Codex Learns to Run on Vertex AI, Claude Code Stops Overwriting Big CLAUDE.md Files, and Copilot Agents Get Their Own Org Identity

At a glance

• Codex now ships a built-in Google Cloud Vertex AI provider that you pick with one config key.
• Fresh Codex profiles start on memory v2, and effort can change mid-conversation on supported models.
• Claude Code no longer overwrites a very large CLAUDE.md or rules file after reading only its start.
• Copilot SDK agents can authenticate as an organization's GitHub App in public preview, with usage billed to the org.

The common thread today is cloud accounts and identity, because each release changes which provider or credentials an agent leans on. OpenAI published Codex at 4:09 PM ET Tuesday, Claude Code reached npm at 2:36 PM ET with release notes following at 4:24 PM ET, and GitHub's announcement landed at 5:36 PM ET.

Each release rewards a short check before its new defaults reach a shared setup, which is why the tutorial below checks first and leaves every install commented out.

Treat today as an identity and provider check day: see whether Codex can use the Google Cloud project you already have, update Claude Code if your team keeps big instruction files, and plan to move any shared Copilot SDK bot onto a GitHub App.

Top Stories

1. Codex v0.165.0 adds a built-in Vertex AI provider and turns on memory v2

OpenAI published Codex v0.165.0 at 4:09 PM ET Tuesday, and npm moved latest to it at 4:14 PM ET. The headline change is a built-in Google Cloud Vertex AI provider, configured with model_provider = "gcp". GPT-6.1 Terra also joins the Codex model catalogs, and the session picker gains a repository filter that makes sessions easier to find across linked worktrees.

Two defaults change as well. Reasoning effort changes during a conversation are now on by default for supported models, and memory v2 is the default for fresh profiles, with support for importing memories from other agents into it. The fixes matter if you run Codex in a shared setup. Fresh TUI sessions now honor the server's default approval reviewer when nothing overrides it, queued agent messages no longer restart work after an interrupt or a switch to Plan mode, Codex stops retrying when a provider reports that a model was not found, and a marketplace URL validation gap that could let Git connect to a different host than the one checked is closed.

Practical dev impact: If your org buys models through Google Cloud, try model_provider = "gcp" in a scratch profile before you touch shared config, because the release notes name the key but say nothing about project, region or auth. Our read: the notes say memory v2 is the default for fresh profiles, so existing profiles likely keep the memory setup they have until you change it.

2. Claude Code v2.1.312 stops Write from overwriting large CLAUDE.md files

Claude Code v2.1.312 reached npm latest and next at 2:36 PM ET Tuesday, while stable stays on v2.1.308. The fix that protects real work covers a very large CLAUDE.md, rules file or @-included file in a subdirectory. Claude could overwrite one after seeing only its start, and Write now asks Claude to read the whole file or edit it in place. Per-directory and paths: skills are also found again after a Bash cd deeper into a git repository.

Enterprise users get changes too. On cloud providers including Vertex and Foundry, model names Claude Code doesn't recognize, such as custom deployment names, now get the shorter system prompt even when they front Sonnet or Haiku, and setting CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0 keeps the full prompt. The Claude apps gateway adds oidc.user_id_claim, so you can identify people by an id_token claim other than sub, such as Microsoft Entra ID's oid. Prompt and agent hooks now run at most 24 at a time per session, sub-agents included, and managed settings are fetched again for the rest of a session after you sign in from the VS Code extension.

Practical dev impact: Teams with large shared CLAUDE.md files should update soon, since the old behavior could replace a file Claude had barely read. If you call Claude through a custom Vertex or Foundry deployment name, compare a few answers after updating and set CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0 if the shorter prompt changes behavior you rely on.

3. Copilot SDK agents can now authenticate as an organization's GitHub App

GitHub announced at 5:36 PM ET Tuesday that Copilot supports server-to-server authentication for GitHub Apps installed on organizations, in public preview. Services and agents built with the Copilot SDK can make requests on behalf of an organization without a signed-in user's credentials. You give a GitHub App the read-only Copilot Requests organization permission, install it on the org, and pass a short-lived installation access token to the SDK.

Usage is attributed and billed to the organization that owns the installation, and the org's model policies, budgets and spending controls apply. Owners can review which apps have access, find app usage in AI usage CSV reports through the app_client_id field, and stop an app by suspending or uninstalling it. GitHub frames this as a workload-owned identity, so automated services stop tying their authentication or usage to one person.

Practical dev impact: A review or triage bot running on one engineer's token is the obvious candidate to move to a GitHub App with the Copilot Requests permission. Because this is a preview, start with one internal bot and confirm its usage shows up under app_client_id before you migrate the rest.

Practical Impact Analysis

These three releases belong together because each one moves an agent's provider or identity closer to the organization and away from one person's setup. Codex gains a built-in path to Vertex AI, Claude Code changes how it treats the custom deployment names enterprises use on cloud providers, and Copilot lets an SDK agent act as a GitHub App with org budgets attached. Codex leads because a built-in provider can change where a whole team's model traffic goes. Our read: the Claude Code fix is the most urgent for anyone whose CLAUDE.md has grown large, since it guards against silent data loss rather than adding a feature.

Elsewhere at our check, Gemini CLI v0.66.0 reached latest at 4:30 PM ET with terminal rendering fixes and opt-in private network endpoints for MCP OAuth flows, while preview moved to v0.67.0-preview.0. Copilot CLI v1.0.105 reached latest at 8:44 PM ET with one fix, so picking Copilot Fusion High or Max in /model now works from long-context sessions.

Tutorial

Read the script before you run it. It only prints and checks unless you uncomment a line.

1. Run step 1 on any machine with Codex or Claude Code.
2. Run step 2 from the root of a repo that has CLAUDE.md files.
3. Read step 3 before you add anything to a Codex profile.
4. Do the Copilot check in your browser, as described after the script.

# We have not run this script end to end on every platform. It only prints and checks unless you uncomment a line.
# 1. Which versions are out, and which do you have? At 5:59 AM ET Wed 28 Oct npm showed
#    @openai/codex latest v0.165.0 and @anthropic-ai/claude-code latest v2.1.312, stable v2.1.308.
npm view @openai/codex dist-tags.latest
npm view @anthropic-ai/claude-code dist-tags --json
command -v codex >/dev/null && codex --version || echo "no codex CLI on PATH"
command -v claude >/dev/null && claude --version || echo "no claude CLI on PATH"
#    Optional npm installs only. Pinned at 5:59 AM ET Wed 28 Oct, latest channel:
# npm install -g @openai/codex@0.165.0
# npm install -g @anthropic-ai/claude-code@2.1.312

# 2. List CLAUDE.md files with their sizes in bytes. These are the files the v2.1.312 Write fix protects.
find . -name CLAUDE.md -not -path "*/node_modules/*" -exec wc -c {} +
echo "CLAUDE.md size check done"

# 3. Does your Codex config already set a provider? Prints the file name only if it does.
grep -l "model_provider" ~/.codex/config.toml 2>/dev/null || echo "no model_provider set in ~/.codex/config.toml"
#    Codex v0.165.0 selects the built-in Vertex AI provider with the key below.
#    Try it in a scratch profile first. The release notes skip setup; PR #56265 says it reads
#    GOOGLE_CLOUD_PROJECT and GOOGLE_CLOUD_LOCATION and uses gcloud application-default auth.
# model_provider = "gcp"
echo "Done. Nothing was installed or changed."

For Copilot, open your organization's GitHub App settings, check which apps already hold the Copilot Requests permission, and compare that list with the bots you expect. Then look for the app_client_id field in your next AI usage CSV report.

Sources:

• Codex v0.165.0 release
• Codex on npm
• Codex PR #56265: built-in Google Cloud Vertex AI provider
• Claude Code v2.1.312 release
• Claude Code on npm
• GitHub Copilot supports server-to-server authentication
• Gemini CLI v0.66.0 release
• Copilot CLI v1.0.105 release

Recommended AI prompt

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