Changelog

All notable changes to DataRobot agent skills are tracked here.

The format follows Keep a Changelog, and the version numbers track the shared plugin version maintained across package.json, .claude-plugin/, .cursor-plugin/plugin.json, .codex-plugin/plugin.json, and gemini-extension.json.

Each entry should be prefixed with the affected skill folder name (for example, `datarobot-predictions`: ...) so it's easy to scan what changed per skill.

Version bumps, [Unreleased] renames, and releases are automated—see CONTRIBUTING.md.

[Unreleased]

• repo: Fixed OpenAI plugin listing: brand color now meets 2:1 contrast on white (Indigo #5C41FF), category set to Coding, and listing text no longer references other AI agents.

[1.14.0] - 2026-10-02

• repo: Added automated public Codex plugin packaging, release attachment, and manifest synchronization.
• datarobot-agent-assist: Expanded integration test coverage for how LLM choices flow from catalog listings into agent_spec.md and .env (gateway vs. deployed routing, setup_and_run failure paths, spec completeness), and added an on-prem / deployed-LLM example to agent-spec-examples.md.

[1.13.0] - 2026-09-30

Added

• datarobot-external-agent-monitoring: Recommend intelligent trace analysis via Tensile (tensile>=0.12.0). Each framework reference file gains an Intelligent Trace Analysis via Tensile section showing where tensile.otel.install(tracer=False) goes for that framework.

[1.12.0] - 2026-09-29

• datarobot-agent-assist: Clarified that the "After Coding" next-steps menu must be reproduced verbatim (same wording, order, and numbers) rather than paraphrased or renumbered, after an observed case where the assistant referenced "option 3" for a menu item that was actually numbered differently.

[1.11.0] - 2026-09-24

• repo: Added Codex plugin metadata and included the Codex manifest in the shared version bump workflow.
• datarobot-workload-api: Never gate CLI commands on a pinned dr version — always run dr self update --force first, matching datarobot-agent-assist's handling. Added references/declarative-cli-deploy.md documenting dr workload config/dr workload up, the CLI-native one-command create/build/deploy/resize path that replaces most manual REST choreography for project-directory deploys — noting up/config can return unknown command even on an otherwise-current CLI, independent of dr self update, with the manual flows as the fallback. Fixed the Code-to-Workload state directory (.datarobot/workload/, not .wapi/), codeRef field placement (imageBuildConfig.codeRef, not directly on the container — PATCH silently drops a misplaced value), dr workload logs --level semantics (minimum threshold, not exact match), and dr artifact code sync/build create/build logs behavior against the live v0.9.0 CLI (three-way diff with remote-wins auto-resolve, --wait, CANCELLED terminal status, structured JSON build logs). Documented the edge's forwarded identity headers and spoof-test results, sharing-role casing (OWNER/USER/CONSUMER, uppercase), the rolling-roll proton-watch trap, and that some instances omit Workload API paths from the public OpenAPI spec. Sharpened the web-UI auth-disable guardrail to require surfacing the tradeoff to the user, not disabling unilaterally.
• datarobot-workload-api: Rewrote SKILL.md and all references/*.md for token economy — terser phrasing, fewer filler words, tighter headings — with no change in technical content. SKILL.md dropped from 6647 to 6262 tokens (still under the 6700 hard cap); reference files dropped ~12% in word count combined.

[1.10.0] - 2026-09-22

• datarobot-agent-assist: Unified and increased timeouts for git commands and long running task commands in helper scripts.

[1.9.0] - 2026-09-21

• datarobot-agent-assist: Increased timeouts to prevent failing some of the long running scripts/tools.

[1.8.0] - 2026-09-02

Added

• datarobot-agent-assist: Dress rehearsal exports a shareable Markdown report to <target_dir>/rehearsal_report/rehearsal_report.md (archived under .datarobot/rehearsal/<session_id>/). DONE always runs --report before any summary or menu; post-design next steps add Review rehearsal report. NOTE: observations persist via rehearsal.py --note. Session state lives under .datarobot/rehearsal/ instead of the system temp dir.

Changed

• datarobot-agent-assist: Bumped application template version to 11.11.6.
• datarobot-agent-assist: Deduplicated SKILL.md by pointing dress rehearsal, clone discipline, and spec validation at the reference files instead of restating them.
• datarobot-agent-assist: Welcome menu infers a clear free-text category (and asks when ambiguous). frontend.type may be held from the Clarification Phase before the first spec draft exists.
• datarobot-agent-assist: Tool/service secrets stay in .env only — coding appends VAR_NAME= and asks the user to paste values in their editor, never in chat, agent_spec.md, or source.
• datarobot-agent-assist: Spec-only / messy-cwd classification treats dress-rehearsal artifacts (.datarobot/rehearsal/, rehearsal_report/) as design-phase files so they do not force a subdirectory clone.

Fixed

• datarobot-agent-assist: Rehearsal reports pair parallel tool returns with the matching call (previously every return was labeled with the last tool). model_substituted / simulation_substituted now compare against the originally requested model so a runtime fallback does not leave a stale flag.

[1.7.0] - 2026-09-02

Added

• datarobot-agent-assist: Support an OpenAI-completions-compatible external LLM (AGENT_ASSIST_LLM_MODEL_NAME, AGENT_ASSIST_LLM_API_KEY, AGENT_ASSIST_LLM_BASE_URL) as a model source alongside the LLM Gateway and deployed models, so list_llm_models.py and setup_template.py --llm-base-url can wire up a template without a DataRobot endpoint or token. Also lets clone_template.py override the template repo via AGENT_ASSIST_TEMPLATE_REPO_URL/_BRANCH/_TAG, with branch now taking priority over tag when both are set.

[1.6.0] - 2026-08-26

Added

• datarobot-llm-gateway: Added a new skill which helps setting up llm gateway, it lists available llm models directly from SKILL.md. The CLI handles auth via its own credential store, and syncs env variables.

[1.5.5] - 2026-08-26

Fixed

• datarobot-model-monitoring: The Pattern 1 health-check example read stats.prediction_count and stats.mean_response_time, which do not exist on the SDK 3.x ServiceStats object (AttributeError). Read them from the .metrics dict instead (stats.metrics["totalPredictions"], stats.metrics["responseTime"]). Also corrected the SDK reference list: model.get_metrics() → model.metrics (a dict of {metric: {partition: score}}); get_metrics() does not exist.

[1.5.4] - 2026-08-25

Fixed

• All skills: helper scripts and examples now call plain dr.Client() instead of re-reading DATAROBOT_API_TOKEN and DATAROBOT_ENDPOINT and passing them back in. The SDK reads those variables itself and then falls back to ~/.config/datarobot/drconfig.yaml, so this fixes setups where only drconfig.yaml is configured — what dr auth login (per datarobot-setup) writes. Previously most sites hardcoded a "https://app.datarobot.com" endpoint default, which supplied an endpoint with no matching token and failed with ValueError: Token must be specified if endpoint is specified (and was missing the /api/v2 suffix); datarobot-model-explainability used os.environ["DATAROBOT_API_TOKEN"] and failed the same setup with KeyError.

[1.5.3] - 2026-08-25

Fixed

• datarobot-model-training: Update the sample code and helper scripts for the current datarobot SDK (3.x). set_target() → analyze_and_model() (sets the target and starts AutoPilot), the removed Project.start(autopilot_on=, max_wait=) → wait_for_autopilot(), project.status → project.stage, model.get_metrics() → model.metrics, and select the recommended model via ModelRecommendation.get(project.id).get_model() instead of a broken max(models, key=lambda m: m.metrics.get("AUC", 0)) that treats a per-partition dict as a scalar. list_models.py now sorts on the validation partition score (null-safe).

[1.5.1] - 2026-08-13

Fixed

• datarobot-agent-assist: LLM_DEFAULT_MODEL now gets the datarobot/-prefixed llm_default_model value, not the catalog llmId the gateway 404s on. setup_template.py refuses an llmId; api_model stays unprefixed for the on-the-wire rehearsal.
• datarobot-agent-assist: The model table leads with LLM_DEFAULT_MODEL (not the unusable llmId) and gives the deployment id its own column, shown only when a deployed entry is present.
• datarobot-agent-assist: The dress rehearsal strips the datarobot/ prefix before reading a provider, so a prefixed spec keeps its cross-provider guard instead of matching nothing and rehearsing against an arbitrary catalog pick.
• datarobot-agent-assist: setup_template.py verifies the model with a direct catalog API call (not dr llm-gateway list, which can fall back to a stored profile and answer about a different instance), rejects characters that would break the .env line, and treats a disabled gateway as a cue to pick a deployed LLM.

[1.5.0] - 2026-08-10

Added

• datarobot-agent-assist: Make a DataRobot-deployed LLM selectable end to end. agent_spec.md gains an optional llm_deployment_id, and setup_template.py gains --llm-deployment-id, writing LLM_DEPLOYMENT_ID, INFRA_ENABLE_LLM=deployed_llm.py, and USE_DATAROBOT_LLM_GATEWAY=0 so the template routes to the deployment instead of the LLM Gateway. Model selection recommends a deployed LLM when the gateway is empty or disabled, which is the normal shape of an on-prem install.

Fixed

• datarobot-agent-assist: DataRobot-deployed LLMs are listed on every dr version the agent template accepts. dr llm-gateway list only reports them from v0.2.79 while the template's minimum is 0.2.77, and a non-empty gateway was enough to skip the direct-API fallback, so the deployed source silently did not exist on an older CLI.
• datarobot-agent-assist: setup_template.py refuses the shared datarobot-deployed-llm placeholder without a deployment id, instead of leaving the template on its gateway configuration to fail later at pulumi up with Model 'datarobot-deployed-llm' not found in catalog. The placeholder is matched case-insensitively, since the value comes from LLM-authored spec text.
• datarobot-agent-assist: The dress rehearsal resolves a deployed LLM through the spec's llm_deployment_id. Resolving on model alone matched whichever deployment the catalog indexed last and reported it as an exact match, so a spec could silently rehearse against a different deployment than the one selected.
• datarobot-agent-assist: list_llm_models.py closes stdin on the dr subprocess, so a credential prompt fails immediately rather than waiting out the timeout, and it names the requested instance alongside the CLI's log lines, making visible the case where the CLI ignored the passed credentials and listed a different instance.
• datarobot-agent-assist: Deployment labels are collapsed to one pipe-free line when rendering the model table; an embedded newline in user-authored label text split a row apart.
• datarobot-agent-assist: SKILL.md documented list_llm_models.py without its required --target-dir, and its Helper Scripts section referenced an undefined <scripts_dir> placeholder instead of the <skill_scripts_dir> the skill resolves. Both made the documented invocations dead instructions.
• datarobot-agent-assist: ensure_env_file printed its progress line and the dr dotenv setup output to stdout, so list_llm_models.py --json produced unparseable output on a target directory with no .env. Both now go to stderr.

[1.4.3] - 2026-08-05

Changed

• datarobot-agent-assist-simulate: Refine UX for swarm simulation.

[1.4.2] - 2026-08-04

• datarobot-model-training: Create/associate a dr.UseCase when creating datasets and projects, so projects aren't orphaned in the DataRobot UI.

[1.4.1] - 2026-08-03

Changed

• datarobot-agent-assist-simulate: UX improvements — domain-aware Q2, grouped scenario list, dr dotenv update in auth setup, live progress narration during swarm run.

[1.4.0] - 2026-07-28

Added

• datarobot-agent-assist: New agent-assist-simulate swarm skill — adversarial scenario generation, multi-turn simulation, convergence loop, and evaluation reporting.

Changed

• datarobot-agent-assist: Move check_codespace.py into agent-assist-build/scripts/ alongside env_utils.py.
• datarobot-agent-assist: Pin ruff==0.15.22 in dev dependencies to prevent silent version drift.

[1.3.10] - 2026-07-28

• datarobot-agent-assist: Merge pre-coding spec gate into pre-coding-checklist.md Bootstrap step 2 (removes duplicate "Before Coding Begins" section). Add missing-spec recovery, path validation, session flags, workspace ask-don't-guess for Code, and journey deduplication. Trim SKILL.md: move CLI setup, helper scripts, plugin tool mapping, dress rehearsal prompt, path resolution, and agent_spec schema into references. Flow polish: Code-no-spec merges with Path resolution step 1, schema read hook in Spec Display, design_to_code guard and Windows stop in pre-coding, welcome menu resets design_messy_cwd.

[1.3.9] - 2026-07-22

• datarobot-workload-api: Add references/web-uis-behind-the-edge.md for serving a browser-facing web app (UI + backend) through the workload endpoint — the edge gateway strips the path prefix (inbound shim + base-path derived from the injected WORKLOAD_ID; proton-id path needs an explicit override), is itself the auth gate and hijacks the Authorization header (so disable the app's own auth and trust the edge), shared-origin _xsrf/CSRF collisions, and WebSocket pass-through. Correct artifact-replacement guidance: same-artifact replacement is allowed for drafts, PATCH /settings/ rolling-redeploys onto a rebuilt image, and imageUri is build-managed (never hand-PATCH it or PATCH the spec mid-build). Trim SKILL.md ~18% to stay within the context-window budget.

[1.3.8] - 2026-07-17

• datarobot-agent-assist: Warn when the ports needed for local agent testing (5173, 8080, 8842) are not exposed inside a DataRobot Codespace, and stop with guidance when Agent Assist runs from an unsupported working directory. New check_codespace.py helper wired into the Pre-requisite Check; no-op outside a Codespace.

[1.3.7] - 2026-07-14

• datarobot-external-agent-monitoring: Fix PydanticAI instrumentation — modern PydanticAI makes OpenTelemetry instrumentation opt-in, so configuring a provider alone emitted no spans; document the required Agent.instrument_all() call.

[1.3.6] - 2028-07-14

• datarobot-agent-assist: Implement pre-coding-checklist, pre-deployment-checklist & workspace-resolution flows.

[1.3.5] - 2028-07-08

• datarobot-agent-assist: Bumped application template version to 11.10.7.

[1.3.4] - 2028-07-01

• datarobot-agent-assist: Improved rehearsal flow to properly handle missing LLM model cases and improved user-facing messaging; refactored Dress Rehearsal instructions into separate references/dress-rehearsal.md file to reduce SKILL.md token count while preserving all behavior and control-flow reliability.
• datarobot-discover: New skill for discovering DataRobot resources — fetches the live catalog from datarobot.com and, if set, from $DATAROBOT_ENDPOINT to surface skills, MCP servers, agents, and platform resources without search index dependency.

[1.3.3] - 2026-06-25

• datarobot-external-agent-monitoring: Support instrumenting existing (brownfield) agents — built on DataRobot or elsewhere — via the "Add tracing to my agent" trigger; resolve agentless invocations to the current IDE workspace; make a DataRobot Use Case the primary telemetry target (validate an existing Use Case ID, or create a net new one via the new create_use_case.py helper) with the shell deployment now optional; extract the OTel config template into reference/dr_otel_config.md; make verify_otel_connection.py accept the experiment_container- (Use Case) entity prefix in addition to deployment-; defer Use Case creation/validation to the post-approval execute step (prerequisites only collect the choice) to avoid premature or duplicate creation; collect the API token via the project .env file rather than chat (never asked for or echoed in the transcript).
• datarobot-setup: Broaden trigger to cover credential failures; add env var and auth validity checks to pre-flight.
• datarobot-workload-api: New skill for the DataRobot Workload API — create/configure, diagnose (CrashLoopBackOff / ImagePullBackOff / OOMKilled / exec format error), observe (logs/traces/metrics/stats), and artifact lifecycle (draft→lock→production, rolling replacement, promote, Code-to-Workload via dr workload code sync when no accessible registry). Modal SKILL.md + bundled scripts/ + deep references/.
• datarobot-setup: Broaden trigger to cover credential failures; add env var and auth validity checks to pre-flight.
• datarobot-model-explainability: Correct SHAP export guidance for datarobot.insights.ShapMatrix (in-memory matrix/columns or classmethod get_as_dataframe/get_as_csv); fix compute_shap_matrix.py --output export; fix anomaly assessment date-range example to use get_explanations() instead of get_latest_explanations(); fix Model diagnostics examples (get_confusion_chart, get_feature_effect); document insights diagnostics (RocCurve, LiftChart, ConfusionMatrix); correct documented SHAP caveats for blenders, the >1000-feature limit, ShapImpact source support, logit-link probability conversion, XEMP contribution wording, XEMP routing guidance, and XEMP max_explanations limit; raise the documented minimum SDK version to datarobot>=3.6.0 when referencing ShapDistributions.

[1.3.1] - 2026-06-02

• datarobot-setup: Corrected issues with setup commands.

[1.3.0] - 2026-05-27

• datarobot-model-explainability: Updated SHAP guidance to use the current datarobot.insights APIs, added data slice and anomaly assessment coverage, added SHAP and XEMP reference docs, and added a compute_shap_matrix.py helper script.

[1.2.0] - 2026-05-20

First tracked release. Skills included:

• datarobot-agent-assist
• datarobot-app-framework-cicd
• datarobot-data-preparation
• datarobot-external-agent-monitoring
• datarobot-feature-engineering
• datarobot-model-deployment
• datarobot-model-explainability
• datarobot-model-monitoring
• datarobot-model-training
• datarobot-predictions
• datarobot-setup