Correction, October 13: Hook blocking depends on the event, Desktop transcript exemptions have version and policy limits, and Vercel’s retention policy protects some active deployments and documents conditional recovery.

Claude Code can stop a tool call when a before-action safety check fails, if you enable the new setting. Vercel gave existing Pro teams an October 23 deadline to keep older deployments available for rollback.

Coding Agents: Broken checks can stop actions

Claude Code hooks can block when they break

A hook is a small script Claude Code runs around an action, such as a command or file edit. Builders can use a hook before an action to refuse risky changes. Version 2.1.295 adds onFailure: "block" for command and HTTP hooks, which call scripts or web services.

According to Anthropic’s changelog, this setting can block an action if its hook cannot start, times out, or exits unexpectedly. Without it, a broken check can let the action through. For a PreToolUse hook, which runs before a tool call, the failure blocks that call. It cannot undo an action through a hook that runs afterward.

The hooks reference also excludes background command hooks and certain completion events. Add the option only to a supported event that can enforce the check you need.

Ask your agent to configure it on checks that protect secrets or production data, then test in a disposable project. Point the test hook at a missing script and request a harmless command covered by it. Confirm that the hook blocks the command when its check fails. Mods run without a separate sandbox, so plugin trust remains a separate concern.

Cursor adds phone access to local agents. Its October 6 update lets the iOS app list and message agents running on your computer. They stay on that computer, which must remain on and online. Remote Control is enabled by default outside Enterprise organizations. Pairing requires account sign-in and approval in the desktop app.

Replying from a phone can unblock work while you are away from your desk. Review pairing requests and investigate unfamiliar computers listed on your account. Cursor’s keep-awake setting requires a plugged-in computer with its lid open. Local execution alone does not establish that all data stays on your machine.

Local transcripts have their own retention policy. Anthropic’s data usage docs give local Claude Code transcripts a 30-day default, adjustable with cleanupPeriodDays. In Claude Code v2.1.248 and later, sessions started or most recently continued in Claude Desktop or Cowork are exempt by default. Managed cleanupPeriodDays settings and the HIPAA configuration, used for regulated health information, impose the ordinary limit instead.

Local transcripts are plaintext files, so keeping them longer also preserves sensitive text you pasted. Save durable project decisions separately and avoid treating this local setting as Anthropic’s server retention policy.

Claude Haiku 5.5 reaches helper tasks. It arrived October 7 in Claude Code, paid GitHub Copilot plans, and Vercel’s AI Gateway. Haiku 5.5 brings lower-cost helper tasks, with Copilot access rolling out gradually.

Copilot local sandboxing is generally available. GitHub’s October 7 announcement covers the command-line tool, Copilot app, and VS Code sessions using Agent Host. A sandbox restricts what agent-run commands can reach, including files, networks, and credentials. It costs nothing extra, and enterprise policies can require it.

Copilot app sandboxing was off by default during its September preview. Check your current project’s policy rather than assuming general availability enabled it. Confirm a harmless file outside the allowed folders stays inaccessible.

Infra & Deployment: Rollback history and spending limits

Vercel gives existing Pro teams a retention deadline

Vercel’s October 9 announcement extends Deployment Storage and Functions Storage billing to all Pro teams at $0.10 per gigabyte of stored data each month. Your team’s email gives its billing start date. The deployment retention window also changes to 30 days, meaning Vercel keeps a shorter history of saved app releases.

Deployments older than 30 days will be deleted starting October 23 unless you opt out in retention settings first. Settings already at 30 days or shorter remain unchanged. Permanently removed deployments cannot support rollback, switching your live app back to an earlier working release.

Vercel’s retention policy exempts currently aliased production releases and certain recent or active branch deployments. An alias is the address pointing visitors to a release. The policy describes conditional recovery for successfully built deployments marked for deletion, while the October 9 announcement warns against relying on restoration. Opt out before the deadline to preserve the history you need.

Review each deployment category’s retention setting before October 23. For an app you rarely update, a fault discovered six weeks later could require rebuilding an old version from code. Weigh the history you need against storage charges, and keep the code needed to reproduce a release.

Ultrafast charges six times the standard rate. Vercel’s AI Gateway update adds faster output for OpenAI’s GPT-6 Astra and GPT-6.1 Sol. AI Gateway routes your app’s model requests to providers. The premium is easier to justify when someone waits for each response than for an unattended summary.

Requests pinned to unsupported processing regions, such as the EU, fall back to standard processing and pricing. The request’s region setting matters, not the user’s location. Compare time saved and total cost on your actual flow. Set a budget before repeated runs, allowing for the request that crosses it.

A spending cutoff also needs a restart plan

Simon Willison called for default hard budget caps on October 3, limits that stop paid usage instead of only sending warnings. Agents make paid services easier to start, while an alert leaves spending running until someone intervenes. The controls he points to have different scopes and consequences.

AWS’s project spend limits require a paid plan and are rolling out to limited customers. Reaching the limit pauses the project and stops its resources. AWS says it permanently deletes project data if you take no action within 90 days. Google Cloud’s Spend Caps are in public preview for specific services, not the whole account.

Google supports the Gemini API, Agent Platform, Cloud Run, and Cloud Run Functions, with one project and service per cap. Its current documentation says usage resumes when you lift the block or the monthly period resets. Persistent-resource charges can continue, and enforcement delays can cause overages.

Spending cutoffs differ in what they pause and how usage resumes. Check which service stops and how to restart it before depending on a cutoff for a live app.

Vercel’s gateway budgets allow the threshold-crossing request to finish, so the bill can exceed the configured figure. Budget alerts require separate setup. For a customer-facing feature, decide what users see when requests stop and who receives the alert.

A cap can also fail incorrectly. On October 7, Anthropic reported organizations incorrectly paused for reaching a spend limit. It marked the incident resolved after requests were refused across the Claude API, Claude.ai, and Claude Code.

Supabase makes backend settings readable by agents. Table layouts and settings can live as project files, so an agent can inspect and change them. Supabase’s Select announcement describes experimental schema tooling, alpha local testing without Docker, and private-alpha Compute services.

Docker runs software in isolated environments on your computer; removing that requirement helps agents test database changes locally. A generated database change can still remove data. Read the change script before applying it, and use a disposable database to rehearse. Confirmation prompts cover a separate agent connection; do not assume they guard file-based changes.

Domain purchases still require confirmation. Vercel says agents can search domains, check prices, and initiate a purchase through its command-line tool. Non-interactive mode, without a person answering prompts, returns an error and hands confirmation back to you. Review the price, renewal choice, and contact details for the person or organization registering the domain.

AI Builders: Check chat access before inviting teammates

v0’s October 6 changelog says personal accounts are now team workspaces. Chats, projects, and settings moved automatically. Its September 10 update established team visibility as the default for new team chats. That earlier sharing change matters when a migrated solo workspace gains its first teammate.

Owners can choose Private under Settings, then Workspace, then Default Chat Visibility. The workspace default applies only to new chats. Existing chats keep their current visibility, which creators can change in each chat’s Invite dialog. Team visibility is distinct from publishing an app on the web.

Check the chat-visibility default in your migrated workspace before inviting a freelancer or client. Choose Private for work that should stay restricted, and inspect older chats separately. If a shared chat contained a secret, replace that credential.

Workflow: Access debates and interrupted releases

Model restrictions could reach hosted access too. Nathan Lambert argues that restrictions on downloadable models must account for capabilities still available through hosted providers. Different visibility into misuse weakens comparisons between open and hosted models.

GitHub Actions failures interrupted releases. GitHub’s resolved October 5 incident affected Actions and hosted runners, the machines executing automated checks and deployment jobs. During the worst roughly 90 minutes, 47% of workflow runs on GitHub-hosted runners failed. Self-hosted runners were unaffected, and existing GitHub Pages sites kept serving.

A failed deployment job does not necessarily mean the live site went down. Check the production version and the failed job before retrying. Confirm required checks complete on the intended commit before releasing; an interrupted run is not a passed review.

What was noise

Codex 0.162.1 fixes a terminal-interface crash on multiline questions and a startup mismatch with a background server. Update if either failure affects your sessions. These fixes do not require redesigning a working agent workflow.

What to watch

  • October 23: Vercel starts deleting older Pro deployments under the changed retention window unless you opt out first.

  • Copilot secret checks: GitHub’s billing notice plans AI Credit charges for new opt-in secret checks. Push protection is in private preview for eligible GitHub Teams and Enterprise Cloud customers with paid Secret Protection or Advanced Security. New secret checks inside the existing security-review command are coming soon. AI-detected alert scanning remains included with those security products at no added charge.

  • Supabase availability: Watch the experimental and alpha features before depending on them for a live backend.

End of article