feat: Add contextStrategy to control conversation growth and prevent token explosion in agent runs

[ibrahimkzmv]

What

Introduces a contextStrategy option to manage conversation history during multi-turn agent runs. Supports sliding-window truncation, LLM-based summarization, and custom compression to prevent unbounded context growth.

Why

Currently, agent conversations grow without limit, causing:

• Context window overflows (especially for smaller/local models)
• Rapidly increasing token costs due to full history being resent each turn
• Degraded model performance from excessively long context

This change ensures context size is actively managed during execution, making long-running agent workflows production-ready and cost-efficient.

Closes #59

Checklist

• npm run lint passes
• npm test passes
• Added/updated tests for changed behavior
• No new runtime dependencies (or justified in the PR description)

[JackChen-me]

Three issues found, two are blockers:

1. summarize produces orphaned tool_result_id after the 2nd compaction. At runner.ts:277, splitAt = Math.max(2, Math.floor(rest.length / 2)) can cut through a tool_use/tool_result pair once rest is odd-length (happens after the first summary injects its own user turn). Anthropic hard-rejects the next call with tool_result block missing corresponding tool_use block. Fix: align splitAt to an even offset, e.g. Math.max(2, Math.floor(rest.length / 4) * 2).

2. Summary adapter.chat() usage isn't added to totalUsage. At runner.ts:311, the summary response's usage is emitted in a trace event but never fed back into the accumulator, so maxTokenBudget is silently bypassed and RunResult.tokenUsage under-reports cost. Fix: return { messages, usage } from summarizeMessages and accumulate at the call site.

3. Synthetic marker/summary messages use role: 'user', producing sequences like [user, user, assistant, user]. Not fatal on Anthropic direct API (silent merge per anthropic-sdk-typescript#565), but breaks Bedrock and blurs prompt semantics on direct API (summary text concatenates onto the original prompt). Fix: tag as role: 'assistant' or fold into the next real user message.

Repros on request.

[ibrahimkzmv]

Appreciate the review, you caught a subtle edge case with the odd-length rest

Fixed a few things:

• safe boundary split for tool_result alignment
• usage tracking now includes summarizeMessages cost
• removed synthetic user rows, moved notes into next turn

Thanks again for the sharp eye @JackChen-me