Bug

When timeout-compaction runs but does not actually reduce the context (e.g., the compaction model produces output but fails to summarize effectively), the system doesn't escalate to a different approach. It logs [timeout-compaction] compaction did not reduce context and falls through to normal error handling, which means the user gets a "Context overflow" error.

Observed log

[timeout-compaction] compaction did not reduce context for azure/grok-4-1-fast-reasoning; 
falling through to normal handling

Root cause

In run.ts, the timeout-compaction path:

1. Detects timeout + high context usage
2. Triggers compaction with the current model
3. If compaction "succeeds" (no error) but context is still over threshold → falls through
4. No escalation to compact_then_truncate, no model switch, no session rotation

The system treats "compaction ran but didn't help" as unrecoverable, when there are still options:

• Try a different model with better summarization capability
• Escalate to compact_then_truncate (aggressive mode)
• Force session rotation with a handoff summary

Expected behavior

After compaction completes but context is still over threshold:

1. Try the next compaction fallback model (if available — see Compaction fails permanently when Azure content filter blocks summarization — no model fallback #64960)
2. If no more models: escalate to compact_then_truncate mode
3. If that still doesn't reduce enough: force a fresh session with a handoff summary from the last assistant message
4. Log a clear warning: "session force-rotated due to irrecoverable context overflow"

Related

• Compaction fails permanently when Azure content filter blocks summarization — no model fallback #64960 (compaction model fallback chain)
• Context overflow causes infinite loop - no circuit breaker #8596 (basic overflow loop fix)
• Compaction late success leaves /status compactionCount stale after timeout #45492 (compaction counter stale after timeout — metadata issue)