Compress per-step agent prompts and add rules-first action routing

[PraneelBhatia]

Problem

The step-level prompt in backend/agent.py is repeatedly sending too much page state:

• large system prompt
• browsing history
• error context
• visible text dump
• interactive element dump

This creates heavy repeated token usage.

Proposal

Compress the per-step context:

• replace raw visible text with a page summary
• send only top-k interactive candidates, not the full element list every time
• send delta history instead of repeated history
• send only new errors since last step
• keep invariant prompt blocks much shorter

Add a rules-first layer before LLM action selection for obvious cases:

• form discovery
• heading checks
• missing alt text
• click target sizing
• dead links
• obvious nav actions

Suggested file targets

• backend/agent.py
• backend/browser_utils.py
• maybe a new summarizer/router helper

Acceptance criteria

• step prompt payload is significantly smaller
• obvious interaction choices are resolved without an LLM call when possible
• logs show reduced context size per step