Bug Description

summarizeChunks() in src/agents/compaction.ts uses retryAsync() with 3 attempts per chunk to handle transient errors during compaction. However, when the summarizer model is persistently unavailable (overloaded, rate-limited, or misconfigured), the retry loop burns tokens across multiple chunks without making progress.

In a multi-chunk compaction scenario (common with long channel conversations), this means:

1. Chunk 1: 3 retries → fail
2. Chunk 2: 3 retries → fail
3. Chunk N: 3 retries → fail

Each retry sends the chunk to the model, consuming input tokens even though the response fails. With large chunks (~4K tokens each), a 5-chunk compaction can waste 60K+ tokens before finally throwing.

Expected Behavior

After N consecutive compaction failures, the system should stop attempting compaction and fall back to the previous summary (or a placeholder). This prevents cascading token waste when the summarizer is down.

A circuit breaker pattern would be appropriate:

• Closed (normal): compaction attempts proceed
• Open (after N failures): compaction is skipped, fallback used
• Half-open (after cooldown): one test attempt allowed

Affected Code

• src/agents/compaction.ts:263 — retryAsync() call in summarizeChunks()
• Called from summarizeWithFallback() and summarizeInStages()