DJ whispher

ハードワークがたたり、メンタル不調になったサラリーマン。
毎日のめまい、頭痛等の症状を抱えながらも、いい仲間に支えられ、周りの方に感謝し、行動することで、新たな環境を作り続けます。
最近はClaudeCodeにハマっており、これで癒しAPLを開発中

[DJ whispher @masayaneg]

Claude Code v2.1.307 added a claude_code.safety_event OpenTelemetry event. It fires for each tool call that a permission rule, auto mode, a safety check, a hook or the sandbox stopped, prompted or sent back. Handy if you want to track how often your guardrails actually kick in.

[DJ whispher @masayaneg]

Claude Code v2.1.305 added sandbox.allowNestedUserNamespaces. It lets sandboxed commands on Linux create user namespaces, which Chromium and rootless podman need. The catch: it also lets them get around the sandbox's file protections, so only enable it when you need those tools.

[DJ whispher @masayaneg]

Claude Code v2.1.304 now warns you on Pro and Max plans when resuming an inactive conversation would eat a noticeable share of your 5-hour usage limit. You can resume anyway, or start fresh and just refer to the old one so Claude can look it up.

[DJ whispher @masayaneg]

Claude Code v2.1.304 added /auto-mode-setup: a guided setup that reads your project and your recent sessions to teach auto mode about your environment. It's also available now on Bedrock, Vertex, Foundry, Claude apps gateway and telemetry-disabled sessions.

[DJ whispher @masayaneg]

Claude Code v2.1.302 turned fork subagents (subagent_type: "fork") on by default in claude -p, the Agent SDK, the desktop app and VS Code, as they already were in the terminal. Don't want them? Set CLAUDE_CODE_FORK_SUBAGENT=0. claude mcp serve never offers them.

[DJ whispher @masayaneg]

Claude Code v2.1.302 added CLAUDE_CODE_AUTO_MODE_MODEL. It picks the model the client-side auto mode classifier uses, the one that judges whether each tool call is safe to run. It does not change the model for server-side auto mode, so check which kind of classifier you're on.

[DJ whispher @masayaneg]

Claude Code v2.1.299 tightened MCP allow rules: mcp__a now matches only the server named a, no longer a server named a__b or a_. To allow every tool on a server, write mcp__<server>__* with the server's whole normalized name. Worth rechecking your rules.

[DJ whispher @masayaneg]

Claude Code v2.1.299 added CLAUDE_CODE_DISABLE_FILE_MENTIONS. With it set, an @ file mention hands Claude only the file's path, not its contents, so Claude reads the file itself with its file tools. Handy if you want @ as a pointer, not a context dump.

[DJ whispher @masayaneg]

Claude Code v2.1.299 added idle compaction in the terminal: a 200K+ token conversation left idle for about 54 minutes is compacted before the one-hour prompt cache expires, so your next message doesn't reprocess it all. Opt out with "idleCompaction": false.

[DJ whispher @masayaneg]

Claude Code v2.1.296 doubled the default limit on MCP tool descriptions and MCP server instructions sent up front, from 2,048 to 4,096 characters. If a server's instructions were getting cut off, they now get twice the room before truncation.

[DJ whispher @masayaneg]

Sonnet 5.5 got cheaper in Claude Code's cost tracking. As of v2.1.296, /cost, the status line, --max-budget-usd and the SDK's cost figures price Sonnet 5.5 cache reads at $0.10 per million tokens, down from $0.20. Long cached sessions show a lower bill.

[DJ whispher @masayaneg]

Claude Code v2.1.296 added an allow_large option to the Read tool. When Claude genuinely needs a whole text file that's past the usual size limits, it can now pull it in with one call, as long as the context has room. No more reading big files in chunks.

[DJ whispher @masayaneg]

Claude Code v2.1.296 added CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL. Set it and every workflow agent runs on that one model, while your other subagents keep their own model settings. Useful for pinning a cheaper model on big fan-out workflows.

[DJ whispher @masayaneg]

Claude Code v2.1.295 added onFailure: "block" for command and HTTP hooks. If a hook can't start, times out, or exits with an unexpected code, the action is now blocked instead of slipping through. Handy for security hooks that must fail closed.

[DJ whispher @masayaneg]

Claude Code v2.1.296 added autoCompactWindow to subagent frontmatter and --agents definitions. A subagent can now auto-compact earlier than the main conversation's window, so a chatty subagent can trim its own context sooner instead of riding the full window.

[DJ whispher @masayaneg]

Claude Code v2.1.293 fixed a subtle post-compaction bug: Claude would sometimes treat its own last actions before a context compaction as still pending, then retract or redo work it had already finished. If long sessions felt like they had do-overs, update.

[DJ whispher @masayaneg]

Claude Code v2.1.292 added an effort parameter to the Agent tool. So when you tell Claude to spin up a sub-agent, you can now also ask for the effort level it should run at, instead of just taking whatever the default is. Handy for cheap lookups vs. hard problems.

[DJ whispher @masayaneg]

Claude Haiku 5.5 (claude-haiku-5-5) just shipped in Claude Code v2.1.293 and is now the default Haiku model on the Anthropic API. It comes with a 1M context window and costs $0.10/$0.50 per Mtok in/out, or $0.50/$2.50 for prompts over 100K tokens.

[DJ whispher @masayaneg]

Claude Code v2.1.288 added --max-findings <n>|all to /code-review. By default it reports a limited number of findings; pass a number for fewer or more, or "all" for everything. The choice sticks until you run --max-findings default. Handy when a big diff gets truncated reviews.

[DJ whispher @masayaneg]

Claude Code v2.1.289 closed two gaps in Bash deny/ask rules under sandbox auto-allow: a rule could be missed when the command sat behind an env var prefix with an expanded value (TZ="$HOME" rm -rf build), or behind a bare variable assignment. If you rely on deny rules, update.