workflow-context-condenser

On a long-running session the 35B model loses coherence as the window fills. The condense_context tool (mcp-observability) applies the OpenHands LLMSummarizingCondenser mechanism: at ≥80% of max_model_len it summarizes the oldest events into one dense digest, always preserving the first 4 messages and the most recent turns. OpenHands measured up to ~2× token-cost reduction at this fill with no quality degradation (arXiv:2511.03690).

When to use it (and when NOT to)

• Use it when a session is genuinely near the limit — long multi-hour work, many tool round-trips, the prompt is dominated by stale early history.
• Do NOT condense routinely or early. Condensing rewrites the prompt prefix, which invalidates the prompt cache — a near-empty context loses more (cache misses) than it gains. Only fire when the window is actually ~80%+ full.

How

condense_context(history) returns {fired, history, tokens_before, tokens_after, ratio, ...}. If fired is false you were under threshold — keep going. If true, continue from the returned (shorter) history; a condenser_fired span records the before/after token counts and ratio so the savings are observable.