Claude Code Daily Briefing - 2026-09-09

Release Summary

Two days after v2.1.263 (9/6), v2.1.265 shipped on 9/8, followed the same day by v2.1.266, which fixed a regression.

Full release notes

New Features & Practical Usage

--plugin-dir now manages an entire plugin folder dynamically (v2.1.265)

--plugin-dir can now point at a single folder containing multiple plugins at once. Any subfolder with a manifest loads as its own separate plugin, and adding or removing subfolders while a session is running is picked up immediately.

claude --plugin-dir ./team-plugins
# ./team-plugins/linter-helper/plugin.json
# ./team-plugins/deploy-checker/plugin.json
# Each subfolder with a manifest loads as a separate plugin,
# and adding or removing folders mid-session takes effect instantly, no restart needed

For teams distributing multiple plugins together, or testing different plugin configurations per project, this lets you manage plugins at the folder level instead of specifying --plugin-dir one at a time. Full release notes

Developer Workflow Tips

Improved /workflows agent detail view — see subagent progress at a glance (v2.1.265)

The /workflows agent detail screen now shows whether each individual tool call is running, failed, or complete, and if a subagent maintains its own task list, that list is displayed alongside it. Pressing Enter expands the listed tool calls so you can inspect their inputs and results directly.

If you’re running multiple agents in parallel with the Workflow tool, this screen alone lets you track progress without digging through logs to figure out where something stalled or failed. Full release notes

A skill that stops coding agents from stalling before they answer — I-have-ADHD (9/9)

A custom skill designed to make coding-assistant replies action-first, replacing long preambles with an immediately actionable task and a numbered procedure. Ten response rules suppress tangential explanation, force the assistant to restate its current state every reply, and require ending on exactly one concrete next step.

If Claude Code’s tendency to pad answers with long background explanations bothers you, this approach of locking response rules into a skill itself is worth referencing for your own CLAUDE.md or custom skills. GeekNews

Security & Limitations

v2.1.265 regression blocked gateway/proxy setups for a day, fixed in v2.1.266 (9/8)

The undocumented environment variable CLAUDE_CODE_USE_GATEWAY used to be ignored unless both ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN were set, but v2.1.265 started forcing Cloud gateway login based on this variable alone. As a result, any LLM gateway or proxy setup that had this variable set alongside an API key, apiKeyHelper, or a custom auth header saw every request fail with a “Not signed in to the Cloud gateway” error.

v2.1.266, released the same day, reverted the variable to being ignored on its own again, and updating alone restores normal behavior with no configuration changes needed. Teams running Claude Code behind their own LLM gateway or proxy should confirm they’ve updated to v2.1.266 or later, since deploying v2.1.265 even briefly could have blocked all requests. Full release notes

Claude service status — no new incidents since 9/3, all services operational

A direct check of the official status.claude.com shows no new incidents registered as of 9/9 since the 9/3 incident (elevated error rates across Claude Mythos 5.1, Fable 5.1, Opus 5, and other models, resolved within about an hour, 13:26-16:23 UTC), with claude.ai (99.4%), Claude Console (99.93%), Claude API (99.51%), Claude Code (99.44%), Cowork (99.43%), and Government (100%) all Operational.

As of the StatusGator check, user reports over the past 24 hours numbered 2, from Maharashtra, India (9/8 13:16, not working) and Provence-Alpes-Côte d’Azur, France (9/7 19:29, down). These read as scattered reports tied to specific regions and times rather than a pattern pointing to a broader service outage. Claude Status · StatusGator

Ecosystem & Plugins

Show GN: running GPT-6 Astra as the main model inside Claude Code (9/8)

A third-party tool that lets you run GPT-6 Astra as Claude Code’s main model while still mixing in Claude for subagents. It works as a local proxy that terminates TLS in front of the Anthropic API using its own CA (certificate authority) so it appears to be api.anthropic.com, connected via the ANTHROPIC_BASE_URL setting.

The upside is being able to mix and match models freely, but since the approach intercepts TLS with a self-issued CA, any organization with internal proxy policies or security requirements should first confirm this doesn’t conflict with audit or compliance requirements before adopting it. GeekNews

Community News

• Mistral raises 3 billion euros in Series D (9/8): Three years after founding, Mistral raised 3 billion euros, pushing its post-money valuation past 21 billion euros — the largest equity raise ever by a European tech company. Samsung Electronics led the round. It’s a sign that the open-weight model camp, which competes with and substitutes for Claude Code, keeps gaining financial firepower. GeekNews
• Meta unveils Muse, a personal AI agent (9/9): Rather than just answering questions, Meta’s personal AI agent carries out multi-step tasks like bookings, purchases, and document drafting across the web and connected apps. It handles web tasks inside Muse Secure VM, a persistent dedicated virtual machine with its own browser, and can be directed through the Muse app or a WhatsApp conversation. It’s a sign that competition over general-purpose autonomous agents — beyond coding — is spreading across Big Tech. GeekNews
• LibreOffice sets a new download record after touting “no AI features” (9/9): LibreOffice 26.8, a free Microsoft Office alternative, logged over 1 million installer downloads within a week of release, making it the most popular update ever. The strategy of touting the absence of built-in generative AI for privacy reasons, rather than despite it, paid off. It’s a counterexample showing that bolting on AI isn’t automatically the answer, worth keeping in mind for developers weighing whether to add AI to their own tools. GeekNews

Minor Changes

All of the following are from v2.1.265 (9/8).

• A 1GB cap was added to how much tool-result data gets saved to disk, and the conversation preview now shows whether a saved file was truncated.
• Fixed an issue where Read, Write, and Edit rejected every file when running inside an AppContainer or a restricted-token sandbox, caused by a change in how symlinks were resolved after permission checks.
• Fixed an issue where http-type MCP servers that only support the legacy HTTP+SSE transport failed to connect at all; it now falls back to SSE automatically as specified in the MCP spec.
• [VSCode] Added an “Archive inactive sessions” setting that automatically archives sessions with no activity for a set period (14 days by default).
• Fixed an issue where syntax highlighting in permission prompts and messages occasionally dropped the character following a Ruby ?, Erlang $, or Perl $ sigil.

Recommended Reads

• Behind the Navier-Stokes solution announcement: a mathematician’s account of the consultation with OpenAI (9/9): Mathematician Tristan Buckmaster published his fluid-equations research with Levent Alpöge and issued a statement laying out the consultation that took place before OpenAI’s Navier-Stokes solution announcement. The key point is that the pair’s work was a personal collaboration that advanced existing mathematicians’ approaches with the help of an LLM. Read alongside Terence Tao’s warning about AI solving math problems solo, covered in the 9/8 briefing, it gives a balanced view of how an AI lab’s announced results relate to the actual researchers’ contributions. GeekNews
• Build vs Buy — AI made development cheaper, but who handles maintenance three years from now? (9/8): AI has sharply cut the upfront cost of building software, but it hasn’t eliminated the operational cost of years of maintenance, security, and incident response. The core point: deciding whether to build in-house has to weigh not just the initial price tag but the blast radius of an outage or breach, who owns long-term operations, how much it differentiates the business, and how easy it is to switch later. Worth a read for any team deciding whether to build internal tools quickly with Claude Code or buy an existing solution. GeekNews
• Designing Grok Bot for the age of persistent agents (9/8): Grok Bot puts a list of Bots, rather than individual conversation histories, at the center of the product, so agents that preserve identity, memory, and execution environment can be trusted with long-term responsibility. The core design organizes the interface around five concepts: Bots, Chats, Prompts, Tools, and Artifacts. If you’re designing long-running agents like Claude Code’s background sessions or agent teams, it’s worth comparing how a competing product structures identity and memory. GeekNews

Interesting Projects & Tools

• Show GN: /take-my-data — a skill for locally scrubbing coding-agent sessions and donating them as an open dataset (9/8): A tool and hub that collects donated session logs from coding agents like Claude Code, Codex, Pi, and OpenCode and turns them into a CC0 public dataset. It started from the observation that almost no public data exists for Korean-language agent trajectories — sessions with tool calls, retrieval, and test results; running /take-my-data inside an agent scrubs sensitive information locally before donating the session. Worth joining for developers interested in building Korean-language coding-agent datasets or benchmarks. GeekNews
• Show GN: Hermeum — an open-source platform that builds and deploys the Hermes agent your team needs in one click (9/8): A project built to solve the silo problem created when every team at a company builds its own agent and wires it into Slack separately. It auto-configures a Hermes agent’s soul, skills, MCP, plugins, and cron to spin up a team-specific agent immediately. Worth a look as a standardized deployment platform if multiple teams are redundantly building their own agents. GeekNews