Summary

Three targeted improvements to the compression system that address the most impactful remaining quality issues documented in #499.

1. Increase serializer truncation limits

The summarizer LLM was working from heavily truncated input:

A 200-line code edit or a full test output was being cut to 3000 chars before the summarizer ever ran. The summary was working from incomplete data. The budget here is the summary model's context window, not the main model's — there's room.

2. Add "Tools & Patterns" section to compression prompts

Both the first-pass and iterative compression prompts now include:

## Tools & Patterns
[Which tools were used, how they were used effectively, and any
tool-specific discoveries (e.g., preferred flags, working invocations,
successful command patterns)]

After compression, the agent retains tool definitions (from self.tools) but loses conversational context of HOW those tools were used effectively. This section preserves working invocations, preferred flags, and tool-specific discoveries.

3. Degradation warning on repeated compressions

After the 2nd compression, the user sees:

⚠️  Session compressed 3 times — accuracy may degrade. Consider /new to start fresh.

compression_count was already tracked (context_compressor.py:145) but never surfaced. Multiple compressions silently compound quality loss — "summaries of summaries" lose information with each pass.

Changes

24 lines changed, 9 added across agent/context_compressor.py and run_agent.py.

Ref #499