← All digests
AI Developer Digest

Sat, Nov 14, 2026

6 signals that cleared the gate17 min read
The Signal — start here
Saturday, November 14: a focused day dominated by one release that will catch automated agent pipelines by surprise. Claude Code v2.1.325 — which shipped last night at 20:40 UTC, after the Nov 13 digest was written — changes two defaults that matter for anyone running Claude Code unattended: sandboxed Bash now requires explicit opt-in to skip permission prompts (autoAllowBashIfSandboxed: true), and scheduled tasks / /loop now pause before running skills that declare allowed-tools or hooks unless you've pre-approved them. Combined with five more security fixes for Bash permission edge cases, this is the third Claude Code release in a week to tighten the permission boundary. The other story worth flagging before the weekend: Gemini 4 Argon introductory pricing ends November 30 — $2/$10 per million tokens doubles to $4/$20 on December 1. If you're using Argon in production, that's a 2× cost event in 17 days.
Must-reads today
1
Claude Code v2.1.325 [BREAKING]: sandboxed Bash and scheduled-task skill permissions changed by default — check your CI and unattended workflows before the next auto-update
2
Gemini 4 Argon pricing — Dec 1: introductory rates end Nov 30; audit your Gemini 4 Argon usage now

🏆 Repo of the Day

1

backnotprop/plannotator — a persistent review workspace that keeps humans in the loop on AI agent plans and diffs

install (npm or npx)

npx plannotator@latest init

open the workspace for the current repo

npx plannotator@latest open

or install once globally

npm install -g plannotator

plannotator open

Claude Code users: add it as a plugin from the Claude Code marketplace or run plannotator mcp to expose its review surface as an MCP server.


Breaking Changes

1
●Breaking

Claude Code v2.1.325 — sandboxed Bash and skill permissions changed by default

What changed
v2.1.323 → v2.1.325: sandboxed Bash no longer runs without prompts by default — it now follows the normal permission flow. Scheduled tasks and /loop now ask before running a skill that declares allowed-tools or hooks.
TL;DR
Claude Code v2.1.325 changes two automation defaults: sandboxed Bash requires a prompt unless autoAllowBashIfSandboxed: true is set, and scheduled / looped runs pause before skills with allowed-tools or hooks unless pre-approved.
Developer signal
Two config changes to audit before the next auto-update:

(1) Sandboxed Bash: If you run Claude Code with sandbox enabled and expected Bash to execute without prompts, add { "sandbox": { "autoAllowBashIfSandboxed": true } } to your highest-priority managed settings source (or project/user settings). Without it, any CI or unattended run that uses Bash will now pause and wait.

(2) Scheduled tasks / /loop skills: If a skill in your loop declares allowed-tools or hooks, add it to your allow list: { "permissions": { "allow": ["Skill(name)"] } }. Without this, recurring agent runs that call those skills will stall. The intent is that skills that can run code or call external tools require explicit authorization before running unattended.

Additional behavior changes to review: --add-dir (not additionalDirectories) is now the correct flag to allow git reads in added directories; claude remote-control now exits with status 4 on self-stop and 69 on registration failure (useful for service managers); CLAUDE_CODE_MAX_CONTEXT_TOKENS now also sets the context window for unrecognized claude-* model IDs (use alongside CLAUDE_CODE_DISABLE_1M_CONTEXT); marketplace ref now accepts any full commit ID on a branch or tag, but clients before v2.1.325 can't load such pins.


Affects you ifYou run Claude Code with sandbox enabled in CI or automated workflows; you use scheduled tasks or /loop with skills that declare allowed-tools or hooks; you use --add-dir / additionalDirectories in git-related configs; you run claude remote-control under a service manager and rely on exit codes.EffortQuick (add one config key per change) / Moderate (audit all skills used in automated loops)
GitHub (anthropics/claude-code v2.1.325) | Date: 2026-11-13 20:40 UTC | Link: https://github.com/anthropics/claude-code/releases/tag/v2.1.325https://github.com/anthropics/claude-code/releases/tag/v2.1.325

Model Releases

No model releases in the last 24 hours.


API & SDK Changes

No new API or SDK releases in the last 24 hours. (Anthropic SDK Python v2.8.0 and TS sdk-v1.8.0 shipped Nov 10 — check the Nov 10 digest.)


Research

arxiv.org DNS-blocked (fifteenth consecutive day). huggingface.co/papers DNS-blocked. No research items confirmed from primary sources this period.


Tooling

1
High

Claude Code v2.1.325 — permission hardening + VS Code prompt suggestions + MCP tool description length increase

What changed
v2.1.322 → v2.1.325 (this release covers the gap from v2.1.323, the last release covered in the Nov 13 digest): VS Code gains tab-to-accept prompt suggestions; MCP tool descriptions on the web expand to 16,384 characters (from 4,096); five additional Bash permission edge cases patched; session resume now retains recent turns after unclean exit; Remote Control sessions exit with predictable status codes.
TL;DR
Claude Code v2.1.325 ships VS Code prompt suggestions (Tab to accept), raises MCP tool description limits to 16,384 chars for web, fixes five Bash permission edge cases, and addresses session resume gaps — in addition to the two behavior changes in Breaking Changes.
Developer signal
VS Code prompt suggestions: Tab accepts the suggested prompt in the VS Code terminal panel. If you find this intrusive, disable it with { "promptSuggestionEnabled": false } in your user settings. MCP tool descriptions: If you maintain an MCP server used via Claude Code on the web, you can now write up to 16,384-character tool descriptions — a 4× increase that enables richer tool documentation without truncation. Session resume: Sessions that ended uncleanly (crash, force-quit) now restore correctly, including recent turns. This fixes a persistent complaint about lost context after a crash. Bash security: Five edge cases patched — compound commands, deny/ask rule misses, allow rules running different programs than named, relative rules not applying after cd, and sandbox writes to the native updater folder. If you rely on the previous behavior of any of these edge cases (unlikely, but auditable by reviewing your .claude/settings.json allow rules), verify the new behavior.


Affects you ifYou use Claude Code in VS Code and want prompt suggestions; you maintain or use MCP servers with long tool descriptions on the web; you rely on session resume in long agentic workflows; you have custom Bash allow/deny rules.EffortQuick (auto-update; add config key if disabling prompt suggestions or restoring sandboxed Bash behavior)
GitHub (anthropics/claude-code v2.1.325) | Date: 2026-11-13 20:40 UTC | Link: https://github.com/anthropics/claude-code/releases/tag/v2.1.325https://github.com/anthropics/claude-code/releases/tag/v2.1.325

Rising Dev Tools

1
Notable

backnotprop/plannotator — annotation and review workspace for AI agent plans and diffs

Developer signal
npx plannotator@latest open starts a local browser workspace linked to the current repo. Annotate plans and diffs inline, share via URL with teammates, and click to send structured feedback to the connected agent. For Claude Code: install from the marketplace or run plannotator mcp to expose it as an MCP server. Run plannotator init to configure which agent session to connect.


Affects you ifYou review AI agent plans or diffs and want an annotation + sharing surface; you work in a team where multiple people review agent-generated plans; you run structured human-approval workflows with coding agents.
GitHub Trending (all languages, daily, Nov 14) | Stars: ~11k (↑ 412 today; ~3.7k March 2026 → ~11k now) | Link: github.com/backnotprop/plannotator

Benchmarks & Leaderboards

lmarena.ai DNS-blocked (thirteenth consecutive day). swebench.com DNS-blocked. No benchmark data confirmed from primary sources this period.


Technical Discussions

hnrss.org high-score Show HN items not confirmable. simonwillison.net most recent post from Nov 11 (outside 24h window). Nothing cleared the quality bar this period from confirmed primary sources.


Quick Hits

  • Anthropic files S-1 with the SEC for a proposed IPO; roadshow launched Nov 10 — not developer-technical, but significant organizational context. Source: anthropic.com/news
  • LangGraph 1.2.22, langgraph-prebuilt 1.1.2, langgraph-checkpoint-postgres 3.2.1, langgraph-checkpoint 4.3.2 (Nov 14, UTC 01:30–01:43): all-performance release — stream child graphs on own threads, skip message conversion for plain messages, pycryptodome bumped to 4.2.0 in checkpoint. No API changes. Source: github.com/langchain-ai/langgraph/releases

Worth Watching (Announced, Not Yet Shipped)

  • Gemini API preview model deprecations — November 17 (3 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 — verify at official URL). Action required if you use any of these previews.
  • Gemini 4 Argon pricing doubling — December 1 (17 days): Introductory rates ($2 input / $10 output per million tokens) confirmed by multiple independent sources to end November 30. Standard rates ($4/$20) take effect December 1. Carbon may follow the same schedule but is unconfirmed — verify on Google's pricing page. Audit your Gemini 4 Argon usage now if cost is a concern.
  • Fast mode removal for Claude Opus 4.8 — November 30 (16 days): speed: "fast" with claude-opus-4-8 will error. Migrate to Claude Opus 5.5. Source: https://platform.claude.com/docs/en/release-notes/overview
  • Claude Sonnet 4.5 retirement — November 30 (16 days): claude-sonnet-4-5-20250929 retires. Migrate to Claude Sonnet 5.5. Source: https://platform.claude.com/docs/en/release-notes/overview


Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.