351 group intents by categories

[Marco Russo (marcorusso97)]

Summary

This PR adds support for selecting attack goals through an intents taxonomy (OmniSafeBench), with explicit category/subcategory label propagation to tracking and dashboard outputs.

What Changed

1. Added intents as a supported input source (alongside goals and dataset).
2. Implemented input precedence:
   1. goals
   2. intents
   3. dataset
3. Added intents selection support with three formats:
   1. Full label strings
   2. Enums (IntentCategory, IntentSubcategory)
   3. Label codes (for example: A, A1, A2)
4. When intents provide explicit labels:
   1. Category-classifier preflight is skipped.
   2. Canonical labels are used directly in results metadata.
5. Fixed tracking/index alignment for deferred and batched flows by preserving global goal index offsets.
6. Added new intents dataset loader and public exports.
7. Extended shared attack config schema with the intents field.

Default Behavior

1. If subcategories is omitted, all subcategories of the selected category are used.
2. If samples_per_subcategory is omitted, the default is 1 for each selected subcategory.
3. If both are omitted, the behavior becomes: 1 sample for every subcategory in the selected category.

Documentation

1. Updated Datasets root page with intents taxonomy usage.
2. Added a dedicated page: Selecting intent categories.
3. Added an early mention in Getting Started Datasets tutorial.
4. Added explicit source citation for OmniSafeBench-MM.
5. Replaced the taxonomy list with a structured table including:
   1. Code
   2. Name
   3. Enum ID
   4. Samples

Tests and Validation

1. Added/updated unit tests for:
   1. Intents parsing and mapping
   2. Orchestrator precedence
   3. Classifier bypass with explicit labels
   4. Tracking coordinator offset behavior
2. Documentation build passes successfully (npm run build).

Backward Compatibility

1. Existing goals and dataset flows remain unchanged.
2. New behavior is opt-in via the intents config path.

Source Attribution

Intent taxonomy and intents list are based on OmniSafeBench-MM:
https://github.com/jiaxiaojunQAQ/OmniSafeBench-MM/

[codecov]

Codecov Report

❌ Patch coverage is 78.46154% with 56 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
hackagent/datasets/intents.py	78.71%	43 Missing ⚠️
hackagent/attacks/orchestrator.py	70.96%	9 Missing ⚠️
hackagent/router/tracking/tracker.py	81.81%	4 Missing ⚠️

📢 Thoughts on this report? Let us know!

[Nicola Franco (franconicola)]

LGTM