"type": "sdk" entries in .mcp.json are now silently skipped with a warning; MCP v2 is now default on Bedrock/Vertex/Foundry — audit your config files todayA quiet-news day for fresh trending tools — reaching into the watchlist for a still-underrated CI/CD tool that pairs exceptionally well with AI agent workflows.
curl -fsSL https://dl.dagger.io/dagger/install.sh | sh
dagger init --sdk=python my-pipeline
dagger call build --source=.
"type": "sdk" in .mcp.json, settings files, plugins, or agent files is now silently skipped with a warning rather than attempted — only an SDK host application can register in-process servers, and Claude Code no longer tries. Second: Bedrock, Vertex, Foundry, and telemetry-disabled installs are now defaulted to the MCP v2 client with the 2026-07-28 protocol negotiation when connecting to direct HTTP servers. The opt-out is MCP_SDK_GENERATION=v1..mcp.json and settings files for "type": "sdk" entries immediately. These were previously attempted and may have produced confusing errors; they now produce an explicit warning and are skipped. If you were relying on in-process SDK server registration via a config file, that path is gone — in-process server registration only works when done by the SDK host application at startup, not via a config file entry. (2) Check your Bedrock/Vertex/Foundry MCP server compatibility with MCP v2 (2026-07-28 protocol). If your MCP server only speaks the older protocol, you may see handshake failures. Set MCP_SDK_GENERATION=v1 in your environment to revert to the v1 client on these platforms while you upgrade your server. (3) MCP error messages improved: 403 insufficient_scope errors from MCP tool calls were previously reported as expired sign-in; they now report the correct cause. 422/4xx errors from HTTP MCP servers also no longer cause hard connection failures. (4) New env vars added: CLAUDE_CODE_MCP_STARTUP_WAIT_MS bounds how long a first non-interactive turn waits for MCP servers to connect (0 = don't wait, useful for scripted/pipeline usage where an MCP server may be slow to start). (5) /code-review now uses leaner inline prompts for models without tuned settings instead of spawning review subagents — if your /review skill invocations spawn many subagents, behavior changes here."type": "sdk" in any .mcp.json or settings file; you run Claude Code on Bedrock, Vertex, Foundry, or with telemetry disabled and have MCP servers that may not speak MCP v2 (2026-07-28); you built tooling around the review-subagent spawning behavior.EffortModerate — remove "type": "sdk" config entries (fix now, no workaround); verify your MCP servers speak the 2026-07-28 protocol or set MCP_SDK_GENERATION=v1 while you upgrade.No new model releases within the 24-hour scan window.
Carry-forward: Claude Fable 5.1 and Mythos 5.1 (September 1), GPT-6 Astra (September 3), Gemini 3.8 Flash (September 2) — see those dates' digests for full entries.
No new API or SDK additions outside of the Claude Code v2.1.274 changes above. Claude Agent SDK TypeScript last confirmed releases were v0.3.235/v0.3.232 (no new release confirmed in the 24-hour window). Anthropic Python and TypeScript SDKs: no new stable releases in window (last covered September 15–16 digest).
off (baseline), exact (reproducible numerical equivalence), fast (maximum throughput with minor floating-point variation), and big (for requests that exceed a single H100). (2) Coverage by family: co-folding and structure prediction (14 packages), structure generation (6), inverse folding (3), hallucination detection (3), genomics (7), protein language models (3). Primary configuration target is NVIDIA H100 80 GB on Linux x86-64; some kits additionally ship A100, H200, B200, or B300 configurations. (3) How Claude produced these: An internal general-purpose research model (not released; separate from the production Claude family) iterated on optimization passes per model, with human review at each stage. The research post documents the methodology. (4) Adjacent: protein design competition — Anthropic co-sponsored a protein design competition with Adaptyv Bio running alongside this release; check the research post for participation details. (5) This is the most concrete demonstration of Claude as a scientific computing accelerator rather than a general assistant — the 36-model scope and four-week timeline are verifiable claims.big mode requires multi-GPU setup and configuration.claude --remote-control sessions, assigning each a branch, and cross-checking their outputs yourself. The Projects coordinator now handles scoping, delegation, output review, and assembly. (2) Architecture: each delegated session runs in Anthropic-managed cloud infrastructure, on its own branch, with its own copy of the repository. Sessions run concurrently. The coordinator session waits for completion and assembles results. (3) Current limits: beta is for select Claude Pro and Max subscribers using cloud sessions — local/IDE sessions are not yet in scope. Access expands over the week following launch; Team and Enterprise access is planned but no date given. (4) How to use it now: if you have access, start a cloud session in Claude Code and request a multi-step goal; Projects will offer to coordinate it. There is no separate CLI flag — it is surfaced contextually in the conversation. (5) What to watch: this is the production version of the architecture Anthropic documented in the "Building a C compiler" post (~2,000 sessions, Linux kernel compilation). The coordinator is the key new piece — the parallel sessions themselves existed before; the structured delegation, progress tracking, and assembly layer is what's new. (6) Cost implications: multiple parallel cloud sessions running concurrently can consume your Pro/Max token budget faster than a single session; Anthropic's daily.dev coverage notes this directly. Monitor usage carefully in early experimentation.npm install -g @anthropic-ai/claude-code@latest — this is a large fix batch covering session stability bugs that were causing stuck sessions in production. (2) Sessions stuck on "unexpected tool_use_id" 400 errors now self-heal: a regression that caused sessions to retry a bad request indefinitely (requiring a manual restart) is fixed. If you've seen this error in the last few days, it was this bug. (3) Streamable HTTP MCP tool calls no longer time out at 5 minutes: a ceiling on long-running MCP tool calls (database migrations, slow test suites, file-system scans called via MCP) is removed. If you've been splitting long MCP operations to stay under 5 minutes, you no longer need to. (4) CLAUDE_CODE_MCP_STARTUP_WAIT_MS: set this to 0 in scripted/CI usage where you don't want the first turn to wait for MCP servers; useful for pipelines where MCP isn't needed. (5) CLAUDE_GATEWAY_DRAIN_TIMEOUT_MS (default 25,000 ms): controls how long the Claude apps gateway waits for in-flight requests to finish on SIGTERM rather than cutting all streams immediately. If you deploy Claude apps gateway and have seen dropped in-flight requests on restarts, set this env var explicitly.npm install -g @anthropic-ai/claude-code@latest; no config changes for most of the fixes; see Breaking Changes above for MCP items that may need attention).Nothing cleared the selection filter this period. The daily-trending list for September 17 was dominated by very-early-stage repos (694–2,071 daily stars, no established star-history curves) and AI-topic aggregate lists rather than concrete developer tools with verifiable momentum. AgentVerse-OS (personal cloud OS for running agents on a single server, Rust) was the most interesting candidate but star-history data was unavailable for momentum verification.
No new leaderboard entries or score changes confirmed in the 24-hour window.
Carry-forward (September 8–16 snapshot): Claude Opus 5 leads SWE-bench Verified at 96.0% (78 models). Claude Fable 5.1 leads SWE-bench Pro at 81.2% (58 models). Claude Fable 5.1-max leads LMArena at 1,513.70 Elo — three models above the historical 1,500 barrier.
Nothing cleared the quality bar this period. arXiv cs.AI/cs.CL/cs.LG were partially accessible via search but no papers from recognized labs with confirmed September 17 submission, associated code, and concrete benchmark numbers surfaced with sufficient verification for the main digest.
(Carried forward from September 11 digest) The gemini-omni-flash-preview endpoint retires September 30. GA replacement: gemini-omni-1.1-flash. Migration: update model string; verify resolution defaults (720p GA, not 360p as in preview). 13 days remaining.
GPT-6 Astra launched September 3, 2026 with free billing through October 4. Starting October 5: $10/$1 (cached)/$50 per MTok standard short-context. Long context (>272K tokens): $20/$75 per request. Batch/Flex: half price. Fast mode: double. Regional data residency: +10%. If you are running GPT-6 Astra in any pipeline, confirm your billing account and cost model before October 5. [platform.openai.com/docs/changelog]
DeepSeek confirmed V4.1 Pro is in development as the V4 Pro replacement. No launch date announced. V4 Pro retirement reversal (September 14) holds; V4 Pro continues at existing pricing until V4.1 Pro ships. [api-docs.deepseek.com/updates/]
Anthropic's Life Sciences Verification Program is initially open to teams and institutions; expansion to individual Claude Pro and Max subscribers is planned but no date given. [anthropic.com/news/life-sciences-verification-program]
Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.