Friday, November 13: two releases in the last 24 hours tighten the application-developer stack from different sides. Claude Code v2.1.323 is a dense security hardening release with one new experimental feature (Code Mode) โ but the real story is a roster of Bash and sandbox permission bypass fixes that read like a production incident report. E2B v2.64.0 lands on the same day with OAuth authentication and per-sandbox CPU controls, making the agent sandbox platform a more serious self-hosted story. The rising signal is GoogleChrome/modern-web-guidance-src, which exploded from 2.7k to 4.6k stars in 24 hours; its own benchmarks show coding-agent task pass rates climbing from 52โ60% unguided to 82โ96% with the guides โ a rare "this is a measurable improvement" story in a category full of vibes.
Must-reads today
1
Claude Code v2.1.323 [HIGH]: Code Mode (experimental) + security hardening across Bash, Write/Edit path deny rules, and sandbox โ plus a 1M context window default change and MCP SDK v1 removal that need attention before upgrading
2
GoogleChrome/modern-web-guidance-src [HIGH]: 1,875 stars today, benchmarks showing 52โ82โ96% pass rate lift for coding agents on modern web API tasks โ today's Repo of the Day
๐ Repo of the Day
1
GoogleChrome/modern-web-guidance-src โ keeps coding agents current on modern web platform APIs
install once (or via your agent's plugin marketplace)
npx modern-web-guidance@latest install
in a Claude Code / Codex / Copilot session, the agent calls:
โ loads current patterns + browser-compat data into context
Breaking Changes
2
โBreaking
Claude Code v2.1.323 โ MCP SDK v1 client removed; 1M context default for unrecognized models
What changed
v2.1.322 โ v2.1.323: MCP_SDK_GENERATION=v1 no longer loads the v1 client โ it now runs the v2 client in compatibility mode (v1 is removed). Separately, the default context window for model IDs Claude Code does not recognize changed from 200K to 1M tokens; subagentStatusLine setting is removed.
TL;DR
Claude Code v2.1.323 removes the MCP SDK v1 client and changes the default context window from 200K to 1M for unrecognized model IDs โ two behavior changes requiring action before upgrading.
Developer signal
Three things to do before upgrading: (1) If you use MCP_SDK_GENERATION=v1, you are now on the v2 client in compatibility mode โ test your MCP server integrations against the v2 protocol. (2) If you use third-party or custom model IDs that Claude Code doesn't recognize, they now assume 1M context; if the actual model has a smaller window, every turn could fail once context fills. Set CLAUDE_CODE_DISABLE_1M_CONTEXT=1 or run /clear proactively. (3) If your config sets subagentStatusLine, remove it โ the setting is ignored and files still load, but it's dead config. Additionally: --permission-mode auto now starts in Manual mode when auto mode is unavailable, rather than a potentially more permissive mode โ verify your CI and unattended workflows start in the expected mode.
Affects you ifYou set MCP_SDK_GENERATION=v1; you run Claude Code against custom or third-party model IDs with context windows smaller than 1M; you rely on subagentStatusLine in your config; you run --permission-mode auto in CI.EffortModerate (audit MCP SDK generation setting, test custom model IDs, remove dead config)
E2B CLI v2.24.0 โ `template list` JSON output shape changed; OAuth auth replaces API key in CLI
What changed
Previously, e2b template list returned a JSON structure keyed on the old shape; v2.24.0 of the CLI changes that JSON output shape (breaking for scripts that parse it) and removes the "Created by" column. Separately, the CLI now authenticates using your login's access token instead of a stored API key, with automatic refresh; E2B_API_KEY still takes precedence.
TL;DR
E2B v2.64.0 breaks the JSON output shape of e2b template list and shifts CLI auth to OAuth access tokens; also adds OAuth to the SDK, resizable templates with configurable CPU counts, and commands.run with a custom shell option.
Developer signal
If you have scripts that parse e2b template list --json, update them for the new shape before upgrading the CLI. For CI: E2B_API_KEY still works and takes precedence, but the new recommended approach is E2B_ACCESS_TOKEN + E2B_PROJECT_ID for token-based auth. For resizable templates: build templates with --max-cpu and pass --cpu on sandbox create to pick the CPU count at runtime โ useful for agent workloads with variable compute needs. The commands.runshell option lets you specify a custom shell (e.g., zsh, fish) for command execution instead of the default sh.
Affects you ifYou parse e2b template list --json output in scripts; you manage E2B auth via CLI and need to migrate to access tokens for CI; you run agent sandboxes and want to vary compute allocation without rebuilding templates.EffortQuick for most users (auth migration is automatic); Moderate for scripts that parse template list JSON (output shape changed)
No model releases in the last 24 hours. (Claude Fable 5.2 / Mythos 5.2 launched November 9 โ check the November 9 digest if not already reviewed.)
API & SDK Changes
No new API/SDK releases from Anthropic SDKs (last release November 10, outside 24h window). See Breaking Changes section for E2B SDK changes.
Research
arxiv.org DNS-blocked (fourteenth consecutive day). papers.cool DNS-blocked. Two arXiv November 2026 papers surfaced in search results (2611.03551: forward-deployed engineer agent evals; 2611.03575: VeriHarm-Bench for coding agents causing unintended harm) but could not be fetched โ both deferred to near-misses.
Tooling
1
High
Claude Code v2.1.323 โ Code Mode (experimental) + 15 security hardening fixes across Bash, Write/Edit, and sandbox
What changed
v2.1.322 โ v2.1.323 (released Nov 12, 22:11 UTC): ships Code Mode as an experimental opt-in feature, fixes 15 permission and sandbox boundary issues, removes MCP SDK v1, changes unrecognized model default context from 200K to 1M, extends Remote Control session identity to 30 days (from 4 hours), and adds plugin version range enforcement via engines["claude-code"].
TL;DR
Claude Code v2.1.323 ships experimental Code Mode (Claude writes JS to call its tools), 15 security fixes โ including Bash wildcard bypass, Write/Edit path deny bypass, and sandbox escape โ and extends Remote Control sessions to 30-day identity retention.
Developer signal
Code Mode (experimental): opt in via settings; Claude writes JavaScript to orchestrate its tool calls rather than making discrete tool calls. The docs at https://code.claude.com/docs/en/code-mode cover enabling and org-level disable. Worth testing for long agentic tasks where tool-call granularity creates overhead. Security hardening: upgrade promptly if your team uses Bash allow rules with wildcards (e.g., Bash(ls*)), sudo wrappers, or runs Write/Edit on paths under Read-deny rules โ all of these had real bypasses now fixed. Remote Control: sessions retain identity for 30 days (up from 4 hours), so long-running background sessions and remote orchestration pipelines will no longer drop connection state. Plugin enforcement:engines["claude-code"] in plugin.json now gates install/update/load โ if you maintain a plugin, add the version range to avoid compatibility surprises for users on older versions.
Affects you ifYou use Bash allow/deny rules with wildcards or sudo; you Write/Edit to paths that have Read-deny rules; you run long Remote Control sessions; you maintain a Claude Code plugin; you use Code Mode (opt-in, experimental).EffortQuick for most (auto-upgrade); Moderate for plugin maintainers adding version ranges and teams auditing Bash rule changes
GoogleChrome/modern-web-guidance-src โ coding-agent guidance that benchmarks a 30โ36 point pass rate lift on modern web tasks
Developer signal
Install with npx modern-web-guidance@latest install or through your agent's plugin marketplace. Run npx modern-web-guidance@latest legacy-scan . on an existing codebase to surface outdated patterns before you start applying guides. If you use Claude Code: install from the GoogleChrome marketplace (the Claude Code built-in marketplace stopped updating this plugin). The agent calls search to find a relevant guide and retrieve to load it โ this is a tool-call pattern, not a prompt injection.
Affects you ifYou use Claude Code, Copilot, Codex, or any coding agent for frontend/fullstack web work; you review agent-generated code for outdated patterns; you build web components with AI assistance.
GitHub Trending (all languages, daily, Nov 13) | Stars: ~4.6k (โ 1,875 today, from 2.7k Nov 12) | Link: github.com/GoogleChrome/modern-web-guidance-src
Benchmarks & Leaderboards
lmarena.ai DNS-blocked (twelfth consecutive day). swebench.com DNS-blocked. No benchmark data confirmed from primary sources this period.
Trends & Emerging Tech
2
Claude Code's security hardening in v2.1.323 reads like a production incident report โ permission-bypass edge cases at scale
What's happening
v2.1.323 fixes 15 distinct permission and sandbox boundary issues in a single release. The scope is unusually broad: Bash allow rules with wildcards skipped the prompt; combined sudo flags skipped the prompt; Write/Edit could write to paths covered by a Read deny rule; plan files with plansDirectory set could reference paths outside the project; the sandbox could switch off mid-session; sandbox.excludedCommands could let unnamed commands run outside the sandbox; project settings were readable through symlinks; .mcp.json approvals weren't enforcing exact name matches. The pattern is: edge cases discovered at scale in production deployments, not theoretical attack paths.
Why watch this
As Claude Code usage grows โ particularly in team and CI environments โ the gap between the permission model's intended behavior and its actual boundaries is narrowing. This release suggests Anthropic is now getting real-world feedback on where those boundaries break down. For teams running Claude Code in production: treat each of these fixes as a signal about where the previous boundary was ambiguous, and audit your allow/deny rules against the new behavior. The trend to track is whether future releases continue this pattern of reactive hardening, or whether the architecture shifts toward a more formally verified permission model.
GoogleChrome/modern-web-guidance-src's growth signals a "pre-training gap" awareness shift
What's happening
The 1,875-star-day for modern-web-guidance-src is notable not just for its size but for what it represents: teams are actively acknowledging that their coding agents don't know what the web platform knows now, and reaching for a tool to bridge that gap. This is a practitioner-level acknowledgment of training data staleness โ something that has been theoretically understood but rarely acted on with concrete tooling. The eval numbers (52โ60% โ 82โ96% pass rate on modern web tasks) are concrete enough to get buy-in from engineering orgs that require numbers, not vibes.
Why watch this
If the "agent knowledge gap" pattern generalizes beyond web APIs โ say, to framework best practices, security patterns, or cloud service APIs that change fast โ it creates a category of tools: "living guidance" that keeps coding agents current in specialized domains. Modern-web-guidance-src is the clearest example of this pattern so far, and its growth today suggests the market is ready for more of them.
hnrss.org inaccessible (high-score Show HN items not confirmable). simonwillison.net not checked. Nothing cleared the quality bar this period from confirmed primary sources.
Quick Hits
tailscale/tailcat (open sourced Aug 2026): Netcat-style tool that runs over Tailscale's data plane without Tailscale's control plane โ WireGuard-encrypted, works through NAT, no account needed on the other side. 9.4k stars, 324 today. Source: github.com/tailscale/tailcat
vercel-labs/agent-desktop (weekly trending): CLI + C library for giving AI agents desktop computer use via the accessibility tree (not screenshots), with an MCP server (agent-desktop mcp). 23.2k stars, +6.4k this week. Source: github.com/vercel-labs/agent-desktop
Worth Watching (Announced, Not Yet Shipped)
Gemini API preview model deprecations โ November 17 (4 days): Gemini 3.1 Flash TTS preview, Gemini Live API preview models, Gemini 2.5 TTS previews retire. Replacement: gemini-3.8-live and gemini-3.8-flash-tts. Source: ai.google.dev/gemini-api/docs/deprecations (DNS-blocked โ verify at official URL).