Claude API adds OpenTelemetry export, Claude Code compacts idle sessions before the cache expires, GitHub Actions gets dependency lockfiles
Today is about seeing and pinning what runs: the Claude API and Claude Code both feed standard OpenTelemetry traces, and GitHub Actions can lock dependencies to exact commits. Meanwhile Copilot’s HydraFusion hands model choice to an orchestration layer, which makes recording what each step actually used more important than ever.
Messages API requests and Claude Managed Agents sessions can be traced in your observability backend, in beta.
The Claude SDK emits spans in Python, TypeScript, C#, Go, Java, PHP and Ruby.
To get spans from the API side, add an OpenTelemetry destination in the Claude Console.
Builder's takeI have been wrapping model calls in my own logging; that layer can now move to standard tracing. I’d start by adding an OpenTelemetry destination for document Q&A in the AI Cloud Drive, so one question shows as a single trace from file parsing to the model’s answer and I can see which step is slow; before production I’d check exactly what the spans carry, then decide which backend they go to.
In the terminal, a 200K+ token conversation left idle for about 54 minutes is compacted before its one-hour prompt cache expires, so the next message doesn’t reprocess it all.
CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS changes the 200K minimum and "idleCompaction": false in settings.json turns it off; both need v2.1.290 or later.
Fixed a sandboxed shell command being able to write into a --plugin-dir folder, whose code Claude Code reloads and runs outside the sandbox.
/usage now warns about max effort, workflows, always-on context and large tool results; the Claude apps gateway no longer serves models blocked by policies such as deniedModels and answers 400 instead.
Builder's takeIdle compaction saves real money: coming back from a meeting to a large conversation used to mean paying full price to reread the context. I’d move the team to 2.1.299 mainly for the --plugin-dir sandbox fix, and raise the threshold for long tasks that need their context verbatim so compaction doesn’t drop details.
The model pool includes Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, and which one handles a task depends on the workflow.
Reasoning effort is set per stage, medium for initial drafts and high for review; the cascade adds a judging step that requests a repair only when the first draft needs it.
Critique now inspects the actual workspace with tools and runs parallel critic models, instead of relying mainly on text summaries.
HydraFusion and HydraFusion Max are in research preview in VS Code, the GitHub Copilot app and Copilot CLI, for Pro, Pro+, Business and Enterprise with experimental or preview features enabled; the page gives no pricing multipliers.
Builder's takeCheap drafts, expensive review, and repair only when needed is a split you can copy in your own agents. I’d try the same structure in PandaClaws content generation, with a mid-tier model drafting and a strong model only checking, but I wouldn’t switch Max on by default for the team until GitHub publishes its billing.
The lockfile at .github/workflows/actions.lock records exact commits for actions, transitive dependencies and reusable workflows, and runs use the recorded commits.
It is generated with the gh actions-lock extension for the GitHub CLI, can be adopted one workflow at a time, and can be checked for staleness in CI without changing files.
Enterprise, organization and repository admins can require lockfiles, also in public preview; when Dependabot updates an action it regenerates the matching lockfile entry in the same pull request.
GitHub notes that locking doesn’t show whether a dependency is safe, so review dependencies before locking them.
Builder's takeThis site’s data pipeline and deploys run on Actions with storage keys in the workflows, so a third-party action’s tag being moved is exactly what worries me. I’d generate lockfiles for the deploy and publish workflows that hold secrets first, add the staleness check in CI, and only then consider requiring it at the organization level.
In about 250,000 US Claude.ai conversations from September 21 to 27, about 5.75% touched politics, public policy or social issues, about 0.51% were about elections or voting, about 0.01% asked whom or what to vote for, and about 0.002% violated election integrity policies.
Later this month Claude.ai will draw official voting information from Democracy Works to give state-specific answers on where to vote and what ID to bring, and will show the latest official results on election night and the days after.
Anthropic says Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 score 99% on its political evenhandedness evaluation.
Reviews by three independent groups found Claude rarely refused but sometimes added unsolicited caveats and counterarguments, and the Foundation for American Innovation found Claude Sonnet 4.5 more likely to soften arguments for right-leaning positions.
Builder's takeThe lesson is the method: measure real usage in a privacy-preserving way first, then decide where to plug in authoritative data and where to add guardrails. My products don’t touch elections, but PandaClaws writes outward-facing content for people and faces the same question of unrequested counterarguments in advocacy copy, so I’d add a no-unrequested-caveats rule to generation and run a matched left and right prompt test.
The MOU covers AI safety, security and model evaluation, research on AI adoption and its impact, and responsible adoption in the public and private sectors, with planned technical exchanges on model evaluation, testing standards and incident response.
In the latest Economic Index, Singapore ranks second of 121 countries in Claude.ai usage relative to population, at nearly six times what its size alone would predict.
More than 1,500 engineers and technical staff at GovTech Singapore have adopted Claude Code, and Grab’s Claude-powered sales assistant BriX saves its sales teams about 40,000 hours a quarter.
Regulated organizations such as financial firms can run Claude with in-region inference on Amazon Bedrock to meet data processing requirements.
Builder's takeFor product builders the useful line is the last one: regulated industries need data to stay in-region, and that is met through the cloud provider’s regional inference. When I scope private deployments of the AI interview product for enterprise clients, I’d check whether in-region inference exists in the target region before discussing model choice, not after.
It is fine-tuned from Qwen3.5-4B with rank-64 LoRA and online supervised fine-tuning; input is text-only English multi-turn conversation and output is JSON with task and user_behavior fields.
On a 1 to 10 simulation fidelity scale it scores 4.95 / 6.02 best-of-1 / best-of-3, against 3.90 / 4.90 for GPT-5.5 and 3.41 / 4.42 for the Qwen3.5-4B base.
Its behavior descriptions re-identify the same user 51.02% of the time versus 25% at random; the card warns the representations are not anonymized data and lists profiling, surveillance and consequential decisions as out of scope.
Builder's takeThe hardest thing to test in the AI interview product is the variety of candidates, and this model turns real conversations into reusable simulated users. I’d try it on de-identified interview transcripts to build a set of candidate personas for regression tests; but its re-identification ability means the descriptions are sensitive data themselves, so they stay in an internal test environment and never get stored next to resumes.
ACE runs at 100 km and the new downscaling model HiROv2 refines its output to 3 km, predicting several high-resolution fields at once rather than only rainfall.
Generating global 3 km data costs about 1/300th of running X-SHiELD, a high-resolution model that explicitly simulates storms; for a single region the savings can exceed 100x.
Given coarsened X-SHiELD data, HiROv2 matches the target’s high-resolution statistics within 3%, including events that occur about once in a million data points; a streamlined version that could be more than 10x faster is in progress.
Builder's takeI don’t work on weather, but the shape is worth borrowing: a cheap coarse model plus a dedicated refinement model, paying for detail only where it’s needed. Read alongside this week’s small-model releases, cheap coverage first and a specialist model for precision second is becoming a general way to cut cost.
New repository_creation:write lets the caller create repositories without the broader repository_administration:write, and installation_repositories:write lets it add or remove repositories from any installation in the organization.
The organization-level APIs for adding and removing installation repositories now support GitHub Apps and fine-grained PATs, which previously required the coarse repo scope.
When an app is installed at the organization level, repository admins can add their repositories without owner approval; organization-level permissions, repository administration and repository creation still need it.
GitHub Apps now support offline_access and receive a refresh token with the expiring access token, and credentials move to a dedicated Credentials tab ahead of workload federation credentials.
Builder's takeThe usual compromise when giving an agent a GitHub token is granting full administration for a single create-repository action. I’d audit the team’s automation PATs and move whatever I can to a GitHub App with fine-grained permissions, alongside the push protection and push attribution from the last two issues.
Researched and drafted with AI assistance; editorial standards and views set by Darius.