fix(memory): use primary model for background memory/skill review

[BukeLy]

What does this PR do?

_spawn_background_review creates a review AIAgent with self.model and self.provider, inheriting the current turn's model. When smart model routing sends the turn to a cheap/weak model (e.g. a local 9B), the background review also runs on that weak model — which typically lacks the tool-calling ability needed to actually save memories or create skills.

Fix: read the primary model config (model.default, model.provider, model.base_url, model.api_key) from config.yaml for the review agent, with fallback to the current model if config is unavailable.

See also: #7491 (complementary — passes base_url/api_key from parent agent; this PR goes further by reading from config since the parent may itself be a cheap routing model)

Related Issue

Fixes #8517

Type of Change

• Bug fix (non-breaking change that fixes an issue)

Changes Made

• run_agent.py: Modified _spawn_background_review → _run_review to read primary model from hermes_cli.config.load_config() instead of inheriting self.model
• tests/run_agent/test_background_review_model.py: New test file with 2 tests covering config-based model selection and fallback on config error

How to Test

1. Enable smart model routing with a cheap model that cannot do tool calls
2. Send enough messages to trigger memory nudge (nudge_interval turns)
3. Verify background review runs on the primary (strong) model, not the cheap model
4. Verify memories are actually saved (check ~/.hermes/memories/USER.md)

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes
• I've tested on my platform: macOS 15.4 (Sequoia), Apple Silicon

Documentation & Housekeeping

• N/A - no config keys or docs changed
• Cross-platform: all platforms affected equally, fix is platform-independent

[BukeLy]

Hi @jbarket @chalkers @leon @intertwine,
could you please take a look at this PR and help enable the CI run? Thanks!

[teknium1]

Thanks for the thoughtful PR, @BukeLy! This is a real concern, but the problem it solves no longer exists on main.

Automated hermes-sweeper review.

• The smart_model_routing feature — which caused the background review to inherit a cheap/weak turn model — was removed wholesale in commit 424e9f36b (PR refactor: remove smart_model_routing feature #12732, April 19 2026). The routing module, its config keys, docs, and tests are all gone.
• The related credential-inheritance issue (review fork not getting api_key/base_url/api_mode) was independently fixed in commit e3901d5b2 (PR fix(run_agent): background review fork inherits parent's live runtime #16099, April 26 2026), which already passes the full parent runtime into the review fork via _current_main_runtime().
• Current _spawn_background_review (run_agent.py line 3257) uses self.model — the parent's actual configured model — with no per-turn model-switching mechanism that could downgrade it.

Closing as already addressed on main. If a new per-turn routing feature lands in the future, the fix pattern here would be a good reference.