hermes - 💡(How to fix) Fix Background reviews currently inherit full tool surface and full history [1 participants]

Summary

Background review currently runs as a structurally heavier agent than necessary.

Even with configurable review-model routing, the review path still inherits:

• a full tool surface when enabled_toolsets is not restricted
• the full conversation history via messages_snapshot

That makes review payload size larger than needed, even when the actual review task is narrowly scoped to memory and/or skills.

Why this matters

This is a separate problem from review-model selection.

Using a cheaper review model helps with cost per token. Reducing review payload helps with the number of tokens sent per review call.

Both matter, but they should be discussed independently:

• PR #13647: configurable review model routing
• PR #13648: minimal toolsets + capped review history

Current behavior

In the current review path, the forked review agent:

• inherits the default full tool loading behavior unless enabled_toolsets is explicitly set
• receives the full messages_snapshot as conversation_history

That means background review can end up much heavier than its actual task requires.

Proposed direction

For background review, use a narrower default review payload:

• only enable the relevant review toolsets
  • memory review -> ["memory"]
  • skills review -> ["skills"]
  • combined review -> ["memory", "skills"]
• pass only a capped recent history window instead of the full conversation history

A minimal implementation can simply cap review history to the last N messages. No token-budgeting or major refactor is required for an initial improvement.

Why this should be separate from model-routing

This issue is about payload size, not model choice.

Even if review continues to use the main model, a smaller review payload is still beneficial:

• fewer prompt tokens
• less structural overhead
• lower likelihood of extra latency and rate-limit pressure

Scope

This is not proposing:

• prompt rewriting
• compression redesign
• max_iterations changes
• skip_context_files
• broader architecture changes

Just:

• smaller tool surface for review
• smaller history window for review

Related

• PR #13647 — configurable review model routing
• PR #13648 — minimal toolsets and capped history
• #5129
• #8506

If this direction is acceptable, I’m happy to keep the payload-size optimization separate from the review-model-routing discussion.