Daily brief

Gemini Loads Tool Schemas Only When a Task Needs Them, Claude Code Moves Its Stable Channel Forward, and GitHub Enterprise Server Finds Secrets Without Leaving Your Network

By ยท 21 Nov 2026

A blank cream dock board with three blank butter yellow, pale aqua and white modules seated in its slots and empty slots left open, seen straight down on a pale apricot table, with no people or text.

At a glance

  • Tool Retrieval is in preview in the Gemini Interactions API, so an agent can keep a large tool library on standby and load schemas only when the current task needs them.
  • The Claude Code stable channel moved forward, and the newest latest build adds organization sign in and fixes long freezes after compaction.
  • The GitHub Enterprise Server release candidate adds AI secret detection that runs on the appliance's CPU with no outbound connection, plus a default Dependabot cooldown.
  • Copilot code review is generally available for Azure Repos. Codex, Gemini CLI and Copilot CLI stay on their current stable releases, and both Cursor changelogs show nothing new.

Google published the Tool Retrieval launch post at 1:00 PM ET Thursday, Claude Code v2.1.330 reached npm at about 2:16 PM ET Friday, and GitHub posted the Enterprise Server release candidate at about 3:41 PM ET.

Treat today as a weekend check in day: try Tool Retrieval on one Gemini agent that already has a long tool list, find out which Claude Code channel your machines follow, and put the Enterprise Server release candidate on a test appliance if you run GitHub on your own hardware.

Top Stories

1. Gemini can load tool schemas only when a task needs them

Google has put Tool Retrieval into preview in the Gemini Interactions API. The launch post says that when an agent connects to dozens or hundreds of tools, sending every full parameter schema on every request inflates input token counts, adds latency, and lowers tool selection accuracy as the list grows. Tool Retrieval keeps the library on standby and loads schemas only when the current task needs them.

To turn it on, you add the API's tool search type to the request and mark the tools you want to hold back. The model still sees the name and description of every deferred tool, but the parameter schemas stay out of context until it asks for a tool by name. The API then loads the full schema and the model calls the function, all inside the same request. Google says Gemini runs the search on the server, with no index to build or host. You can keep your most used tools loaded up front and defer the long tail, and the same mark on a remote MCP server defers every tool it exposes.

The docs say it is in preview, requires the v1beta API version, works only in the Interactions API rather than the older generate content method, and that the API surface may change before general availability. Loaded tools stay in context for the rest of the conversation and cannot be unloaded. Google says there is no separate charge and tokens are billed at the standard rates for the model you call, but the docs add that each server side search triggers an extra model call billed at the input rate, and that with small libraries or very long contexts that search on every turn, loading tools up front can cost less. The launch post says it works with Gemini 3.6 and later models, including all Gemini 4 models, while the docs list Gemini 4 Carbon, Neon, Hydrogen and Argon, plus Gemini 3.8 Flash and Gemini 3.6 Flash.

Practical dev impact: If your Gemini agent already carries a long tool list, try Tool Retrieval on a test project and compare token use before changing production traffic. Stay on the Interactions API and the preview API version the docs name, and keep tools you mention by name loaded from the start.

2. Claude Code moves its stable channel forward as v2.1.330 lands

Between our Friday morning check and this morning, the npm stable tag for Claude Code moved from v2.1.311 to v2.1.323, while the latest tag moved to v2.1.330 on Friday afternoon. Anthropic's setup docs describe stable as a version that is typically about one week old and skips releases with major regressions, and you choose your channel in the config menu.

The changelog says Claude Code no longer freezes for minutes after compacting a very long conversation, the worktree exit dialog no longer deletes a worktree that another running session is using, and settings are no longer lost when several Claude Code processes save them at the same time. It also adds sign in to organizations managed in the Anthropic Enterprise Admin Console, after which Claude Code applies the organization's settings and its cloud provider or gateway.

Practical dev impact: Find out which channel your shared machines follow before you debug a change nobody expected. If you run parallel worktree sessions, try v2.1.330 on a test machine for the worktree fix.

3. GitHub Enterprise Server v3.23 release candidate brings AI secret detection inside your network

GitHub released a release candidate for GitHub Enterprise Server v3.23. For enterprises with GitHub Secret Protection, it adds AI secret detection for unstructured secrets that runs entirely inside your network, including environments with no internet connection. GitHub says a self contained tuned model runs on the appliance's CPU with no GPU, no hosted model and no outbound connection, and the feature is in public preview.

The release also widens malware advisories beyond npm to PyPI, Maven, RubyGems, NuGet, Go, Rust and Composer, and Dependabot now waits until a package version has been on its registry for at least three days before opening a version update pull request. Access tokens that expire after eight hours, with refresh tokens, are now the default for new OAuth apps and opt in for existing ones, in public preview, and every client must handle refresh tokens before you require expiration.

Practical dev impact: If your security team kept AI scanning off because it needed an outside service, plan a test of this preview on a staging appliance. Tell developers about the three day Dependabot wait before they wonder where their update pull requests went.

Practical Impact Analysis

Our read: Tool Retrieval leads because a Gemini API change reaches anyone building agents with a large tool library. Claude Code comes second because a stable channel move touches every machine on that channel at once. The Enterprise Server release candidate comes third, for teams that run GitHub themselves, but it carries the clearest security change of the day.

The thread tying them together is sending less than you used to. Gemini holds schemas back until a task needs them, Claude Code moves stable users without asking them to switch, and GitHub brings AI scanning to teams that cannot send code outside their walls. Elsewhere, Microsoft says Copilot code review is generally available for Azure Repos. Codex stays on v0.171.0, Gemini CLI on v0.70.0 and Copilot CLI on v1.0.114, so their newer alpha, preview and prerelease builds belong on test machines only. Cursor's main and CLI changelogs show no new entries since our last check.

Tutorial

  1. Open a test shell. Every command below only reads versions, public package data or local diagnostics.
  2. Script step 1 prints the Claude Code npm tags to compare with our snapshot.
  3. Script step 2 prints your Claude Code version and runs the read only doctor check.
  4. Script step 1 prints the current Codex, Copilot and Gemini CLI latest tags to compare with our snapshot.
  5. For Tool Retrieval, follow Google's official docs. This script does not call the API.
# We have not run this script end to end on every platform. It prints and checks, and changes no files.
# Run it in a test shell. Nothing is written to your shell profile.
# This script does not call the Gemini API. For Tool Retrieval, use Google's docs.

# 1. Snapshot at 5:58 AM ET Sat Nov 21 2026: claude-code latest 2.1.330, stable 2.1.323,
# codex latest 0.171.0, copilot latest 1.0.114, gemini-cli latest 0.70.0.
python3 -c 'import json, urllib.request; d = json.load(urllib.request.urlopen("https://registry.npmjs.org/-/package/@anthropic-ai/claude-code/dist-tags")); print("claude-code latest", d["latest"], "stable", d["stable"])'
python3 -c 'import json, urllib.request; d = json.load(urllib.request.urlopen("https://registry.npmjs.org/-/package/@openai/codex/dist-tags")); print("codex latest", d["latest"])'
python3 -c 'import json, urllib.request; d = json.load(urllib.request.urlopen("https://registry.npmjs.org/-/package/@github/copilot/dist-tags")); print("copilot latest", d["latest"])'
python3 -c 'import json, urllib.request; d = json.load(urllib.request.urlopen("https://registry.npmjs.org/-/package/@google/gemini-cli/dist-tags")); print("gemini-cli latest", d["latest"])'

# 2. Claude Code on this machine, then its read only diagnostics.
if command -v claude >/dev/null; then
  claude --version
  claude doctor
else
  echo "no claude on PATH"
fi

# 3. Other latest tags already printed above. Nothing is installed or changed.
echo "Compare these tags with the snapshot in the comments."
echo "For Tool Retrieval, read the official Gemini Tool Retrieval docs."
echo "Done. Nothing was installed or changed on disk."

If your Claude Code version is older than the stable tag, check the release channel in the config menu before you update a shared machine. For Tool Retrieval, follow Google's docs and compare token use on a test project before you change production traffic.

Sources:

Recommended AI prompt

Copy this paragraph into ChatGPT, Claude, Gemini, Grok, or whatever AI you use.

Prompt

You are my developer tooling lead for Saturday 21 November 2026. Story 1: Tool Retrieval is in preview in the Gemini Interactions API, requires the v1beta API version, does not work in the older generate content method, and loads parameter schemas only when the current task needs them. Google says there is no separate charge, but the docs say each server side search triggers an extra billed model call, loaded tools stay in context for the rest of the conversation, and with small libraries or very long contexts that search every turn, loading tools up front can cost less. Story 2: the Claude Code stable channel moved from v2.1.311 to v2.1.323, and the latest build v2.1.330 adds Enterprise Admin Console sign in and fixes freezes after compaction, worktree deletions and lost settings. Story 3: the GitHub Enterprise Server v3.23 release candidate adds AI secret detection that runs on the appliance with no outbound connection and a three day Dependabot cooldown. Ask which of us build Gemini agents with long tool lists, which Claude Code channel our machines follow, and whether we run Enterprise Server, then give a short checklist for each story and link the official posts instead of inventing settings or flags.