GeezerVison

Closes #1779

Description

_set_skills_declarations() / _final_skill_declaration() in packages/gooddata-eval/src/gooddata_eval/core/agentic/conversation.py and _check_visualization_skill_activated in core/evaluators/visualization.py previously derived the active skill set solely from the set_skills tool call's arguments (what the agent requested).

However, what the agent requests is not always what actually becomes active:

• The service drops requested names it does not recognise.
• The service pulls in skills declared as dependencies of the ones requested.

The tool call's own result echoes back the authoritative post-replacement active set (e.g. skills_to_activate). Reading the result is strictly truer to what actually became active than re-deriving it from the request arguments.

Changes

• Implemented _extract_active_skills(tc) to prefer the authoritative post-replacement set in tc.parsed_result() (skills_to_activate, skill_names, skills), falling back to tc.parsed_arguments().get("skill_names") when the result is absent, unparseable, or errored.
• Updated _set_skills_declarations() in conversation.py to use _extract_active_skills().
• Updated _check_visualization_skill_activated() in visualization.py to use _extract_active_skills().
• Added unit tests in both test_agentic_conversation.py and test_visualization_evaluator.py covering:
  1. Unrecognised name requested (dropped by result -> not credited).
  2. Implicit dependency pulled in by result (not in arguments -> credited).
  3. Result missing (fallback path).
  4. Result unparseable (fallback path).
  5. Result errored (fallback path).
  6. Explicit empty result (deactivates skills).

Summary by CodeRabbit

• Bug Fixes
  • Skill activation and evaluation now reflect the skills reported by a successful tool result, including dependencies and explicit empty results.
  • When results are missing, unreadable, or indicate failure, evaluation falls back to the requested skills.