Subagents: JS, UI

[alex-w-99]

Issues addressed:

• LLMModel type alias is not callable at runtime #124

Other updates:

• New AbstractSpecialist base class with @specialist_tool decorator for declarative tool registration. Supports conversational and autonomous modes with finalize gating.
• Added JSSpecialist for analyzing JavaScript sources, identifying dynamic values, and building JS-based operations.
• Added InteractionSpecialist for browser DOM interactions (clicks, form fills, navigation). Uses interactions_data_store for managing captured user interactions.
• Added runner scripts for standalone specialist execution (run_js_specialist.py, run_interaction_specialist.py).
• Refactored js_utils.py with cleanup and improvements to JavaScript code generation.

Note: Existing specialist agents like NetworkSpyAgent do not yet inherit from AbstractSpecialist but should be migrated in the future.

Next steps:

• Consider having an AbstractDataStore / AbstractDataLoader. Data stores share a lot of code. Additionally, methods like search_by_regex should be available in all data stores, but is currently only available in JSDataStore.

[alex-w-99]

@claude Review my PR please.

[dimavrem22]

looks like the interaction agent replies with a summary but not the structures result via finalize tool

[alex-w-99]

looks like the interaction agent replies with a summary but not the structures result via finalize tool

Good catch, looks like this was just because the agent was getting confused due to the complexity of the output data model we expect/require of it. I've updated some prompts in commit 8969e4a and seems to work now.

[claude]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.

[claude]

Code Review

Found 1 CLAUDE.md import grouping violation in bluebox/llms/infra/js_data_store.py at line 14.

Issue: Import Grouping Violation

The third-party import import regex is placed between stdlib imports, violating CLAUDE.md's import grouping rule.

Location: bluebox/llms/infra/js_data_store.py:10-22

All stdlib imports should be grouped together, followed by a blank line, then third-party imports.

Current structure:

import fnmatch
import json
import threading

import regex  # third-party
from collections import Counter  # stdlib
from dataclasses import dataclass, field  # stdlib
from pathlib import Path  # stdlib
from typing import Any  # stdlib
from urllib.parse import urlparse  # stdlib

Correct structure:

import fnmatch
import json
import threading
from collections import Counter
from dataclasses import dataclass, field
from pathlib import Path
from typing import Any
from urllib.parse import urlparse

import regex

[dimavrem22]

LG