0.9.1

What's Changed

🐛 Bug Fixes

• fix(jobs): cap Kubernetes client below incompatible v37 by @mckornfield in #3694

Full Changelog: 0.9.0...0.9.1

0.9.0

What's Changed

🚨 Breaking Changes

• feat(evals)!: add taskset run submit contracts by @ngoncharenko in #2895

✨ New Features

• feat(builder): let a service submit builds for a person by @albcui in #2760
• feat(scaled-evals): bind a separate verifier image for harbor_opensandbox tasks by @JashG in #2556
• feat(studio): plot tokens on the experiment Pareto chart by @marcusds in #2765
• feat(studio): draft a customization job with AI by @htolentino-nvidia in #2654
• feat(studio): enable trace graph feature flag by default [ASTD-767] by @marcusds in #2792
• feat(studio): show insights analysis run status and progress by @marcusds in #2774
• feat(nemo-agents): forward per-model extensions metadata to Fabric by @benmccown in #2747
• feat(studio): Enable switchyard optimizations w/ form by @steramae-nvidia in #2809
• feat(evaluator): publish failed scores to Intake as FAILED rows by @SandyChapman in #2595
• feat(evaluator): run a wheels-v1 Gym host offline when the deployment allows no egress by @SandyChapman in #2587
• feat(agents): add mini-swe adapter by @mmogallapalli in #2836
• feat(studio): pick the evaluator config from a fileset by @marcusds in #2812
• feat(studio): compare evaluation runs side by side on the experiment page by @marcusds in #2817
• feat(studio): auto-refresh the traces list by @marcusds in #2820
• feat(studio): explain why a finished customization job is still active by @walston in #2838
• feat(evals): add taskset run records and frozen plan storage by @ngoncharenko in #2844
• feat(studio): add an evaluation scorecard to the experiment page by @marcusds in #2819
• feat(builder): ship it with the platform, and publish its two images by @albcui in #2674
• feat(builder): add it to the platform's Helm chart by @albcui in #2681
• feat(scaled-evals): add the compose services spec for harbor_opensandbox by @JashG in #2801
• feat(studio): Switchyard optimization results view by @steramae-nvidia in #2839
• feat(docker): collect source for every open source component in the builder images by @SandyChapman in #2911
• feat(builder): accept any archive name, share reads, submit from the typed client by @SandyChapman in #2909
• feat(studio): show Data Designer preview rows in a table by @marcusds in #2821
• feat(scaled-evals): add an OpenSandbox server extension that runs compose services by @JashG in #2802
• feat(ci): add generic /helix-* PR-comment agent dispatch trigger by @benmccown in #2710
• feat(scaled-evals): run compose tasks on harbor_opensandbox by @JashG in #2803
• feat(studio): show failed evaluator results from Intake by @SandyChapman in #2596
• feat(Studio): Remove the "continue" button on the Fine-tune/DD job creation page by @htolentino-nvidia in #2868
• feat(evals): add image builds for task revisions by @SandyChapman in #2881
• feat(helm): ship the OpenSandbox compose services extension with the chart by @JashG in #2987
• feat(scaled-evals): add an OpenSandbox driver for harbor_opensandbox by @JashG in #2855
• feat(studio): Enable platform guard for platform readiness by @steramae-nvidia in #3010
• feat(scaled-evals): support Harbor 0.13.2 on harbor_opensandbox by @JashG in #2857
• feat(studio): run model evaluation as a full-page wizard by @nv-odrulea in #2956
• feat(evals): add idempotent taskset run submission by @ngoncharenko in #2896
• feat(evals): add taskset run read routes by @ngoncharenko in #2904
• feat(builder): copy published images, and use copies for prebuilt task images by @SandyChapman in #2986
• feat(studio): run agent evaluation as a full-page wizard by @nv-odrulea in #2958
• feat(scaled-evals): enforce task disk limits on harbor_opensandbox by @JashG in #2919
• feat(studio): Enable guard for platform readiness by @steramae-nvidia in #3047
• feat(scaled-evals): support agent bundles on harbor_opensandbox by @JashG in #3074
• feat(jobs): add cooperative pause mechanism by @soluwalana in #2793
• feat(studio): Programmatic rendering of optimization strategies by @steramae-nvidia in #3061
• feat: format docs for CLI by @mikeknep in #3123
• feat(studio): override feature flags from an in-app editor by @aray12 in #2708
• feat(jobs): add cooperative pause to Kubernetes and Volcano backends by @anubhutivyas in #2875
• feat(agent-hardener): let the safety model use its own endpoint and key by @koralchapnik in #3093
• feat(insights): update Compass integration and insight model by @shanaiabuggy in #2722
• feat(tools): add a manual acceptance check for a running instance by @ironcommit in #2787
• feat(evals): run Harbor taskset runs as Kubernetes task Jobs - k8s MVP by @ngoncharenko in #3084
• feat(evals): publish taskset runs to an Intake Experiment by @ngoncharenko in #3087
• feat(evals): size taskset run sandboxes with execution.sandbox by @ngoncharenko in #3089
• feat(evals): add --harbor-git task source to runs submit by @ngoncharenko in #3090
• feat(jobs): add cooperative pause to Docker and subprocess, and release rerun by @anubhutivyas in #2877
• feat(jobs): reclaim stopped job storage and record image digests by @anubhutivyas in #2869
• feat(studio): add the model details feature flag by @htolentino-nvidia in #3094
• feat(files): Filter/Return created_by and updated_by for filesets by @nakolean in #3149
• feat(studio): create a hook to track a model's related jobs, fileset and deployment by @htolentino-nvidia in #3096
• feat(studio): Recent Jobs add to Dashboard by @nakolean in #3060
• feat(evals): let agents install themselves on no-network Harbor tasks by @SandyChapman in #3121
• feat(jobs): add attempt-aware APIs and safe updates by @ironcommit in #3105
• feat(studio): render reasoning and tool calls in model chat by @steramae-nvidia in #3131
• feat(customization): add pause, resume, and rerun with tracker continuity by @anubhutivyas in #2886
• feat(customizer): update error classification library (#3165) by @anubhutivyas in #3234
• feat(files): add timing spans to DuckDB log queries (release/0.9) by @mckornfield in #3482
• feat(evals): support a custom evaluation task image by @mckornfield in #3505

🔧 Improvements

• refactor(sdk): remove the last Stainless SDK remnants and compat surface by @maxdubrinsky in #2678
• refactor(scaled-evals): claim evaluations and builds without row locks by @arpitsardhana in #2445
• refactor(scaled-evals): submit evaluation jobs from the API after commit by @arpitsardhana in #2446
• refactor(scaled-evals): remove Switchyard integration by @arpitsardhana in #2447
• refactor(scaled-evals): build evidence and archives from Platform Jobs by @arpitsardhana in ...

0.8.1

What's Changed

🐛 Bug Fixes

• fix(studio): keep date filter edits draft until Apply by @mckornfield in #3017
• fix(agents): allow workflow-only Fabric resolution warning by @mckornfield in #3204
• fix: backport CA certificate updates to 0.8 by @svvarom in #3215
• fix: address oidc token issues (#3163) by @mckornfield in #3222
• fix: backport LiteLLM security pins to release/0.8 by @svvarom in #3253
• fix(rl): skip unused Mooncake wheels in image builds by @mckornfield in #3265

📝 Documentation

• docs: add a known issue about sandboxing by @gabwow in #3147

Full Changelog: 0.8.0...0.8.1

0.8.0

What's Changed

🚨 Breaking Changes

• refactor(evaluator): move Harbor runtime modules into runtimes/harbor by @ngoncharenko in #2444
• refactor(evaluator)!: HarborRunnerTarget selects its agent from a source union by @SandyChapman in #2455
• fix(evaluator): propagate Gym env secrets through a shared source by @ngoncharenko in #2469
• refactor!: delete the generated nemo_helix SDK by @maxdubrinsky in #2635
• refactor(evaluator)!: GymRunnerTarget selects its agent from a source by @SandyChapman in #2482
• feat(cli)!: replace nemo agent with nemo describe by @mikeknep in #2738
• feat(optimization)!: add strategy titles and requires_config; rename legacy to config-tuning by @stefan-kickoff in #2910

✨ New Features

• feat(evaluator): add Harbor taskset integration test and notebook by @ngoncharenko in #2136
• feat(studio): Add QuickstartPanel for sandbox Storybook by @nakolean in #2403
• feat(evaluator-sdk): run a complete Fabric config in the Harbor agent by @SandyChapman in #2357
• feat(optimization): pass judge inference params to tunable_rag_evaluator by @marcusds in #2330
• feat(agents): expose enabled agent deployment methods [ASTD-668] by @marcusds in #2385
• feat(evaluator): propagate env_secrets from the Harbor runner by @ngoncharenko in #2399
• feat: orchestrate between the numeric and prompt optimizations by @gabwow in #1860
• feat(example-plugin): show the shared CLI pattern in middleware-configs commands by @mikeknep in #2457
• feat(customizer): serve saved job templates from the customization router by @htolentino-nvidia in #2099
• feat(studio): Customizer/Data designer start page redesign by @htolentino-nvidia in #2238
• feat(studio): Banner for Sample workspace by @nakolean in #2456
• feat(agent-optimization): show strategy descriptions in list-strategies by @mikeknep in #2492
• feat(studio): rename agent creation to registration [ASTD-681] by @marcusds in #2521
• feat(studio): turn on FF VITE_FF_AGENT_OVERVIEW_ENABLED by @nv-odrulea in #2524
• feat(studio): save a customization job as a template, and start from one by @htolentino-nvidia in #2516
• feat(evaluator): evaluate a registered agent through the Fabric runner by @SandyChapman in #2320
• feat(models): add family and model_providers filters to list models by @marcusds in #2537
• feat(evaluator): evaluate a registered agent through the Harbor runner by @SandyChapman in #2358
• feat(studio): Sample quickstart dashboard by @nakolean in #2514
• feat(jobs): add OpenShell-compatible nhx-tasks image variant by @maxdubrinsky in #2496
• feat(agents): add openshell deployment mode on OpenShell 0.1.2 by @maxdubrinsky in #2473
• feat(scaled-evals): Files-backed artifact storage backend + tests (1/4) by @benmccown in #2081
• feat(evaluator): live metric evaluation endpoint by @nv-odrulea in #2400
• feat(studio): trim the agent Details tab [ASTD-682] by @marcusds in #2522
• feat(scaled-evals): route artifact storage through the Files service (2/4) by @benmccown in #2122
• feat(studio): deploy agents on create with a method the platform can run [ASTD-659] by @marcusds in #2425
• feat(evaluator): propagate Fabric env secrets in localhost, sandbox, job runs by @ngoncharenko in #2510
• feat(plugin): add deployments HTTP consumer client by @benmccown in #2172
• feat(skills): add release test scope workflow by @vevaidyanath-lang in #2173
• feat(builder): core entities and contracts by @albcui in #2363
• feat(builder): compile a build request into a job and submit by @albcui in #2364
• feat(studio): Optimization Form behind FF by @steramae-nvidia in #2051
• feat(intake): represent failed evaluator results by @SandyChapman in #2593
• feat(studio): pick the eval wizard dataset from a fileset [ASTD-721] by @marcusds in #2651
• feat(studio): Deploy HPO job parameters in new agent by @steramae-nvidia in #2649
• feat(builder): signing and completion by @albcui in #2431
• feat(studio): build an agent image before deploying to Docker or Kubernetes [ASTD-714] by @marcusds in #2616
• feat(builder): build plane steps with sandbox supervisor by @albcui in #2432
• feat(builder): run it end to end on minikube by @albcui in #2509
• feat(evaluator): evaluate a registered agent through the Gym runner by @SandyChapman in #2490
• feat(setup): add api endpiont to create sample workspace and demo agent by @mmogallapalli in #2663
• feat(switchyard): add the switchyard optimization strategy by @stefan-kickoff in #2591
• feat(nhx-sandbox): add minimal shared sandbox package by @JashG in #2474
• feat(agents): add a prompt-master optimization plugin by @stefan-kickoff in #2229
• feat(evaluator): FieldMapping allows array element matches, with limts by @nv-odrulea in #2577
• feat(cli): add searchable setup model picker by @mckornfield in #2622
• feat(studio): show a blocking screen when NeMo Helix is unreachable by @aahunt-nv in #2623
• feat(studio): accept Parquet and CSV datasets in Run evaluation [ASTD-749] by @marcusds in #2667
• feat(scaled-evals): install Harbor runners from per-version requirements with OpenSandbox support by @JashG in #2475
• feat(rename): add reusable plugin rename profiles and Evals invocation by @gabwow in #2665
• feat(studio): Enable submission of optimization job through UI by @steramae-nvidia in #2656
• feat(agents): add remote agent fabric adapter by @mmogallapalli in #2687
• feat(studio): Support options for optimization strategies by @steramae-nvidia in #2680
• feat(studio): deploy HPO trials that tune an agent's system instructions by @marcusds in #2700
• feat(nemo-agents): add Pi (nvidia.fabric.pi) harness support by @benmccown in #2653
• feat(studio): send agent ethos with analysis runs and add Run analysis now by @marcusds in #2718
• feat(evaluator): filter tasks and tasksets by kind, metadata, metrics, members, tags and text by @SandyChapman in #2513
• feat(agents): add support for nooa fabric adapter by @mmogallapalli in #2698
• feat(scaled-evals): add OpenSandbox Harbor environment and ownership-scoped cleanup by @JashG in #2476
• feat(agent-hardener): install by default and rename the menu entry to Agent Hardening [ASTD-755] by @marcusds in #2696
• feat(builder): prefix job names, return image_ref, and filter images by set by @albcui in #2721
• feat(studio): Add sample workspace creation modal by @nakolean in #2724
• feat(scaled-evals): add harbor_opensandbox runtime backend by @JashG in #2477
• feat(cli): Remove --agent-mode by @mikeknep in #2740
• feat(anonymizer): run GLiNER PII detector in-process via Files HF pull-through by @benmccown in #2709
• feat(studi...

0.7.3

What's Changed

🐛 Bug Fixes

• fix(helm): preserve legacy Postgres claim labels by @svvarom in #2950

Full Changelog: 0.7.2...0.7.3

0.7.2

What's Changed

🐛 Bug Fixes

• fix(deployments): correctly track startup timeout failures by @mmogallapalli in #2920

📝 Documentation

• docs(release): add 0.7.1 patch notes by @mckornfield in #2824

🧰 Maintenance

• chore(docs): update dpo docs for known issue by @anubhutivyas in #2832

Full Changelog: 0.7.1...0.7.2

0.7.1

What's Changed

🐛 Bug Fixes

• fix(packaging): remove Insights Git dependencies from PyPI metadata by @svvarom in #2743
• fix(agents): replace lookalike domain in email triage tutorial by @svvarom in #2788

🧰 Maintenance

• build(evaluator): pin nemo-sandboxed-gym to the published 0.1.0 by @SandyChapman in #2689

Full Changelog: 0.7.0...0.7.1

0.7.0

What's Changed

🚨 Breaking Changes

• refactor!: rename Iron Swarm to Agent Hardener by @koralchapnik in #1848
• refactor(cli)!: drop the generated Stainless command tree and the CLI generator by @maxdubrinsky in #2004

✨ New Features

• feat(eval-author): harden trace environment generation by @asutermo in #1899
• feat(auth): implement access keys rotate by @anastasia-nesterenko in #1693
• feat(evaluator): add typed trial measurements by @ngoncharenko in #1896
• feat(plugin): Spec filter for jobs by @steramae-nvidia in #1890
• feat(studio): Customizer landing page by @htolentino-nvidia in #1931
• feat(evaluator): capture Relay's OTLP export so Fabric trials carry an OTLP trace by @SandyChapman in #1910
• feat(evaluator): capture OTLP inside the Fabric sandbox for container trials by @SandyChapman in #1915
• feat(studio): add insights analysis config, fix test util by @steramae-nvidia in #1774
• feat(studio): Agent Optimizations behind FF by @steramae-nvidia in #1941
• feat(studio): move the Create Deployment wizard to a dedicated page by @aray12 in #1947
• feat(evaluator): expose agent_kwargs on HarborRunnerTarget by @SandyChapman in #1996
• feat(studio): offer a deploy CTA when a model's chat is unavailable by @aray12 in #1948
• feat(studio): Agent Details with contextual CTAs by @steramae-nvidia in #1960
• feat(agents): record which GitHub commit a deployment is running [ASTD-604] by @marcusds in #2009
• feat(studio): add insights analysis config panel to agent details by @steramae-nvidia in #1776
• feat(studio): DeltaText component by @steramae-nvidia in #2030
• feat(studio): add trace format detection and ingest helpers by @steramae-nvidia in #1775
• feat(studio): Agent Optimization Job Details Route by @steramae-nvidia in #2029
• feat(scaled-evals): add benchmark-wide Harbor archive downloads by @asutermo in #2035
• feat(eval-author): add trace-derived tool-call access by @asutermo in #2010
• feat(studio): add Import traces modal by @steramae-nvidia in #1777
• feat(eval-author): measure failure-case coverage by @mstaats-nvidia in #1854
• feat(studio): create an agent from one seen in imported traces by @rrhyne in #2024
• feat(cli): typed-client plumbing for the Stainless-free CLI by @maxdubrinsky in #1986
• feat(evaluator): route Harbor agent credentials through env_secrets by @SandyChapman in #2032
• feat(studio): create an agent from a GitHub repository and show its commit [ASTD-593] by @marcusds in #2046
• feat(eval-author): help users build a working starter suite by @jessbuhl in #1943
• feat(eval-author): add authoring loop, per-check proof, and bounded repair to trace-environment by @asutermo in #2066
• feat(studio): upload an agent as loose files, not only as a whole directory [ASTD-610] by @marcusds in #2050
• feat(sandboxed-gym): vendor the Hugging Face cache into writable work by @SandyChapman in #2034
• feat(agents): upgrade to fabric 0.3.0 by @mmogallapalli in #2095
• feat(agents): Add ExecuteAgentJobConfig.image and agents config fields by @mikeknep in #2071
• feat(scaled-evals): run builds and evaluations with platform jobs by @arpitsardhana in #1887
• feat(studio): show deployment status for custom models and adapters by @aray12 in #2052
• feat(scaled-evals): project evaluations into entity store by @arpitsardhana in #2043
• feat(evaluator): run a NeMo Fabric deepagents harness on Nemotron inside Harbor by @SandyChapman in #2079
• feat(auth): add scopes to access keys creation by @anastasia-nesterenko in #1988
• feat(studio): collapse thinking blocks in assistant chat [ASTD-586] by @marcusds in #2132
• feat(automodel): wire deployment_config through to auto-deploy by @aray12 in #2063
• feat(cli): show provider status in inference providers list by @benmccown in #2159
• feat(inference-gateway): route LoRA adapters through the base model's VirtualModel by @benmccown in #2160
• feat(studio): package an agent image from Studio [ASTD-472] by @marcusds in #1659
• feat(studio): offer Deploy on custom models that are not being served by @aray12 in #2083
• feat(customization): link completed CLI submit to Studio job details by @benmccown in #2155
• feat(rl): support auto-deploy after a training job completes by @aray12 in #2064
• feat(studio): open the New Agent modal on Upload agent [ASTD-618] by @marcusds in #2175
• feat(auth): add ZITADEL Kubernetes IdP reference with opaque-token support by @ironcommit in #2070
• feat(studio): Add FeatureFlag for Dashboard Sandbox in Studio by @nakolean in #2185
• feat(models): validate model provider URL scheme at create time by @benmccown in #2157
• feat(agents): add nemo agents redeploy to update a deployed agent in one step by @benmccown in #2168
• feat(auth): add stable account identity resolution by @ironcommit in #2084
• feat(sandboxed-gym): release nemo-sandboxed-gym independently of the platform by @SandyChapman in #1906
• feat(studio): Chart other series on a customization detail page by @htolentino-nvidia in #2188
• feat(plugin): add job token usage reporter by @mckornfield in #2107
• feat(studio): Expand StatTile for actions by @nakolean in #2191
• feat(studio): storybook DashboardPanel by @nakolean in #2210
• feat(data-designer): Add user-facing check-models command by @mikeknep in #2214
• feat(evaluator): submit a Gym runner with secret refs and a placement by @SandyChapman in #2211
• feat(studio): deploy the base model for adapter fine-tuning jobs by @aray12 in #2085
• feat(evaluator): run Fabric harnesses on Harbor tasks that ship no Python by @SandyChapman in #2209
• feat(studio): surface error message when trying to delete fileset in use by @dmariali in #1942
• feat(studio): navigation consistency, persisted header state, and chevron polish (ASTD-582) by @walston in #2244
• feat(scaled-evals): run evaluations through Platform Jobs by default by @arpitsardhana in #2230
• feat(studio): non-sandbox Quickstart Dashboard by @nakolean in #2258
• feat(insights): use trace-intel for analysis and reconciliation by @shanaiabuggy in #2282
• feat(data-designer): report job token usage by @mckornfield in #2108
• feat(intake): import Gym rollout trajectories by @asutermo in #2226
• feat(anonymizer): report job token usage by @mckornfield in #2110
• feat(studio): offer deployment for runs that emit their own model by @aray12 in #2086
• feat(customizer): combine fileset upload with model and dataset creation by @anubhutivyas in #2249
• feat(safe-synthesizer): report job token usage by @mckornfield in #2111
• feat(evaluator): execute stor...

nemo-sandboxed-gym-v0.1.0

What's Changed

✨ New Features

• feat(studio): gate fine-tuning deploy UI on DEPLOYMENTS_ENABLED by @aray12 in #2329
• feat(evaluator): submit agent evaluations from the async plugin SDK by @SandyChapman in #2312
• feat(release): link nightly instance in Slack notification by @svvarom in #2328
• feat: add the api for general compute and prompt optimization by @gabwow in #1857
• feat(studio): enable the Inference provider feature flag by default [ASTD-644] by @marcusds in #2265
• feat(auth): support configurable Studio bearer tokens by @a2bondar in #2227
• feat(studio): launch agent HPO studies from the Optimize button [ASTD-650] by @marcusds in #2323
• feat(auth): support configurable CLI device flow by @a2bondar in #2228
• feat(studio): model evaluation form with steps and test by @nv-odrulea in #2116
• feat: implement ga prompt optimization by @gabwow in #1858
• feat: nemo chat supports configured default model by @mikeknep in #2342
• feat(setup): Create sample workspace and agent as part of setup flow by @mmogallapalli in #2334
• feat(studio): filter the Traces list by agent by @marcusds in #2348
• feat(studio): carry over the cookbook hyperparameters automodel now accepts by @htolentino-nvidia in #2380
• feat(studio): let Run evaluation set how many rows run in parallel by @marcusds in #2349
• feat(agents): strategy-based nemo agents optimize by @stefan-kickoff in #1862
• feat(evaluator): report job token usage by @mckornfield in #2113
• feat(agents): report job token usage by @mckornfield in #2114
• feat(plugins): stop installing Agent Hardener by default [NMP-244] by @marcusds in #2379
• feat(switchyard): add native IGW routing host by @tylersbray in #2325
• feat(optimization): backport numeric and prompt orchestration to release/0.7 by @gabwow in #2493
• feat(studio): give the fine-tuning form its own route by @aray12 in #2383
• feat(studio): Insights tab for Agent Details by @steramae-nvidia in #2606
• feat(studio): surface the evaluator's run outcome and coverage by @SandyChapman in #2584

🔧 Improvements

• perf(auth): fix permission merge and xdist groups in integration tests by @mckornfield in #2322
• refactor: take typed clients from get_nemo_client in non-job routes by @maxdubrinsky in #2333
• refactor(guardrails): remove generated SDK surface by @ironcommit in #2332
• refactor(jobs): hand the job path typed NemoClient clients by @maxdubrinsky in #2335
• refactor(intake): move last intake callers to typed clients and delete the generated SDK surface by @maxdubrinsky in #2337
• refactor: move service tasks and controllers off the Stainless SDK client by @maxdubrinsky in #2389
• refactor: finish stainless -> NemoClient migration for plugins, services, and scripts by @maxdubrinsky in #2390

🐛 Bug Fixes

• fix(studio): Fine tuning terminology by @htolentino-nvidia in #2268
• fix(evaluator): report why a sandboxed Gym host failed instead of a timeout by @SandyChapman in #2317
• fix(optimization): make the hermes chat-only example run on the platform by @marcusds in #2324
• fix(studio): remove dead Evaluate CTA from fine-tune results (ASTD-489) by @walston in #2336
• fix(scaled-evals): keep plugin out of default --all-packages installs by @ngoncharenko in #2341
• fix: restore valid CODEOWNERS teams by @svvarom in #2344
• fix(agents): make the email-security-triage example evaluable from Studio by @marcusds in #2346
• fix(tests): stop non-finite-expiry params drifting between xdist workers by @SandyChapman in #2376
• fix(studio): Use normalized entity name in constant by @mikeknep in #2361
• fix(insights): let analysis enable take the model pair and fail cleanly by @marcusds in #2347
• fix(auth): allow customization task service principals by @ironcommit in #2419
• fix(auth): ignore external OIDC scopes by @ironcommit in #2362
• fix(studio): render create-entity modals outside wizard forms by @aray12 in #2382
• fix(evaluator): emit empty rubric label when the judge reply does not parse by @SandyChapman in #2423
• docs(evaluator): state the Python version cloudpickle metric bundles need by @SandyChapman in #2427
• fix(evaluator): keep Gym runtime evidence inside the agent-eval bundle by @SandyChapman in #2429
• fix(evaluator): honor num_repeats and concurrency for sandboxed Gym jobs by @SandyChapman in #2428
• fix(auth): allow public API documentation by @a2bondar in #2466
• fix(evaluator): enforce Gym startup and collection timeouts on sandboxed runs by @SandyChapman in #2486
• fix(auth): forward external runtime scopes to the PDP by @ironcommit in #2525
• fix(train): align name max len to w/ deployment max suffix len by @soluwalana in #2468
• fix(sandboxed-gym): reuse the gym-host image's prebuilt component venvs by @SandyChapman in #2519
• fix(gym): set limits when requests is set by @soluwalana in #2481
• fix(evaluator): send LLM judge prompts in a shape the endpoint accepts by @SandyChapman in #2511
• fix(evaluator): fall back to the next reranker route on a gateway-wrapped 404 by @SandyChapman in #2512
• fix(studio): Improve schema and error messages when generating a DD config with AI by @htolentino-nvidia in #2463
• fix(evaluator): run each Codex trial in its own throwaway Codex home by @SandyChapman in #2555
• fix(cli): wait longer for services readiness and stop the child on timeout by @tylersbray in #2502
• fix(platform): report controller health under the --controllers names by @tylersbray in #2504
• fix(intake): retry local ClickHouse setup without blocking platform startup by @tylersbray in #2503
• fix: NOOA client None backend_format (#2548) by @mikeknep in #2585
• fix(insights): Allow service principals to read/write insights (#2570) by @mikeknep in #2586
• fix(studio): refresh workspace access check after first-login creation by @aray12 in #2571
• fix(agents): show container deployment URLs in the deployments list by @tylersbray in #2611
• fix(cli): clarify nemo setup API key validation UX by @tylersbray in #2615
• fix(evaluator): keep Fabric and Harbor agent-eval evidence inside the downloadable bundle by @SandyChapman in #2597
• fix(evaluator): score NaN when an optional regex judge group is unfilled by @SandyChapman in #2588
• fix(evaluator): fail eval jobs that produced no usable scores by @SandyChapman in #2567
• fix(sandboxed-gym): read OpenSandbox output text and lifecycle state by @SandyChapman in #2637
• fix(deps): bump wandb-core for gym images by @mckornfield in #2554
• fix(docs): update customizer tutorials and fileset docs to the typed files client by @anubhutivyas in #2646
• fix(train): resolv...

nemo-sandboxed-gym-v0.1.0-a1

What's Changed

🚨 Breaking Changes

• refactor!: rename Iron Swarm to Agent Hardener by @koralchapnik in #1848
• refactor(cli)!: drop the generated Stainless command tree and the CLI generator by @maxdubrinsky in #2004

✨ New Features

• feat(eval-author): harden trace environment generation by @asutermo in #1899
• feat(auth): implement access keys rotate by @anastasia-nesterenko in #1693
• feat(evaluator): add typed trial measurements by @ngoncharenko in #1896
• feat(plugin): Spec filter for jobs by @steramae-nvidia in #1890
• feat(studio): Customizer landing page by @htolentino-nvidia in #1931
• feat(evaluator): capture Relay's OTLP export so Fabric trials carry an OTLP trace by @SandyChapman in #1910
• feat(evaluator): capture OTLP inside the Fabric sandbox for container trials by @SandyChapman in #1915
• feat(studio): add insights analysis config, fix test util by @steramae-nvidia in #1774
• feat(studio): Agent Optimizations behind FF by @steramae-nvidia in #1941
• feat(studio): move the Create Deployment wizard to a dedicated page by @aray12 in #1947
• feat(evaluator): expose agent_kwargs on HarborRunnerTarget by @SandyChapman in #1996
• feat(studio): offer a deploy CTA when a model's chat is unavailable by @aray12 in #1948
• feat(studio): Agent Details with contextual CTAs by @steramae-nvidia in #1960
• feat(agents): record which GitHub commit a deployment is running [ASTD-604] by @marcusds in #2009
• feat(studio): add insights analysis config panel to agent details by @steramae-nvidia in #1776
• feat(studio): DeltaText component by @steramae-nvidia in #2030
• feat(studio): add trace format detection and ingest helpers by @steramae-nvidia in #1775
• feat(studio): Agent Optimization Job Details Route by @steramae-nvidia in #2029
• feat(scaled-evals): add benchmark-wide Harbor archive downloads by @asutermo in #2035
• feat(eval-author): add trace-derived tool-call access by @asutermo in #2010
• feat(studio): add Import traces modal by @steramae-nvidia in #1777
• feat(eval-author): measure failure-case coverage by @mstaats-nvidia in #1854
• feat(studio): create an agent from one seen in imported traces by @rrhyne in #2024
• feat(cli): typed-client plumbing for the Stainless-free CLI by @maxdubrinsky in #1986
• feat(evaluator): route Harbor agent credentials through env_secrets by @SandyChapman in #2032
• feat(studio): create an agent from a GitHub repository and show its commit [ASTD-593] by @marcusds in #2046
• feat(eval-author): help users build a working starter suite by @jessbuhl in #1943
• feat(eval-author): add authoring loop, per-check proof, and bounded repair to trace-environment by @asutermo in #2066
• feat(studio): upload an agent as loose files, not only as a whole directory [ASTD-610] by @marcusds in #2050
• feat(sandboxed-gym): vendor the Hugging Face cache into writable work by @SandyChapman in #2034
• feat(agents): upgrade to fabric 0.3.0 by @mmogallapalli in #2095
• feat(agents): Add ExecuteAgentJobConfig.image and agents config fields by @mikeknep in #2071
• feat(scaled-evals): run builds and evaluations with platform jobs by @arpitsardhana in #1887
• feat(studio): show deployment status for custom models and adapters by @aray12 in #2052
• feat(scaled-evals): project evaluations into entity store by @arpitsardhana in #2043
• feat(evaluator): run a NeMo Fabric deepagents harness on Nemotron inside Harbor by @SandyChapman in #2079
• feat(auth): add scopes to access keys creation by @anastasia-nesterenko in #1988
• feat(studio): collapse thinking blocks in assistant chat [ASTD-586] by @marcusds in #2132
• feat(automodel): wire deployment_config through to auto-deploy by @aray12 in #2063
• feat(cli): show provider status in inference providers list by @benmccown in #2159
• feat(inference-gateway): route LoRA adapters through the base model's VirtualModel by @benmccown in #2160
• feat(studio): package an agent image from Studio [ASTD-472] by @marcusds in #1659
• feat(studio): offer Deploy on custom models that are not being served by @aray12 in #2083
• feat(customization): link completed CLI submit to Studio job details by @benmccown in #2155
• feat(rl): support auto-deploy after a training job completes by @aray12 in #2064
• feat(studio): open the New Agent modal on Upload agent [ASTD-618] by @marcusds in #2175
• feat(auth): add ZITADEL Kubernetes IdP reference with opaque-token support by @ironcommit in #2070
• feat(studio): Add FeatureFlag for Dashboard Sandbox in Studio by @nakolean in #2185
• feat(models): validate model provider URL scheme at create time by @benmccown in #2157
• feat(agents): add nemo agents redeploy to update a deployed agent in one step by @benmccown in #2168
• feat(auth): add stable account identity resolution by @ironcommit in #2084
• feat(sandboxed-gym): release nemo-sandboxed-gym independently of the platform by @SandyChapman in #1906
• feat(studio): Chart other series on a customization detail page by @htolentino-nvidia in #2188
• feat(plugin): add job token usage reporter by @mckornfield in #2107
• feat(studio): Expand StatTile for actions by @nakolean in #2191
• feat(studio): storybook DashboardPanel by @nakolean in #2210
• feat(data-designer): Add user-facing check-models command by @mikeknep in #2214
• feat(evaluator): submit a Gym runner with secret refs and a placement by @SandyChapman in #2211
• feat(studio): deploy the base model for adapter fine-tuning jobs by @aray12 in #2085
• feat(evaluator): run Fabric harnesses on Harbor tasks that ship no Python by @SandyChapman in #2209
• feat(studio): surface error message when trying to delete fileset in use by @dmariali in #1942
• feat(studio): navigation consistency, persisted header state, and chevron polish (ASTD-582) by @walston in #2244
• feat(scaled-evals): run evaluations through Platform Jobs by default by @arpitsardhana in #2230
• feat(studio): non-sandbox Quickstart Dashboard by @nakolean in #2258
• feat(insights): use trace-intel for analysis and reconciliation by @shanaiabuggy in #2282
• feat(data-designer): report job token usage by @mckornfield in #2108
• feat(intake): import Gym rollout trajectories by @asutermo in #2226
• feat(anonymizer): report job token usage by @mckornfield in #2110
• feat(studio): offer deployment for runs that emit their own model by @aray12 in #2086
• feat(customizer): combine fileset upload with model and dataset creation by @anubhutivyas in #2249
• feat(safe-synthesizer): report job token usage by @mckornfield in #2111
• feat(evaluator): execute stor...