Read the Large File Once. Compact the Subagent Early. Stable Stepped to v2.1.287.

At a glance

• Claude Code's newest build lets Claude read a large file in one call and lets a subagent compact early, and it was not on npm's stable channel at 6:04 AM ET today.
• Claude Code's stable channel stepped forward one build after Friday morning, to the release that added Claude Mods.
• Copilot CLI's new release, npm's latest at 6:04 AM ET today, brings native Entra sign-in on macOS out of pre-release.
• Codex fixed a crash and a startup failure, Gemini CLI only patched its preview, and Cursor's changelog showed nothing new.

Friday brought no new model, but it changed how a long Claude Code session spends its context. Claude Code v2.1.296 lets Claude pull a whole oversized file in one read and lets a subagent compact on its own schedule, while npm's stable channel stepped from v2.1.286 to v2.1.287 at some point after 8:44 AM ET Friday, so stable fleets updating now get the Claude Mods build we covered on 2 Oct. Treat today as the day to canary Claude Code v2.1.296 on machines that run long subagent work, reread the v2.1.287 notes before your stable machines restart, and move Copilot CLI to v1.0.95 where Mac users sign in with Microsoft Entra.

Top Stories

1. Claude Code v2.1.296 reads big files once and compacts subagents early

Anthropic published Claude Code v2.1.296 to npm at 12:58 PM ET Friday and posted the GitHub release at 3:28 PM ET. Of its 79 changelog entries, two change how a long session spends context, and a third changes which model workflow agents use. The Read tool gains an allow_large option, so Claude can read a text file past the usual size limits in one call when it needs the whole file and the context has room, instead of paging through it with offset and limit. Subagent frontmatter and --agents definitions gain autoCompactWindow, so a subagent can auto-compact earlier than the main conversation's window. A new CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL variable runs every workflow agent on one model while other subagents keep theirs. The documented CLAUDE_CODE_SUBAGENT_MODEL_FORCE is broader, since it also forces subagents and teammates.

The release also adds CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS for a longer backoff on overloaded (529) retries and doubles the default limit on MCP tool descriptions and server instructions to 4,096 characters. Claude Code's /cost, status line, --max-budget-usd and SDK cost figures now price Sonnet 5.5 cache reads at $0.10 per million tokens instead of $0.20. Anthropic's pricing page already lists $0.10, so this fixes what Claude Code displays, not what you pay. Fixes worth knowing cover managed PreToolUse and prompt hooks that refused a call without ending the turn, and Bash permission checks that auto-approved some commands using BASH_ARGV0.

The subagents, tools and environment variable docs did not cover any of the four additions at 6:05 AM ET today, so the changelog is the only official source, and it gives no unit or range for the subagent field. The docs do describe a top-level autoCompactWindow setting in tokens, from 100,000 to 1,000,000; our guess that the subagent field works the same way is an inference, not Anthropic's statement.

Practical dev impact: If claude --version prints v2.1.287 or older, none of this is on your machine. Canary v2.1.296 where agents chew through large logs or generated reports, give the subagent that grows fastest a smaller window than its parent once you have confirmed the value format, and use the workflow variable to put workflow agents on one model without editing each definition. Recheck any spend alert tuned to Claude Code's old Sonnet 5.5 cache figure.

2. Claude Code stable stepped to v2.1.287

At 6:04 AM ET today npm stable pointed to v2.1.287, latest and next both pointed to v2.1.296, and Anthropic's native installer pointers matched. Our checks last saw stable on v2.1.286 at 8:44 AM ET Friday, and npm keeps no history of tag moves, so the exact time is unknown. The v2.1.287 build was published on 1 Oct and added Claude Mods and the built-in You should know watcher.

Practical dev impact: A machine that was on v2.1.286 and updates on stable this weekend picks up mods for the first time, so revisit the plugin decision you made after our 2 Oct brief before those machines restart. They still lack everything from v2.1.288 through v2.1.296, including the onFailure hook setting we covered Friday.

3. Copilot CLI v1.0.95 takes Entra broker sign-in to latest

GitHub released Copilot CLI v1.0.95 at 12:04 PM ET Friday, npm had it the same minute, and it was npm latest at 6:04 AM ET today. On macOS it uses the native Microsoft Entra broker for sign-in when available, with a browser fallback, and copilot config supports sandbox credential injectHosts keys, with key completion in Bash, Zsh and Fish. Both first appeared in Thursday evening's pre-releases. The final release adds two more lines: --context now applies to new and resumed ACP sessions instead of silently using the default or a previously saved context tier, and managed plugin setup retries hourly or after policy changes instead of on every message failure.

At 6:04 AM ET npm prerelease pointed to v1.0.96-2, posted at 10:13 PM ET Friday, which makes model IDs in /model and /config case-insensitive and saves the canonical ID.

Practical dev impact: Move Mac users who sign in with Entra to v1.0.95, and test ACP integrations that pass --context, because our reading is that a resumed session may now run at a different tier than before. The notes do not give the injectHosts key path, so use shell completion on copilot config to find it. Keep the v1.0.96 pre-releases off fleet machines.

4. Codex v0.162.1 fixes a crash and a startup mismatch

OpenAI released Codex v0.162.1 on GitHub at 3:44 PM ET Friday and on npm at 3:50 PM ET, and it was npm latest at 6:04 AM ET today. It fixes a TUI crash when asynchronous questions span several lines, and startup failures when a running background server's feature settings differed from the CLI defaults. The v0.163.0-alpha.5 build, npm alpha at 6:04 AM ET, has no release notes.

Gemini CLI npm latest was still v0.63.0 at 6:04 AM ET, and only preview moved, to v0.64.0-preview.1, a single cherry-picked patch. Cursor's changelog still led with Remote Control for local agents at 6:05 AM ET, which we covered on Wednesday.

Practical dev impact: Move to Codex v0.162.1 if you run a persistent background server next to the CLI, and keep the alpha builds off fleet machines. Gemini CLI and Cursor give you no reason to touch a pin today.

Practical Impact Analysis

The thread through today is context budget. Claude Code v2.1.296 gives you two levers that work together: a whole log can arrive in one read, which fills a subagent faster, and that subagent can compact before its parent has to. Set the windows on purpose, so the specialist that grows gets the smaller window while the parent that holds your plan keeps the larger one.

The stable step matters for a different reason. Teams that stayed on stable to avoid new surfaces now receive mods without the fixes that followed, which is a reason to reread the v2.1.287 notes, not to jump everyone to latest. Copilot CLI and Codex moved toward fewer surprises at sign-in and startup, so those updates are low risk, and the one behavior change worth testing is --context on resumed ACP sessions. Check versions rather than trusting the channel you think a machine follows.

Tutorial

Check the channels first, because these tags move without notice. The install lines are commented out so you can run the checks safely and uncomment one on purpose, on the channel it names.

#!/usr/bin/env bash
# 1. Read the channel pointers and local versions.
npm view @anthropic-ai/claude-code dist-tags --json
npm view @github/copilot dist-tags --json
npm view @openai/codex dist-tags --json
claude --version   # allow_large and autoCompactWindow need v2.1.296 or later
copilot --version
codex --version

# 2. List the subagent definitions you might give their own compact window.
for d in "$HOME/.claude/agents" ".claude/agents"; do
  [ -d "$d" ] && ls "$d"
done

# 3. Optional canaries, npm installs only. Leave stable fleets (npm stable was v2.1.287 at 6:04 AM ET) alone.
# Claude Code, npm latest and next at 6:04 AM ET:
# npm install -g @anthropic-ai/claude-code@2.1.296
# Copilot CLI, npm latest at 6:04 AM ET:
# npm install -g @github/copilot@1.0.95
# Codex, npm latest at 6:04 AM ET:
# npm install -g @openai/codex@0.162.1

Restart each tool after a change and trust its --version output over the tag; native installs follow their own pointers. Do not mix npm and native Claude Code installs on one machine.

The changelog names autoCompactWindow and where it goes but not its value format, so we are not printing a sample definition. On a canary, add it to the frontmatter of the subagent that grows fastest, then start a session with --debug, which in v2.1.296 names unrecognized frontmatter fields in custom agent files. Then run that subagent on a large log and note when it compacts. This check is our suggestion, not Anthropic's, and we have not run it.

The changelog lists no accepted values for CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL; the documented CLAUDE_CODE_SUBAGENT_MODEL takes an alias such as sonnet or a full model name, and we assume the new one does too. By our reading, Claude sets allow_large, not you, so say in your prompt when Claude needs a whole file.

Recommended AI prompt

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