Context Management Exam Questions: Domain 5 Deep Dive

Domain 5 of the CCAR-F exam, Context Management and Reliability, carries 15% of your total score. That translates to roughly nine of the 60 items on any given sitting. Context management exam questions are not abstract: every item presents a scenario, a symptom, and a set of plausible fixes, and you must pick the one that addresses root cause rather than surface noise. This guide works through the concepts, the decision logic, and the trap answers the exam consistently uses.

What does Domain 5 actually test?

Domain 5 tests your ability to diagnose and fix context-related failures in production Claude deployments. The domain covers five broad skill clusters: recognising when a context window is degrading, choosing between session-continuation and session-reset strategies, injecting summaries correctly, managing stale context in long-running agents, and applying the right reliability pattern when the window is the constraint rather than the model.

Per Anthropic's published exam guide, Domain 5 accounts for 15% of the CCAR-F exam weight. Because each sitting draws four scenarios at random from a bank of six, you will almost certainly see at least one scenario that is primarily a context-management problem, even if it is framed as an agentic or orchestration question.

Why does context management matter more than candidates expect?

Most candidates over-invest in Domain 1 (Agentic Architecture, 27%) and under-invest in Domain 5. The mistake is treating context management as a narrow "token counting" topic. In practice, context failures are the most common silent failure mode in production agents: the model does not error, it just answers with stale or diluted information.

The attention dilution problem is the conceptual anchor for this domain. Once you understand that adding tokens to a window does not add proportional value, the exam's preferred answers become predictable: trim aggressively, summarise early, and isolate subagent contexts rather than sharing one giant window.

What are the core concepts behind context management exam questions?

The table below maps the six most-tested concepts in Domain 5 to their exam-relevant decision rule and the most common wrong answer.

Each of these maps to a concept in our Context Management and Reliability library, which covers 174 atomic concepts across all five CCAR-F domains.

How does the exam frame context window scenarios?

The exam never asks "what is the context window size?" It asks what you should do when a specific symptom appears. The four most common symptom patterns are:

1. Attribution loss: the model cites the wrong source or conflates two documents. Root cause is almost always attention dilution or a missing structured-context-passing pattern.
2. Contradictory outputs across turns: the model gives a different answer to the same question ten turns later. Root cause is stale context, not model inconsistency.
3. Silent truncation: early instructions disappear from the model's effective attention even though they are technically within the window. Root cause is the lost-in-the-middle effect.
4. Runaway token spend: an agentic loop accumulates tool results without pruning. Root cause is missing tool-result trimming or a missing compaction step.

For each symptom, the exam expects you to trace to root cause before selecting a fix. A proportionate fix for attribution loss is structured context passing, not a larger model. A proportionate fix for stale context is a session reset with summary injection for fresh sessions, not a retry.

What is the stale context problem and how does the exam test it?

Stale context occurs when facts established early in a session are contradicted by later tool results, but the model continues to reason from the earlier facts because they occupy a prominent position in the window. The exam tests this with scenarios where an agent has been running for many turns and begins producing outputs that are internally consistent but factually wrong relative to the current state of the world.

The correct diagnostic sequence is:

1. Identify the turn at which the contradiction was introduced.
2. Determine whether the stale fact is in the system prompt, the conversation history, or an injected tool result.
3. Choose between forking the session (if the stale fact is in the history) or patching the system prompt (if the stale fact is in a static instruction).

The exam consistently rewards root-cause tracing over symptomatic fixes. If a scenario describes an agent that is citing an outdated database schema, the correct answer is to reset the session with a fresh schema injection, not to add a "please use the latest schema" reminder to the next user turn.

How does session management strategy appear in exam questions?

Session management options is one of the highest-yield concepts in Domain 5. The exam presents three strategies: resume, fork, and fresh start. Each has a specific use case.

The when to resume vs fork vs fresh start decision tree is worth memorising as a flowchart. The exam will give you a scenario and three answer options that each correspond to one of these strategies; the correct answer is always the one that matches the signal in the scenario, not the one that sounds most cautious.

What reliability patterns does Domain 5 expect you to apply?

Reliability in Domain 5 is not about uptime; it is about output consistency under context pressure. The exam tests three reliability patterns specifically:

Independent review instances: when a task requires reviewing a long document, spawn an independent Claude instance with only the document and the review criteria, rather than passing the full conversation history. This prevents prior conversation bias from contaminating the review.

Structured handoff: when handing off between agents or between a Claude session and a human reviewer, the handoff payload must be a structured object (JSON or XML), not a prose summary. Prose summaries lose precision; structured objects preserve field-level facts.

Compaction before delegation: before spawning a subagent, compact the context to the minimum necessary for that subagent's task. The subagent context isolation pattern is the exam's preferred answer whenever a scenario describes a coordinator passing "the full conversation" to a subagent.

How should you approach multiple-response items in Domain 5?

Several Domain 5 items are multiple-response: the stem tells you to select two or three correct answers. The trap is selecting answers that are individually plausible but redundant. The exam rewards answers that address distinct failure modes.

A worked example: a scenario describes an agent that is producing inconsistent outputs after 40 turns. The stem asks you to select two actions that together address the root cause. The options are:

• A. Increase the model's temperature to reduce determinism.
• B. Reset the session and inject a structured summary of the prior 40 turns.
• C. Add a "be consistent" instruction to the system prompt.
• D. Trim tool results to remove redundant intermediate outputs before the next session.

The correct pair is B and D. B addresses the stale context by resetting with a summary. D addresses the token accumulation that caused the degradation. A is wrong because temperature affects output variance, not context staleness. C is wrong because a prompt instruction cannot fix a structural context problem.

This pattern, two answers that address distinct layers of the same root cause, appears repeatedly in Domain 5 multiple-response items.

How do Domain 5 concepts connect to other exam domains?

Context management does not exist in isolation. The exam frequently embeds Domain 5 concepts inside Domain 1 (Agentic Architecture) scenarios. The coordinator responsibilities concept, for example, includes managing the context budget across subagents, which is a Domain 5 skill tested under a Domain 1 heading.

Similarly, Domain 4 (Prompt Engineering and Structured Output, 20%) intersects with Domain 5 when the question is about how to structure a summary injection. A well-formed summary injection uses the same principles as a well-formed system prompt: explicit field labels, ordered by importance, with the most critical facts at the top to counter the lost-in-the-middle effect.

The five-domain weight breakdown is worth keeping in mind as you allocate study time:

At 15%, Domain 5 is the smallest domain by weight, but its concepts bleed into every other domain. A candidate who cannot diagnose context degradation will lose points in Domain 1 scenarios, not just Domain 5 ones.

What does a well-structured summary injection look like?

The exam does not ask you to write code, but it does ask you to recognise correct versus incorrect patterns. A correct summary injection for a fresh session looks like this:

{
  "session_summary": {
    "task_objective": "Migrate the payments service from PostgreSQL 13 to 15.",
    "completed_steps": [
      "Schema audit complete: 3 deprecated columns identified.",
      "Test environment provisioned on 2026-06-01."
    ],
    "open_decisions": [
      "Whether to run parallel writes during cutover."
    ],
    "constraints": [
      "Zero downtime required.",
      "Rollback window: 4 hours post-cutover."
    ],
    "last_known_state": "Awaiting sign-off on parallel-write proposal."
  }
}

An incorrect pattern is a prose paragraph: "We have been working on migrating the payments service. So far we have audited the schema and found some issues. We need to decide about parallel writes." The prose version loses the structural distinction between completed steps, open decisions, and constraints, which means the model cannot reliably retrieve individual facts under attention pressure.

The exam will present both patterns and ask which one is more reliable for a fresh session. The structured JSON is always correct because it preserves field-level precision and positions critical facts where attention is strongest.

How should you prepare for context management exam questions specifically?

Three preparation actions have the highest return for Domain 5:

1. Work through the context management concept library systematically. Each atomic concept maps to a task statement, and the task statements map directly to item stems. If you can state the decision rule for each concept in one sentence, you can answer the exam item in under 90 seconds.

2. Practice distinguishing root cause from symptom. For every practice scenario, write down the symptom, the root cause, and the proportionate fix before looking at the answer options. Candidates who jump to the options first are more likely to select a plausible-sounding symptomatic fix.

3. Review the session management decision tree until it is automatic. The resume/fork/fresh-start decision appears in multiple scenarios, sometimes disguised as an orchestration question. If you can apply the decision tree in under 30 seconds, you will not lose time on these items.

Our platform's adaptive engine uses Bayesian Knowledge Tracing with a 0.90 mastery threshold, so it will keep routing you back to Domain 5 concepts until you have demonstrated consistent accuracy, not just a single correct answer. Practice exams mirror the real format: 60 questions, scored 100 to 1000, with 720 as the passing bar.