✨ support auto-compaction

[xyliugo]

Change Type

• ✨ feat
• 🐛 fix
• ♻️ refactor
• 💄 style
• 👷 build
• ⚡️ perf
• 📝 docs
• 🔨 chore

Description of Change

Adds automatic context compaction to the agent execution loop to prevent context-window overflows during long tool-heavy runs. Introduces a new CLI flag --compaction-token and plumbs it through the evaluator into the agents. When enabled, the agent summarizes and rewrites the in-flight conversation to system + initial user + summary once the prompt token count reaches the configured threshold.

Additional Information

• --compaction-token 999999999 disables compaction (default).
• Token counting is done pre-call using LiteLLM token_counter; Claude native path uses Anthropic’s /v1/messages/count_tokens.
• Compaction summaries are included in token_usage; Claude token counting calls are not.
• messages.json now reflects the compacted on-the-fly message state (not the full pre-compaction transcript).

[zjwu0522]

lgtm