Daily brief
The Session Went Idle. Claude Compacted It. Copilot Stopped Waiting for Nobody.
14 Oct 2026

At a glance
- Claude Code now shrinks a big conversation you walked away from before its cache runs out.
- A custom Copilot agent can carry its own hooks, and autopilot no longer stalls on a question nobody will answer.
- Admins can make Copilot clients fetch current policy at startup and fail closed if they cannot.
- Push protection now blocks Anthropic keys by default, and a new Gemini model joined the Copilot picker.
Tuesday afternoon and evening moved three channels at once, and the common thread is what happens when nobody is watching the session. Claude Code treats an idle conversation as a cost problem it can solve before you come back, Copilot CLI's autopilot gives up on a permission question no person is there to answer, and GitHub's managed settings can stop a client whose policy might be out of date. Treat today as the day to try idle compaction on one long Claude Code session, give a custom Copilot agent its own hooks, and decide whether stale policy should block startup on your fleet.
Top Stories
1. Claude Code v2.1.299 compacts the idle conversation before the cache runs out
Anthropic published Claude Code v2.1.299 to npm at 8:23 PM ET Tuesday, and it took both the latest and next tags. At 6:05 AM ET Wednesday the stable tag had not moved and still pointed at v2.1.287.
The headline line is idle compaction in the terminal. When a conversation of 200K tokens or more sits idle for about 54 minutes, Claude Code compacts it before the one hour prompt cache expires, so your next message does not reprocess the whole history. The CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS environment variable changes the 200K minimum, and "idleCompaction": false in settings.json turns the feature off. The changelog says both controls need v2.1.290 or later, which reads oddly next to a v2.1.299 feature, so we quote it as written.
The release also adds CLAUDE_CODE_DISABLE_FILE_MENTIONS, which gives Claude the paths of @ file mentions to read with its own file tools instead of the files' contents. A sandbox fix closes a gap where a sandboxed shell command could write into a --plugin-dir folder whose code Claude Code later reloads outside the sandbox. Another fix stops a mod that submits a prompt at the end of every turn from carrying a headless run past --max-turns, --max-budget-usd or an interrupt.
Practical dev impact: If you leave a big session open over lunch or a meeting, the history should already be compacted when you return, so the first prompt back is not a full cache rebuild. The tradeoff is that compaction summarizes, and a session where you need exact earlier turns, such as a long debugging trail, may be better with the feature off or with a higher token floor. The file mentions switch is worth testing in repositories with large generated files, where pasting contents can burn context the agent never needed.
2. Copilot CLI v1.0.97 gives custom agents their own hooks and unsticks autopilot
GitHub published Copilot CLI v1.0.97 to npm at 2:37 PM ET Tuesday, and it is now latest. By 6:05 AM ET Wednesday the prerelease tag had already moved to v1.0.98-2, published at 12:26 AM ET, so keep that one off shared machines.
The first line adds agent-scoped declarative hooks for custom agents, so a hook can belong to one agent instead of every session; the notes link no docs, so we name no key. A related fix changes how autopilot treats a preToolUse hook that returns "ask": autopilot now denies the call instead of stopping on a permission dialog nobody is watching. Resumed sessions also rebuild plugin hooks from the right plugin list, so hooks from repository-local marketplace plugins keep running after --resume and hooks from plugins you disabled stay off.
Several smaller lines matter for teams. /usage now shows billed AI Credits for token-billed sessions, including after a resume. In /config, managed permission rules are read-only and hidden unless your organization set them. Streamed responses behind TLS-intercepting proxies now complete when the proxy omits close_notify.
Practical dev impact: Agent-scoped hooks let you put a strict check, such as a test gate or a path rule, on the one agent that edits production code without slowing every other agent down. The autopilot change turns a silent hang into a visible denial, which is the right failure for an unattended run, but a hook that answers "ask" will now block that work in autopilot, so review any hook that leans on a human prompt. Developers behind a corporate proxy that inspects TLS have a concrete reason to update this week.
3. GitHub lets enterprises refuse a stale Copilot policy, and adds a model
GitHub's Changelog post at 11:20 AM ET Tuesday adds forceRemoteSettingsRefresh to enterprise managed settings. Set to true through server managed, MDM managed or file based settings, it makes Copilot CLI, VS Code and the GitHub Copilot app fetch current managed settings at startup instead of trusting a cached policy, and fail closed if they cannot confirm it. The same post adds a minimum required client version to the sandbox key, so users on older clients must update before they can keep using Copilot once that key is present.
Two more posts landed the same day. At 12:12 PM ET secret scanning added detectors for Anthropic service and user API keys and turned on push protection for both by default. At 2:39 PM ET GitHub made Gemini 4 Argon generally available in Copilot for Pro+, Max, Business and Enterprise users, selectable in Copilot CLI, the major IDEs and the Copilot app, with a gradual rollout.
Practical dev impact: Failing closed is a real availability tradeoff, since a developer on a flaky network or a VPN that blocks the settings endpoint may find Copilot will not start, so pilot it with one group before applying it everywhere. The sandbox version floor means any team that pins Copilot CLI for stability should confirm the pin meets the minimum before admins add the key. If Anthropic keys sit in test fixtures or sample files in a repository, expect pushes to be blocked and move those values into a secret store.
Practical Impact Analysis
All three stories change what a tool does when you are not looking at it, which is where a quiet default can surprise a team. Claude Code now rewrites your idle history for you, Copilot autopilot now says no where it used to wait, and managed clients can now refuse to start rather than run on old rules, so in our view each default deserves a deliberate choice.
The rollout order follows from who is affected. Individual developers can take Claude Code v2.1.299 from latest and decide per project whether idle compaction helps, while teams that want slower change can stay on stable, which read v2.1.287 at 6:05 AM ET Wednesday. Copilot CLI v1.0.97 is a reasonable fleet update, but audit autopilot hooks that answer "ask" first, and administrators should hold forceRemoteSettingsRefresh until they know how reliably clients reach the settings endpoint. At the same 6:05 AM ET check, Gemini CLI latest was v0.64.0, published to npm at 4:48 PM ET Tuesday with preview at v0.65.0-preview.0, and Codex latest held at v0.162.1 with alpha at v0.163.0-alpha.11. Cursor's main changelog led with an Origin app marketplace, and the Cursor CLI changelog still ends at v2026.10.08.
Tutorial
We have not run this script end to end on every platform, so run it before you change anything today. It prints the Claude Code channel pointers and your local version, and every install or setting line is commented out and labeled, so nothing changes unless you uncomment it on purpose.
#!/usr/bin/env bash
# 1. Channel pointers. At 6:05 AM ET Wed 14 Oct npm showed Claude Code stable v2.1.287, latest and next v2.1.299.
npm view @anthropic-ai/claude-code dist-tags --json
# 2. What this machine runs
claude --version
# 3. Claude Code latest channel, npm installs only, npm latest at 6:05 AM ET Wed 14 Oct:
# npm install -g @anthropic-ai/claude-code@2.1.299
# 4. Optional, documented in the v2.1.299 changelog: raise the idle compaction floor for this shell only
# export CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS=300000
If latest no longer reads v2.1.299, a newer build shipped after 6:05 AM ET, so read its changelog before you update. Native installs follow their own pointers, so do not mix npm and native installs of the same tool on one machine. To turn it off for one project instead of raising the floor, add "idleCompaction": false to that project's settings.json.
One suggestion of ours, not something the changelog documents: if you want a rough sense of whether compaction helps, glance at /usage before you step away from a large session and again after your first message back. Treat it as an informal look, not a measurement, since the changelog never says /usage reports cache rebuild cost.
Recommended AI prompt
Copy this paragraph into ChatGPT, Claude, Gemini, Grok, or whatever AI you use.
You are my developer tooling lead for Wednesday 14 October 2026. Claude Code v2.1.299 is now the npm latest release and compacts a conversation of 200K tokens or more after about 54 minutes idle, before its one hour prompt cache expires, with CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS to change the floor and "idleCompaction": false in settings.json to turn it off. Copilot CLI v1.0.97 adds hooks scoped to a single custom agent and makes autopilot deny a preToolUse hook's ask decision instead of waiting, and GitHub added a managed forceRemoteSettingsRefresh setting that makes Copilot clients fetch current policy at startup and fail closed if they cannot. Ask me how long my sessions usually run, whether we use Copilot autopilot with hooks, and how reliable our network access to GitHub is. Then recommend which of these defaults to keep, change or pilot, and link me to the official changelog or documentation for each setting instead of inventing names.