Make background review toolsets configurable

[hunterbohm]

Summary

• Add memory.review_toolsets with the existing safe default of ["memory", "skills"].
• Resolve the background self-improvement review whitelist from config instead of hardcoding memory/skills.
• Whitelist both visible tool schemas and raw pre-Tool-Search schemas so configured MCP/plugin toolsets work when collapsed behind tool_search/tool_call.

Tests

• python -m pytest tests/run_agent/test_background_review.py tests/run_agent/test_background_review_toolset_restriction.py tests/run_agent/test_background_review_summary.py tests/run_agent/test_background_review_cache_parity.py tests/hermes_cli/test_plugins.py::TestThreadToolWhitelist -q -o 'addopts='
• python -m pytest tests/hermes_cli/test_config.py tests/hermes_cli/test_config_validation.py -q -o 'addopts='
• python -m ruff check agent/background_review.py hermes_cli/config.py tests/run_agent/test_background_review_cache_parity.py

[teknium1]

Thanks for making the runtime whitelist configurable. The current patch needs two corrections before it can preserve main's safety and Tool Search contracts.

Problems

• agent/background_review.py:510 replaces the current profile gate at agent/background_review.py:802-808 with a default containing memory. Commit 437dcacbb added that gate because a memory-disabled profile could otherwise receive the built-in memory tool; please retain it when composing configured toolsets.
• The MCP/plugin claim is incomplete for a review-only configured toolset. The fork still inherits the parent scope (agent/background_review.py:711-712), and agent/tool_executor.py:239-256,402-415 rejects an unwrapped tool_call outside that scope. The whitelist alone does not change it.
• The added test only mocks schema collection and captures the whitelist (tests/run_agent/test_background_review_cache_parity.py:304-326); it does not exercise actual bridge dispatch or the memory-disabled profile case.

Suggested changes

• Compose configured toolsets with the existing memory-enabled/user-profile-enabled gate.
• Add real Tool Search dispatch coverage for an in-scope configured deferred tool, plus a memory-disabled regression test.

Automated hermes-sweeper review.

[teknium1]

Blocking: this replaces main's current memory-disable gate (agent/background_review.py:802-808, commit 437dcacbb). Because the new default contains memory, a profile with both _memory_enabled and _user_profile_enabled false will again whitelist the built-in memory tool. Compose the configured list with that existing gate instead of replacing it.

[teknium1]

Blocking coverage gap: this test mocks schema resolution and stops before any tool dispatch. It cannot verify the claimed deferred MCP path, whose real executor scope comes from the inherited agent.enabled_toolsets (agent/tool_executor.py:239-256,402-415), nor catch the memory-disabled-profile regression.