The one thing
01A long-context request returned another tenant’s data after every smoke test passed.
Cloudflare says 155 responses from its GLM 5.3 Workers AI deployment contained data from earlier customers’ requests between August 27 and September 9. Every affected prompt exceeded 681,280 tokens. The model advertised a 1,048,576-token context window, while the prefill KV pool held only 681,280 tokens. Cloudflare knew the capacities differed, but long-context tests returned HTTP 200 with streaming output. 1
The failure crossed two nodes. SGLang rejected an oversized prompt and replaced it with a one-token placeholder so the batch could drain. Prefill transferred that one page and reported completion; decode had reserved pages for the original prompt and treated them as overwritten. It then attended to stale pages left by previous requests. A successful response therefore hid the isolation failure. 1
Test the state boundary, not the status code. Operators using disaggregated SGLang should update to 0.5.22 or later, assert from the running engine that KV capacity covers the advertised context, and send canary-marked requests from separate tenants through rejection and abort paths. The acceptance check is that an over-capacity request fails before allocation or transfer and no marker from tenant A appears in tenant B’s output. Cloudflare says it now checks capacity directly and validated its fixes with the researcher’s reproduction and canary markers. 1
Shipped
03Claude Code 2.1.311 fixes stale instructions after compaction
Version 2.1.311 fixes standard-context sessions that stopped above 200K tokens with a 1M-context credit error; they now compact and continue. It also refreshes MEMORY.md and CLAUDE.md after automatic compaction, removes freezes caused by a very long list item, and lets an interrupt take effect while an MCP resource read or skill download is stuck. Upgrade machines that run long or unattended sessions, then repeat one session across compaction and confirm the edited instruction file is the one Claude follows. 2
Copilot Assisted permissions can auto-approve tool requests
Assisted permissions are generally available in Copilot CLI, the Copilot app and VS Code. A model classifies requests to run commands, access paths or open URLs; requests it calls clearly safe proceed, while the rest still ask. Enterprises can disable Assisted, block Allow All and choose the default for new sessions. Pilot it on a repository with an audit trail and replay known safe and unsafe requests before making it the default. 3
AI SDK 7.0.154 exposes step failures, aborts and package versions
AI SDK 7.0.154 adds onStepError to generateText, streamText, ToolLoopAgent and telemetry integrations, plus onError and onAbort for ToolLoopAgent. Telemetry now includes aiVersion, including in bundles with more than one SDK version. Upgrade agent loops where a failed step currently becomes a generic run failure, wire each callback to the run ID and step, and verify an abort is recorded separately from an error. Experimental steering also landed; keep that behind an explicit opt-in. 4
The conversation
01Automated permission decisions need a record of the request, policy decision and resulting action.
- Ethan Mollick (@emollick)Commentary · not a measured product study
Mollick argues that shorter reasoning summaries from Claude and ChatGPT make it harder for a person to intervene during work, pushing use toward automation rather than augmentation. The public post offers an operator’s interpretation, not a measured error rate. 5
Do not use a model’s narrated reasoning as the control record. Log the requested command, path or URL, the permission mode and decision, the executed action and its result. For an Assisted-permissions pilot, review false approvals and unnecessary prompts from that event stream before expanding scope.
The same rule applies below the agent layer: Cloudflare’s HTTP success concealed stale shared state. Observability has to cover the boundary whose failure would cross users, not merely the outer request.
The Mollick post is attributed commentary and could not be independently measured; GitHub documents Assisted behavior but publishes no classification error rate.
Skip this
04Cursor Agent SDK The SDK is available, but its new hosted and self-hosted operating contract needs a first-run evaluation before a recommendation.
Pydantic AI 2.62.0 Its compatibility changes span model settings, errors and deferred capabilities; compressing them would hide migration work.
Supabase scoped PAT announcement The feature became generally available on October 6, so a fresh social post is not a new release in this window.
Free coding-agent repository roundup The Reddit post is promotional discovery and does not establish a new upstream product or operating decision.
Everything we saw
9898 candidates scanned · 5 used in this issue — the rest, with the reason each one was left out
| Item | Source | Signal | Call |
|---|---|---|---|
| Cloudflare: cross-request KV cache exposure in GLM 5.3 | Vendor postmortem | Primary incident report with dates, counts, mechanism and remediation | One Thing: direct isolation failure with a reproducible operator check. |
| Claude Code 2.1.311 | GitHub release | Dated primary release notes | Shipped: compaction and interruption fixes change long-session operation. |
| GitHub Copilot Assisted permissions GA | GitHub changelog | Dated primary product announcement | Shipped: permission automation needs a bounded pilot. |
| AI SDK 7.0.154 | GitHub release | Dated primary release notes | Shipped: new callbacks make step failures and aborts observable. |
| Ethan Mollick on reduced reasoning summaries | X | Attributed commentary; not independently measured | Conversation: a useful operator concern, kept explicitly unverified. |
| Cursor Agent SDK | Vendor documentation | Available SDK and deployment documentation | Watch: evaluate a first project and access model before recommending adoption. |
| LangGraph 1.2.18 | GitHub release | Default delta snapshot frequency changed from 1000 to 100 updates | Watch: faster reads trade against storage and need a workload-specific measurement. |
| Pydantic AI 2.62.0 | GitHub release | Broad compatibility notes | Dropped: migration surface is too broad for the short edition. |
| OpenAI Python 3.39.0 | GitHub release | Adds the GPT-6.1 Terra model identifier | Dropped: no immediate upgrade trigger for clients that do not select that model. |
| Free Claude Code repository roundup | Promotional community post | Dropped: no verified upstream release or operating change. |
End of feed. That is everything from the window worth your time.
Explore more news, analysis and practical guides on NO FEED.