaw
2026.11.2 → 2026.11.4 — changed defaults, a renamed variable, delegated change approval
2026.11.4 (2026-11-16; 220 commits, 125 pull requests, 90 contributors) lists several things to act on before or after you update (release notes). IRC: an empty sender list now denies all senders, so remove the override to inherit access or list the intended senders. PostgreSQL: writes now inherit your configured commit durability, which can add latency, and asynchronous commits can lose recent acknowledged writes after a crash. fs-safe 0.26.7: if you set FS_SAFE_PYTHON_MODE or OPENCLAW_FS_SAFE_PYTHON_MODE without an explicit native policy, replace it with FS_SAFE_NATIVE_MODE or OPENCLAW_FS_SAFE_NATIVE_MODE, keeping the same value, before updating. Telegram: resend or replace callback buttons delivered or queued before the fix. Updates: the process running the update must already include the fix, older updaters may need repair first, and a newer CLI may refuse to migrate a database used by an identified older Gateway service, in which case run openclaw update or use that service's CLI. Copilot: restart OpenClaw after an SDK update before repeating setup verification.
New and changed. A standing approval setting, autoApproveChanges, covers permitted delegated configuration changes, and changes that enable host code execution can be authorized this way, so treat it as a security decision; revoked grants are rechecked, even mid-run. Policy plugins can inspect images and block a run before they reach the model. cron.skipMissedJobs: late scheduler wakeups no longer silently skip recurring jobs, while missed slots after downtime or a clock jump stay skipped. The Control UI and TUI now wait up to five minutes for a preparing agent.
Security and privacy fixes. Page-initiated navigation to policy-blocked addresses between guarded browser actions is blocked (a same-origin exception applies to direct connections); recognized secrets are redacted in Gateway and Doctor diagnostics, client logs and OpenTelemetry attributes; paired-node screenshots and photos no longer attach to outgoing messages automatically; commands beginning with a NAME+=value assignment now prompt; and saving a skill API key no longer replaces an environment-backed or referenced key with plaintext. Doctor no longer truncates MEMORY.md after a failed legacy memory merge. We read the notes' headline sections; the page is long, so check it for your channels. No 2026.11.3 release appears in the feed, and 2026.11.5-beta.4 (Telegram widgets) is a prerelease we do not count.
Release notes →
Affects: /openclaw/, /openclaw/setup/, /openclaw/configuration/, /openclaw/security/, /openclaw/troubleshooting/
2026-11-16
Hermes
0.23.1 → 0.23.2 — a roll-up patch, curated notes deferred to 0.24.0
Hermes Agent v0.23.2 (2026-11-16) is a patch release that rolls up about 1,989 merged pull requests since v0.23.1 (7,918 non-merge commits across 5,211 changed files, 940 closed issues) into one stable tag for the GitHub release (CLI and installers), the Microsoft Store app, Docker and Hermes Cloud. Its notes say the full curated notes for this window ship with v0.24.0, so the details below come from the one paragraph the release does give.
Named as included but not documented in the release. Durable subagents that survive a gateway restart or crash and can be continued on their own conversation; one prompt-cache clock per session; hermes fallback check for fallback-chain readiness; hermes update --force to stop gateways at once instead of draining; one-time browser_vault codes entered from Telegram and Discord; opt-in shared Telegram DM threads; the Gemini Interactions API by default for Gemini 3 and later plus Gemini 4 Carbon; fast mode for Claude Mythos 5.2; per-job cron options (wrap_response, a fresh thread per run); and new plugin APIs (ctx.fork_session, ctx.redact_text(), ctx.register_idle_task). We have not yet checked these command and option names against Hermes's own docs, so treat them as the release notes' wording until the 0.24.0 notes land. A release-candidate fix also lets desktop chats created by an August build with a bare custom provider send again after updating. Release candidates rc.10 to rc.13 and a channel tag also appeared; we do not count them.
Releases →
Affects: /hermes/, /hermes/setup/, /hermes/tasks/, /hermes/memory/, /hermes/discord-gateway/
Self-hosted Claude Managed Agents launched on 2026-11-12. You can run the whole Managed Agents service on your own infrastructure, not only the self-hosted sandboxes, and store agents, sessions and session history there. The server sends model requests to the provider you choose, with Anthropic as the default, and access is by request through your Anthropic account team (docs).
Text watermark (2026-11-16). Text generated by Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 carries Anthropic's text watermark on supported platforms. A request that sets temperature, top_p or top_k to a non-default value might not carry it (docs). Claude Meetings (beta, 2026-11-12) is rolling out on Team and Enterprise plans in Claude and Claude Desktop: Claude can join a Google Meet call as a visible participant, take notes, answer questions and build artifacts such as docs, whiteboards and slides during the call. The platform notes for these dates list no new model IDs, prices or retirement dates.
API notes →
Affects: /claude-cowork/, /claude-cowork/vs-api/, /claude-cowork/projects/
2026-11-12
Models
Google — Gemini 4 Hydrogen
Gemini 4 Hydrogen (gemini-4-hydrogen) became generally available on 2026-11-12 at $1.50 input and $7.50 output per million tokens (output includes thinking tokens), cached input $0.075, with a batch rate of $0.75 / $3.75 and a priority rate of $2.70 / $13.50. Unlike Gemini 4 Carbon and Argon, Google's pricing page gives Hydrogen no introductory end date, and it describes a unified 1M-token context window with no separate long-context tier. Carbon and Argon stay at $2 / $10 through 2026-11-30 and move to $4 / $20 from 2026-12-01. Google's changelog also lists gemini-decision-1 on 2026-11-11 without a price here. On the OpenAI side, the only changelog entry since our last run is omni-moderation-csam-latest (2026-11-16), a no-cost moderation classifier for approved organizations.
Pricing →
Affects: /tools/model-pricing/, /tools/cost-calculator/
2026-11-13
NemoClaw
0.0.142 → 0.0.144 — rebuild for the new OpenClaw runtime, Switchyard bindings, MCP timeout check
v0.0.143 (2026-11-10) updates managed OpenClaw and its official plugins to OpenClaw 2026.10.5 and removes the deprecated Model Router from new onboarding. Rebuild existing sandboxes after installing it to get the new runtime (NVIDIA's Update Sandboxes guide covers upgrade and recovery); resumed onboarding rejects the retired routed and nvidia-router selections and does not migrate them, so choose a supported provider for a new attempt. OpenClaw start and recover now wait through temporary startup failures within the gateway recovery deadline.
v0.0.144 (2026-11-13) makes managed Hermes onboarding with --switchyard-config create and attach OpenShell providers for the selected hosted profiles, and a dedicated judge can replace a target that fails qualification. Published Switchyard bindings can't be reused for routing or credential changes: to change them, preserve the sandbox and onboard again under a new name. Rebuilds and installer upgrades now validate the MCP tools timeout first; NEMOCLAW_MCP_TOOLS_LIST_TIMEOUT_MS must be an integer from 1500 to 10000 ms or unset, so fix the value if a rebuild or upgrade is rejected. Deep Agents Code starts in Manual mode when its auto-approval capability is disabled, even if its native config selects Auto or Yolo, the installer prepares the OpenShell replacement before retiring the old gateway, and uninstall now reports why OpenShell refused sandbox cleanup. The repository's commit feed also shows about 20 commits since our last run (CI, backup-state and Hermes fixes); we count releases, not commits.
Release notes →
Affects: /nemoclaw/, /nemoclaw/setup/, /nemoclaw/switching-providers/, /nemoclaw/policy/
Not counted as news
Kilo Code 7.8.30 is tagged pre-release, so it is not counted; its notes say Mercury Edit 2 autocomplete is retired in the VS Code extension and Codestral becomes the default autocomplete model (saved Mercury or Inception selections move to Codestral), and the SDK drops client.kilo.edit. We will count it when a stable release carries it. OpenClaw 2026.11.5-beta.4 and Hermes release candidates rc.10 to rc.13 are prereleases. A Qwen4 Max row in the OpenRouter catalogue (2026-11-16) is not confirmed by its vendor. xAI's models page changed (it carries a daily date) while its release notes list nothing newer than 2026-11-02, and OpenAI's pricing page changed without a dated announcement; neither changes a rate we list. An OpenClaw blog post about an event in Shenzhen is not a release. IronClaw has had no release since 1.4.1.
See all releases
Browse the full changelog index for the complete history across all platforms, or the daily one-liner for the most recent state of each agent.