Step 1 of 3
Audit context stability
Find every byte that changes between turns - each one is a cache killer.
Audit my agent's context assembly for cache hostility. Capture the exact assembled context for three consecutive turns and diff them. Classify every difference: EXPECTED GROWTH (new turns appended at the end - fine), PREFIX MUTATIONS (anything changing before the append point - cache killers), and VOLATILE INJECTIONS (timestamps, counters, randomized ordering, per-turn IDs inside stable sections). Common culprits from harness analyses: current-time in the system prompt, re-rendered plan blocks at the top, tool definitions re-serialized in unstable order. Produce the diff report with each mutation's source located in code.
Expected after this step
A three-turn diff report locating every prefix mutation.
Should not happen
- ✕A timestamp in the system prompt silently killing the cache every turn
- ✕Re-sorting or rewriting history each call so no prefix ever matches
- ✕Injecting the dynamic plan reminder at the top instead of the tail
- ✕Claiming cache wins without reading the usage fields that prove them
Verify before continuing
Do not move on until every check is true. The complete button stays locked until then.
Do not continue if…
- !A timestamp in the system prompt silently killing the cache every turn
- !Re-sorting or rewriting history each call so no prefix ever matches
- !Injecting the dynamic plan reminder at the top instead of the tail
- !Claiming cache wins without reading the usage fields that prove them
If the AI messes this up
Use this when the AI fakes progress or breaks the feature. It forces a real fix.
The diff is empty but bills are still high. Verify you are diffing the exact serialized request payload (not your in-memory structures) - serialization order and whitespace count too.