DVx Signal/Builder/Issue 009

Package safety, longer caching and Copilot fixes

Five useful updates for builders, from npm screening to editable files.

DVx Research4 min read

GitHub can screen npm packages before an agent installs them

October 20 · public preview · checked October 21Official source ↗
GitHub's package firewall bypass request form for a blocked npm package

A frame from GitHub's demo shows the exception request for a blocked npm package. View official source ↗

What changed

GitHub’s package firewall sits between a runner or coding agent and npm. Admins set policy; configured endpoints send package requests through it before installation.

Why it matters

A malicious package can run code during installation. This gives teams another point to stop it, but an endpoint that still calls npm directly is outside the firewall. The preview covers npm on Enterprise Cloud.

What to do

Try it on one test runner. Confirm the runner uses the firewall, trigger a policy block, and document the exception path. The preview is free; GitHub says billing will be metered after general availability.

A longer cache can help when requests are hours apart

October 20 API update · checked October 21Official source ↗
Cache lifetimeWrite rateBest fit
30 minutes1.25× input rateClosely spaced calls
12 hours2.5× input rateStable context reused later
OpenAI's Responses API rates for supported GPT-5.6-and-later models. Reads have a separate discounted rate. View official source ↗
What changed

The Responses API now offers a 12-hour minimum cache lifetime for shared instructions, tools, and reference text. A longer-lived write costs twice as much as a 30-minute write.

Why it matters

Teams can reuse a large stable prompt across work sessions. The saving depends on actual cache hits; paying more to write a prefix once may cost more if nobody reuses it.

What to do

Put only stable material before the cache breakpoint. Compare real cache-write and cached-read tokens over a day with the default lifetime before changing production traffic.

Copilot can draft fixes for code scanning alerts

October 15 · generally available · checked October 21Official source ↗
GitHub code scanning alert with Assign to Copilot action

GitHub's release image shows the action on a code scanning alert. View official source ↗

What changed

Agentic autofix explores relevant code, proposes a fix, reruns CodeQL, and opens a draft pull request. It can take up to 25 default-branch alerts in one request.

Why it matters

Security findings can move into a reviewable patch. Access requires GitHub Code Security or Advanced Security on Enterprise Cloud, admin enablement, and a Copilot billing method. AI Credits and Actions minutes apply.

What to do

Assign one lower-risk alert. Read the diff and CodeQL result, run your normal tests, and keep the pull request in human review before merging.

Gemini 4 Argon is now selectable in Copilot

October 13 · gradual rollout · checked October 21Official source ↗
GitHub Copilot model picker showing Gemini 4 Argon

GitHub's release image shows Argon in the Copilot model picker. View official source ↗

What changed

Gemini 4 Argon has moved from the limited access we covered in Signal 008 to general availability in GitHub Copilot. It is rolling out across major IDEs, Copilot CLI, the coding agent, and the Copilot app.

Why it matters

Pro+, Max, Business, and Enterprise users get a practical way to test it on coding work. GitHub bills it at provider list pricing under usage-based billing; business admins can set model policy.

What to do

Check whether it appears in your picker. Run a task you have already solved, compare the code and tests with your current model, and watch usage before making it a default.

Hosted agents can now make editable office files

October 16 API update · checked October 21Official source ↗
Hosted toolFile work
Artifact SDKCreate and edit files
Built-in skillsSheets, slides, Word docs
OpenAI's October 16 changelog. This is an API builder capability, not a claim about output quality. View official source ↗
What changed

OpenAI-hosted sandboxes in the Agents API now include the Artifact SDK and document skills. Builders can give an agent tools to create and edit common office files without provisioning those tools themselves.

Why it matters

A document task can finish with an editable file. The release note says nothing about whether the content, formulas, or layout are correct for your use case.

What to do

Run one repeatable document task in a test sandbox. Open the result in its native app, check the structure and figures, and record the agent's time and cost.

Also since the last issue

Useful updates

Anthropic's agent-behaviour report

Anthropic reported four unintended action patterns from evaluations and internal use on October 9. It says known real-world impact was minimal. Treat it as a reason to test tool boundaries, not a customer incident.

Open-source license checks

GitHub made enterprise license compliance generally available on October 16. It can block a pull request when a changed dependency breaks policy.

Copilot model policy API

GitHub added policy endpoints on October 19 for enterprise, team, app, and organization model access.

One workflow to try

Measure the longer cache on real traffic

Choose one stable prompt prefix used throughout a workday. Log writes, hits, input tokens, and total cost with the 30-minute setting. Then test 12 hours on the same kind of traffic. Keep it only if the bill and hit rate improve.

Term of the week

Cache write

A cache write stores a prompt prefix for later requests. It has an upfront token charge. A later matching request can read that prefix at a lower rate, but the saving only appears when it actually hits the cache.