Diagnosing a Long-Session Context Failure for CCAO-F

The capstone diagnosis

This knowledge point is where the whole context-management task statement comes together. The CCAO-F exam tests it at the evaluate level: given a long session where early instructions no longer hold, can you correctly diagnose the cause, choose the right response, and recognise the better upfront strategy? It draws on recognising degradation and choosing between restart and summarise and turns them into a single judgement.

The scenario is specific and common: detailed instructions given at the start of a long session lose force much later in the same session. The evaluative move is to attribute this to context compression -- the instructions were condensed as the window filled -- rather than to Claude deciding to disregard them. That diagnosis is not academic; it determines which response fixes the problem and points toward a prevention strategy that avoids the failure entirely.

Diagnosing correctly

The correct diagnosis rests on a single observation: the instructions were followed at the start and stopped being followed later, in the same session. That timeline rules out the model never understanding them -- it plainly did, early on. What changed between early and late is not the model's grasp of the instruction but the state of the context window, which filled over the session and compressed the earliest content, the instructions included. So the loss of force is compression, not disregard.

Getting this right matters because the two diagnoses point to opposite actions. If you conclude the model is disregarding instructions, you will try to make it comply -- often by repeating the instruction in place, which is the very trap to avoid, because it treats a context-state problem as a compliance problem and does nothing about the filled window. If you correctly diagnose compression, you address the context state instead: restart, summarise, or persist, chosen per the three responses and the restart-versus-summarise judgement. The diagnosis is what makes the fix land on the real cause.

Choosing the response and preventing recurrence

Once compression is diagnosed, the response follows from the earlier analysis. If the thread is still coherent and worth continuing, summarise the state and restart into a fresh window so the early instructions -- captured in the summary -- are back in active context. If the session has drifted beyond recovery, restart clean. If the instructions are the kind that should hold across every session, persist them to Memory or the knowledge base so they survive indefinitely. The diagnosis selects among these; it does not leave you repeating yourself into a full window.

Better still is preventing the failure before it happens. Breaking a large task into smaller segments, with interim saves to the knowledge base at segment boundaries, reduces exposure to compression because no single session runs long enough to fill and compress critically. Planning those segment boundaries in advance is a stronger strategy than extending one session indefinitely and hoping the window holds. The evaluative insight is that the robust practice is architectural -- design the work into segments -- rather than reactive. This is the same segment-and-save discipline that supports persisting durable facts.

What the CCAO-F exam trips candidates on

Two errors are tested. The first is concluding the model changed its mind about instructions rather than recognising context compression as the cause. The credited reading uses the early-followed-then-dropped timeline to identify compression and rejects the "model disregarded it" framing.

The second is responding to the symptom by repeating the same instruction again in place instead of addressing the underlying context state. Repeating the instruction into an already-full window does not fix the compression and the same failure recurs. The credited response addresses context state -- restart, summarise, or persist -- and, better, prevents recurrence by segmenting the task with interim saves.

Common misreadings to avoid

How this shows up on the exam

Questions describe a long session where early instructions fade and ask for the diagnosis and fix, with distractors that blame the model or repeat the instruction in place. Use the timeline -- followed early, dropped late -- to diagnose compression, then prescribe a context-state response and, for full credit, the segment-and-save prevention strategy. The reliable answer never treats it as a compliance problem to be re-pasted away.

This capstone knowledge point draws together the finite context budget, recognising degradation, the three responses, and choosing between restart and summarise.