Daily brief
Copilot CLI's Sandbox Lets Localhost In, and Claude Code Ships a Fix Only Build
By Steven Sorensen · 19 Oct 2026

At a glance
- Copilot CLI test builds let sandboxed commands reach services on your own machine and network.
- Those builds also add a switch to stop the CLI uploading client telemetry.
- Claude Code's newest build is fixes only, and the slower channel is far behind it.
- VS Code is moving new chats to the Copilot harness by default.
Sources are thin today. The weekend produced one new main channel build with no listed changes, and our other lead rests on prerelease notes GitHub published Wednesday to Friday, which have not reached Copilot CLI latest. Treat today as a defaults and channels check: test what the Copilot CLI sandbox can reach on your network before these builds ship, decide which Claude Code channel each machine follows, and look at the harness name on your next VS Code chat.
Top Stories
1. Copilot CLI prereleases let the sandbox reach localhost and your private network
Copilot CLI latest has held at v1.0.98 since 3:53 PM ET Wednesday 14 October. On the npm prerelease channel, GitHub shipped six builds between 8:59 PM ET that Wednesday and 6:46 PM ET Friday 16 October, ending at v1.0.100-1. Everything here comes from those builds' release notes on GitHub's releases page, not the CLI's changelog file.
The headline is in v1.0.99-0: the sandbox now allows local network access by default, so commands can reach localhost and private network services. The same build lets /sandbox keep sandboxing on for the current session without enabling it globally, installs and updates trusted repository plugins automatically before chat runs, and applies path specific instruction files under .github/instructions when the agent reads or edits a matching file.
The later builds add control. v1.0.99-1 lets sandboxed git on github.com use a GITHUB_TOKEN configured in credentials.envVars, and adds --trust-folder for one off prompt mode sessions started with -p or piped input. v1.0.99-2 shows a notice when administrator managed settings override your selected model and turns on semantic code search by default for online sessions. v1.0.99-3 adds --disable-client-telemetry, and v1.0.100-0 stops sessions stalling for minutes when an MCP server hangs while loading tools.
Practical dev impact: If you count on the sandbox to keep an agent away from a local database, a dev server or an internal admin page, that boundary loosens by default once these builds reach latest. GitHub has not said when that happens or whether every item ships unchanged, so in our view the useful move is to test on a spare machine from prerelease now and write down which local services a sandboxed command can reach.
2. Claude Code v2.1.303 is a fix only build, so the real choice is your channel
Claude Code v2.1.303 reached npm latest and next at 5:15 PM ET Sunday 18 October. Its changelog entry reads, in full, "Bug fixes and reliability improvements." The news is the gap: at our 5:51 AM ET check Monday, stable still served v2.1.292, published 1:10 PM ET Tuesday 6 October, and nine published releases are newer than stable, including v2.1.302 with the subagent and permission mode defaults Saturday's brief covered.
Anthropic's setup docs describe two values for autoUpdatesChannel. "latest", the default, gets new features as soon as they are released. "stable" uses a version that is typically about one week old and skips releases with major regressions. You can set it in settings.json or through /config under Auto-update channel.
The second key is minimumVersion, a floor that background auto-updates and the manual update command refuse to go below, so moving to "stable" does not downgrade a machine already on a newer "latest" build. Switching from "latest" to "stable" in /config asks whether to stay on your current version or allow the downgrade, and choosing to stay sets minimumVersion for you. Switching back to "latest" clears it.
Practical dev impact: Two teammates on different channels can see different behavior from the same prompt this week, because one has Friday's v2.1.302 defaults and the other does not. Our advice is to pick one channel for the team on purpose, and if you move machines to stable, let /config set the floor rather than typing a version you have not checked.
VS Code v1.142, dated Wednesday 14 October in its release notes, which give no time, is also worth one look apart from the policy changes Sunday's brief covered. Microsoft says it is progressively rolling out the Copilot harness, built on the Copilot SDK and run through Agent Host, as the default harness for new sessions, and chat titles now name the active harness, such as Copilot or Codex. Windows sandboxing for terminal commands in Local agent sessions is deprecated in favor of Agent Host sessions. In our view, check the chat title before you compare behavior across machines or file a bug.
Practical Impact Analysis
Both lead stories are about defaults that move without anyone on your team choosing them. The Copilot CLI change carries the security weight. A sandbox that reaches localhost and private network services by default changes what a sandboxed command can touch on a developer laptop. Because it is all still on prerelease, in our reading this is a week to test, not to roll out. Claude Code v2.1.303 changes nothing you can see, so its value is as a prompt to settle the latest versus stable question before the next feature build widens the gap.
Elsewhere, at our 5:51 AM ET check Monday, Codex latest held at v0.163.0 with alpha at v0.164.0-alpha.3 from 12:14 AM ET Saturday, whose GitHub release has no notes. Gemini CLI held at v0.64.0 with preview at v0.65.0-preview.0, and its 9:43 PM ET Sunday nightly carries the same commit hash as Friday's and Saturday's, so in our reading it adds no new code. Cursor's main changelog still tops out at Friday's cloud agent canvases update and its CLI changelog at v2026.10.08, and the GitHub Changelog had no posts after 10:00 AM ET Friday.
Tutorial
We have not run this script end to end on every platform, so read it before you run it. It prints versions and changes nothing; installs and the settings example are comments only.
- Run steps 1 and 2 and compare the pointers with this brief.
- To test the Copilot CLI sandbox, uncomment the step 3 install on a spare machine only.
- Inside a Copilot CLI session on that machine, try a sandboxed command against one local service you expect to be off limits.
- For Claude Code, read step 4 and pick one channel. Prefer
/configfor the switch.
# We have not run this script end to end on every platform. Nothing below installs or changes anything unless you uncomment it.
# 1. Copilot CLI pointers. At 5:51 AM ET Mon 19 Oct npm showed latest v1.0.98, prerelease v1.0.100-1.
npm view @github/copilot dist-tags --json
if command -v copilot; then copilot --version; else echo "copilot not on PATH"; fi
# 2. Claude Code pointers. At 5:51 AM ET Mon 19 Oct npm showed stable v2.1.292, latest and next v2.1.303.
npm view @anthropic-ai/claude-code dist-tags --json
if command -v claude; then claude --version; else echo "claude not on PATH"; fi
# 3. Optional, spare machine only. Copilot CLI prerelease channel, npm installs only, at 5:51 AM ET Mon 19 Oct:
# npm install -g @github/copilot@1.0.100-1
# Flags from the prerelease notes: --trust-folder (v1.0.99-1), --disable-client-telemetry (v1.0.99-3).
# 4. Optional, Claude Code, npm installs only. Pick one channel, at 5:51 AM ET Mon 19 Oct:
# latest and next channel:
# npm install -g @anthropic-ai/claude-code@2.1.303
# stable channel:
# npm install -g @anthropic-ai/claude-code@2.1.292
# Settings example from the Claude Code setup docs, not applied by us.
# Set the floor to the version this machine already runs:
# {
# "autoUpdatesChannel": "stable",
# "minimumVersion": "2.1.303"
# }
Homebrew installs choose a Claude Code channel by cask name, and apt, dnf and apk installs choose by repository, so neither reads autoUpdatesChannel. Do not mix npm and native installs of Claude Code on one machine.
Correction: our Oct 15, Oct 16 and Oct 17 briefs said Copilot CLI prereleases had no changelog notes. They do have release notes.
Recommended AI prompt
Copy this paragraph into ChatGPT, Claude, Gemini, Grok, or whatever AI you use. You are my developer tooling lead for Monday 19 October 2026. Copilot CLI prerelease builds v1.0.99-0 to v1.0.100-1, on the npm prerelease channel between 8:59 PM ET Wednesday 14 October and 6:46 PM ET Friday 16 October and still behind latest v1.0.98, let sandboxed commands reach localhost and private network services by default, let /sandbox stay on for one session, let sandboxed git use a GITHUB_TOKEN from credentials.envVars, and add --trust-folder and --disable-client-telemetry. Claude Code v2.1.303, on latest and next since 5:15 PM ET Sunday 18 October, is a fix only build, while stable still serves v2.1.292, and autoUpdatesChannel with minimumVersion controls which one a machine follows. VS Code v1.142 is rolling out the Copilot harness as the default for new sessions. Ask me which local services our developers run, whether we use the Copilot CLI sandbox, and which Claude Code channel and install method each machine has. Then give me a short checklist to test the Copilot CLI sandbox against our local services, choose one Claude Code channel for the team, and check the harness name in VS Code chats, and link me to the official release notes or docs for each step instead of inventing names.
New: The AI-Dev Prompt Pack, 40 tested prompts for Claude Code, Cursor, and Gemini CLI.
See the Prompt Pack