files.write buffering contract in JS: streams are no longer held in memory, and browser code on Firefox/Safari now throws. Separately, Agno wiring Claude Code and Codex permission requests into a managed approval UI is a small but notable signal: "human-in-the-loop for agent actions" is evolving from a team-by-team implementation problem into a framework-provided platform feature.gzip + useOctetStream: false combos now rejected with InvalidArgumentError; Firefox/Safari browser code needs Blob conversion. Audit before upgrading./analytics/usage, /analytics/cost) replace the old API — if you use usage/cost data programmatically, start the migration now.curl -fsSL https://raw.githubusercontent.com/JetBrains/thinkrail/main/install.sh | bash
powershell -c "irm https://raw.githubusercontent.com/JetBrains/thinkrail/main/install.ps1 | iex"
thinkrail ~/code/my-repo
files.write now streams byte-chunk iterables directly to the sandbox instead of buffering them in memory. Three breaking behavioral changes shipped alongside this in JS and Python.retainOutput option to command execution.gzip: true, useOctetStream: false (JS) or gzip=True, use_octet_stream=False (Python), it now throws InvalidArgumentError / InvalidArgumentException instead of silently gzip-overriding. Pick one: either keep gzip and let it default to octet-stream, or drop gzip. Do this before upgrading.ReadableStream and AsyncIterable file writes now throw in browsers without streaming request body support. Convert to Blob first: await new Response(stream).blob(). Chromium-based browsers continue to stream fine.TemplateError. If you manage your own templates, update envd before deploying SDK 2.62.0.Iterable[bytes] works only in Sandbox (sync); AsyncIterable[bytes] works only in AsyncSandbox (async). Mixing types throws InvalidArgumentException.retainOutput=False: For high-volume or long-running commands where buffering all stdout/stderr is costly, set retain_output=False and handle output exclusively via on_stdout/on_stderr callbacks. Callbacks still receive all output even with buffering off.e2b JS or Python SDK with files.write; you upload files into E2B sandboxes from browser contexts (Firefox, Safari, iOS); you use gzip combined with useOctetStream: false; you have long-running commands where buffering command output wastes memory.EffortModerate (audit files.write call sites, fix gzip+octet-stream combinations, add Blob conversion for browser code; straightforward once you know the patterns).analytics-usage-cost-2026-08-24 beta header: /v1/organizations/analytics/usage (token counts and request counts) and /v1/organizations/analytics/cost (spend in USD). Simultaneously, the existing Usage and Cost API and the Claude Enterprise Analytics API's usage_report, cost_report, user_usage_report, and user_cost_report endpoints were designated legacy — still supported, migration guides provided.anthropic-beta: analytics-usage-cost-2026-08-24 header; the Anthropic SDK and ant CLI send this automatically. The most significant capability upgrade is actor-level filtering: you can now see exactly which user or service account consumed tokens, enabling per-user quota enforcement and cost attribution at the API layer. Time-series buckets (daily, hourly, per-minute) make it straightforward to build spend dashboards directly from the API rather than exporting from the Claude Console.ClaudeAgent and CodexAgent permission requests now surface in the AgentOS approval UI and pause the run until a human approves or rejects. Previously, CodexAgent with no approval settings would let requests go unanswered; v3.1.12 changes the default to the on-request policy, bounding the pause with hitl_timeout (default: 900 seconds). Routines can now be created with enabled=False and activated later from the UI without redeployment. ChartTools.create_chart() gains an optional preformatted headline value shown beside the chart title. Bug fixes include: FastAPI tracing no longer polluting Agno traces, routine cancellations recorded only after the run stops, and AccuracyEval's default judge updated from the retired o4-mini to gpt-5.6-terra.ClaudeAgent or CodexAgent in Agno, the default behavior changes: runs now pause at permission request checkpoints and wait up to 900 seconds for approval. Review your hitl_timeout setting — for agents that frequently trigger permission requests in production, 15 minutes per checkpoint may be too long (or you may want this as a safety gate). The initially-disabled routines feature is useful for setting up scheduled agent jobs you want to validate before going live: create with enabled=False, test manually, then flip the switch from the AgentOS UI.ClaudeAgent or CodexAgent; you run agents that trigger Claude Code or Codex permission requests; you build scheduled automations in Agno Routines.EffortQuick (upgrade Agno; review hitl_timeout if agents trigger frequent approval requests; no code changes required).git worktree add management; you work across multiple concurrent agent tasks.ultrareview workspace-trust requirement (Oct 21), Anthropic's audit log expansion, and E2B's retainOutput pattern for surfacing command output without buffering it all. Each of these is a different mechanism for giving humans visibility or control over what agents do, without requiring teams to build the instrumentation themselves.Nothing cleared the quality bar this period — community sources (hnrss.org, simonwillison.net) DNS-blocked this scan.
getDiffContextSnippet, ambiguous-width terminal character handling, confirm_action flow no longer creates duplicate history entries or telemetry. Nightly build only; stable channel unchanged. (GitHub releases)Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.