The Context Pollution Debugger

My agent's behaviour degrades over long runs. Help me diagnose it.

Run characteristics: [LENGTH, WHEN DEGRADATION STARTS]
Symptoms: [DESCRIBE — repeats itself, forgets constraints, loses
the goal, contradicts earlier work, gets slower]
Transcript excerpt: [PASTE FROM THE DEGRADED PORTION]

Diagnose:
1. Is this context exhaustion, context pollution, or something else?
2. What is consuming context that should not be?
3. Is anything in the context actively misleading — stale results,
   superseded decisions, failed attempts presented as facts?
4. Has the goal drifted, and where did it start drifting?
5. Are errors accumulating — does an early mistake compound?
6. Is the agent re-reading things it already knows?

Then recommend:
7. The single change most likely to fix it
8. What to measure to confirm the fix
9. Whether this needs a context strategy change or a prompt change

Be specific about which content in the transcript is the problem.