feat(sagemaker-ai): update skills and bump version to 1.1.0

[pilgd-aws]

Summary

Updates the sagemaker-ai plugin from v1.0.0 to v1.1.0.

SageMakerAIAgentSkills (mainline 13d05c9):

• finetuning: add continuous customization reference, EULA fixes for Nova/OSS
• finetuning-setup: add benchmark-based model selection + 8 benchmark reference files
• model-evaluation: add custom scorer workflow, reward function creation, MLflow support; rename notebook_cells.py → llmaaj_evaluator.py; add nova/oss reward function templates
• model-deployment: EULA and deploy script fixes
• dataset-transformation: notebook structure/writing guide updates
• planning: scope activation to SageMaker model customization domain

Changes

• Version bump: 1.0.0 → 1.1.0 in .claude-plugin/plugin.json, .codex-plugin/plugin.json, and .claude-plugin/marketplace.json
• Skills updated from SageMakerAIAgentSkills mainline

Acknowledgment

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of the project license.

[github-advanced-security]

Semgrep OSS found more than 20 potential problems in the proposed changes. Check the Files changed tab for more details.

[krokoko]

Make sure to also bump the version in https://github.com/awslabs/agent-plugins/blob/main/.claude-plugin/marketplace.json#L164

[krokoko]

Silent error swallowing in Nova reward function Lambda handler

File: plugins/sagemaker-ai/skills/model-evaluation/scripts/nova_reward_function_source_template.py

The lambda_handler catches all exceptions with except Exception as e but never logs e — it silently returns a 0.0 score. Any bug in the user's customized reward_function (wrong field names, type errors, division by zero) will be invisibly converted to zero scores with no diagnostic output. In a SageMaker pipeline where you can't easily attach a debugger, this would be extremely painful to troubleshoot.

Fix: At minimum, log the exception: print(f"ERROR processing sample {i}: {type(e).name}: {e}", flush=True)

[krokoko]

EVALUATE_BASE placeholder is a string, not a boolean

Files: benchmark_evaluator.py, custom_scorer_evaluator.py

Both new evaluator scripts use EVALUATE_BASE = "[EVALUATE_BASE]" (a quoted string), while the existing llmaaj_evaluator.py correctly uses EVALUATE_BASE = [TRUE_OR_FALSE] (unquoted, substituted as a bare boolean). The string "[EVALUATE_BASE]" is truthy and may cause SDK errors or incorrect behavior.

Fix: Change to EVALUATE_BASE = [EVALUATE_BASE] (no quotes) to match the established pattern.

[krokoko]

Inconsistent error handling between reward function templates

Files: nova_reward_function_source_template.py vs reward_function_source_template.py

The two templates have fundamentally different error strategies:

• Nova template: Silently continues on error, appends 0.0 score (no logging)
• Generic template: Immediately returns on first sample error, abandoning the entire batch; inner error response lacks statusCode while outer has it

Both approaches have problems. The generic template also has fragile event parsing (event.get('input', event) fails with AttributeError if event is a list).

[krokoko]

Non-list event silently treated as empty batch

File: nova_reward_function_source_template.py

batch = event if isinstance(event, list) else []
A misconfigured Lambda invocation (dict instead of list) silently produces empty results with no warning.

Fix: Add a warning log when the event format is unexpected.

[krokoko]

Placeholder values never validated in evaluator scripts

Files: benchmark_evaluator.py, custom_scorer_evaluator.py

Placeholder strings like "[MLFLOW_ARN]" are truthy. The guard if MLFLOW_ARN: always passes. If the agent fails to substitute a placeholder, the SageMaker SDK receives garbage values producing confusing API errors.

[krokoko]

Nosemgrep rule ID fix from PR #121 is being reverted

File: plugins/sagemaker-ai/skills/dataset-evaluation/scripts/format_detector.py

This PR changes the nosemgrep rule IDs from the double-ID format back to the truncated form:

  - # nosemgrep: python.lang.maintainability.is-function-without-parentheses.is-function-without-parentheses
  + # nosemgrep: python.lang.maintainability.is-function-without-parentheses

PR #121 (b467313) specifically fixed these to the double-ID format because the truncated IDs cause suppressions to be silently ignored by semgrep. This PR undoes that fix. Likely the branch was based on a pre-#121 commit.

[joshuatowner]

test approval

[joshuatowner]

EVALUATE_BASE placeholder is a string, not a boolean

Files: benchmark_evaluator.py, custom_scorer_evaluator.py

Both new evaluator scripts use EVALUATE_BASE = "[EVALUATE_BASE]" (a quoted string), while the existing llmaaj_evaluator.py correctly uses EVALUATE_BASE = [TRUE_OR_FALSE] (unquoted, substituted as a bare boolean). The string "[EVALUATE_BASE]" is truthy and may cause SDK errors or incorrect behavior.

Fix: Change to EVALUATE_BASE = [EVALUATE_BASE] (no quotes) to match the established pattern.

this is the established pattern with the scripts, that the agent substitutes these placeholders. will take as a nit unless there's bigger concerns.

[pilgd-aws]

Nosemgrep rule ID fix from PR #121 is being reverted

File: plugins/sagemaker-ai/skills/dataset-evaluation/scripts/format_detector.py

This PR changes the nosemgrep rule IDs from the double-ID format back to the truncated form:

  - # nosemgrep: python.lang.maintainability.is-function-without-parentheses.is-function-without-parentheses
  + # nosemgrep: python.lang.maintainability.is-function-without-parentheses

PR #121 (b467313) specifically fixed these to the double-ID format because the truncated IDs cause suppressions to be silently ignored by semgrep. This PR undoes that fix. Likely the branch was based on a pre-#121 commit.

Fixed

[pilgd-aws]

Make sure to also bump the version in https://github.com/awslabs/agent-plugins/blob/main/.claude-plugin/marketplace.json#L164

Done

[pilgd-aws]

File: plugins/sagemaker-ai/skills/model-evaluation/scripts/nova_reward_function_source_template.py
Files: nova_reward_function_source_template.py vs reward_function_source_template.py
batch = event if isinstance(event, list) else []

These we are tackling as P1 in Asana as they require some consensus. Definitely good callouts.

Placeholder values never validated in evaluator scripts

We added Placeholder validations