Claude Code Lets Subagents Think Less, Copilot Now Remembers Your Repo, and OpenCode Sizes Compaction to Every Model

At a glance

• Claude Code subagents can now run at a lower effort level than your main session.
• Opus 4.8 fast mode is on a 30 day clock, and requests will fail rather than fall back.
• Copilot Memory is generally available, and an opt-in preview proposes instruction changes as pull requests.
• OpenCode can size its compaction buffer as a share of each model's context window.

This is a thin Saturday with one real overnight release. Claude Code v2.1.316 reached npm at 8:42 PM ET Friday, a few hours after v2.1.315, which has no changelog section of its own. OpenCode shipped v2.2.4 and v2.2.5 on GitHub during Friday. Codex, Copilot CLI and Gemini CLI kept the same latest tags, and neither Cursor changelog moved.

So the second story is not overnight news. It is a GitHub Universe announcement from Wednesday that was easy to miss among the bigger Copilot launches this week. Our read: it is the item most likely to change how your team writes and reviews instruction files.

Treat today as a context and cost tune-up: try a lower subagent effort level in one repo, find anything pinned to Opus 4.8 fast mode before 30 November, and decide who reviews what Copilot remembers and proposes.

Top Stories

1. Claude Code v2.1.316 gives subagents their own effort level

The most useful addition is CLAUDE_CODE_SUBAGENT_EFFORT. The changelog says it sets the effort level subagents run at when nothing else names one, instead of inheriting the session's. A high effort main session no longer forces the same level on subagents that only search and read files. Our read: a subagent whose definition already sets its own effort: keeps that value, since the variable only fills the gap when nothing else names a level. The subagents docs say an exported CLAUDE_CODE_EFFORT_LEVEL beats a subagent's effort:, so it likely beats the new variable too.

A second switch, CLAUDE_CODE_DISABLE_ULTRATHINK, turns the ultrathink keyword off, for teams that want effort set in one place. In VS Code, ultrathink now runs its turn at xhigh effort, as in the terminal. Neither new variable appears on Anthropic's environment variables page yet, so treat the changelog as the only reference for now.

The release deprecates Opus 4.8 fast mode as of 30 October and removes it on 30 November. After removal, fast mode requests on Opus 4.8 return an error and do not fall back to standard Opus 4.8. The changelog says to migrate to Opus 5.5 to keep the speedup.

Among the fixes, --restricted sessions no longer upload files from outside the working directories when a Read allow rule covered them, PowerShell hooks now block on exit code 2 when the command begins with a type cast or param, Grep and Glob no longer hang on Windows, and SSL certificate errors now name likely causes.

Practical dev impact: If any script, alias or saved default runs Opus 4.8 in fast mode, change it this month, because after 30 November those requests stop with an error instead of quietly running at standard speed.

2. Copilot Memory is GA, and dreaming can propose instruction changes

GitHub announced this at Universe on Wednesday, so it is this week's news rather than today's. Copilot Memory is now generally available on paid Copilot plans. Copilot can remember personal preferences and repository conventions across sessions and clients, including the Copilot app, Copilot CLI, VS Code, cloud agent and code review. A repository memory can hold the right test command, so you stop repeating it. You can view or delete personal memories in your Copilot settings.

Copilot dreaming, in public preview, reflects on past sessions to improve future ones. It offers three dream types: personal memory curation, repository customization, which manages repository memories and proposes improvements to existing skills and custom instruction files, and code review customization, which turns recurring accepted review feedback into guidance for Copilot code review. Changes to checked-in skills and instructions arrive as pull requests for your team to review, while memories and code review guidance update without one. Dreaming is opt-in and not available on Copilot Pro. Administrators enable it for organizations through Copilot policies, and dream runs consume AI credits.

Practical dev impact: Before anyone turns on repository dreams, put your instruction and skill files under code owners, so a pull request that Copilot proposes gets the same review as any other config change.

3. OpenCode v2.2.5 sizes compaction to every model

OpenCode, the open source terminal and desktop coding agent, shipped two releases on Friday. In v2.2.5, compaction.buffer in opencode.json accepts a percentage of the model's context window, such as "20%", in addition to a token count. One setting can now keep the same share of the window free whether you switch to a small or a large model.

Errors got more useful too. When a read misses because a folder is missing, the error names that folder and the nearest one that exists. When an edit cannot find the text it was told to replace, the error shows the file's closest lines or says the change is already there. LiteLLM context window errors now trigger compaction instead of stopping the session, and the earlier v2.2.4 made desktop terminals survive server restarts.

Practical dev impact: If you move OpenCode between models with very different context sizes, replace a fixed token buffer with a percentage so compaction stops starting too early on small models or too late on large ones.

Practical Impact Analysis

The through-line today is that effort and context are becoming settings you own rather than defaults you inherit. The Claude Code subagent variable is the quickest win, because one line in a test shell can lower what every subagent spends without touching your main session. Try it in one repo for a day. The Opus 4.8 fast mode deadline is the only item with a hard date, so it belongs on this week's list even if you skip everything else. Our read: the failure after 30 November is loud, but a nightly job nobody watches can still fail for days.

If you are on a paid Copilot plan, take five minutes to see what Memory has stored for you. Dreaming is a team decision, because it writes proposals against shared files and spends credits, and that choice is easier before it is switched on. OpenCode's percentage buffer is small, but it removes a real annoyance for anyone who switches models, and the clearer errors should cut down on repeated failed edits. Elsewhere this week, Goose v1.57.0 began requiring HTTPS for its allowlist and validating redirects.

Tutorial

1. Open a new terminal in the root of one test repo. Nothing below changes files.
2. Step 1 prints versions, and step 2 lists any settings or instruction file that still names Opus 4.8.
3. Step 3 sets the subagent effort level for this shell only. Our read: medium is a valid value, because the documented CLAUDE_CODE_EFFORT_LEVEL accepts low, medium, high, xhigh and max, but the new variable is not documented yet.
4. Start claude from that shell, run one task that spawns subagents, and check their level in /tasks. Close the shell to undo it.

# We have not run this script end to end on every platform. It prints and checks, and changes no files.
# Run it from the root of one test repo, in a new terminal you can close afterwards.

# 1. Which Claude Code version is out, and which do you have?
#    At 5:57 AM ET Sat 31 Oct npm showed @anthropic-ai/claude-code latest v2.1.316.
npm view @anthropic-ai/claude-code dist-tags.latest
if command -v claude >/dev/null; then claude --version; else echo "no claude CLI on PATH"; fi
#    Optional install only. Pinned at 5:57 AM ET Sat 31 Oct, npm registry channel:
# npm install -g @anthropic-ai/claude-code@2.1.316

# 2. Does anything still name Opus 4.8? Lists matching files only.
grep -l "opus-4-8" .claude/settings.json .claude/settings.local.json CLAUDE.md 2>/dev/null
grep -l "opus-4-8" "$HOME/.claude/settings.json" 2>/dev/null
echo "If a file was listed, check whether it also turns on fast mode, and plan a move before 30 Nov."

# 3. Subagent effort for this shell only. Nothing is written to your profile.
#    Our read: "medium" matches the documented CLAUDE_CODE_EFFORT_LEVEL values.
#    Changelog only so far. An exported CLAUDE_CODE_EFFORT_LEVEL may win; check /tasks.
export CLAUDE_CODE_SUBAGENT_EFFORT=medium
echo "Subagent effort for this shell: $CLAUDE_CODE_SUBAGENT_EFFORT"
#    Now start claude from this shell, run one task with subagents, then open /tasks.
#    To undo, close this terminal or run: unset CLAUDE_CODE_SUBAGENT_EFFORT

# 4. OpenCode only. Merge into the project opencode.json, then switch models to compare:
#    "compaction": { "buffer": "20%" }
echo "Done. Nothing was installed or changed on disk."

Sources:

• Claude Code changelog (v2.1.316)
• Claude Code environment variables
• GitHub intelligence puts your code, context, and history to work
• About Copilot Memory
• About Copilot dreaming
• OpenCode v2.2.5 release
• OpenCode v2.2.4 release
• Goose v1.57.0 release

Recommended AI prompt

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