Question 150 A software review finds that a single agent’s attention weakens after reading thousands of lines, even though the model’s context window is large enough. What is the best response? Remove test requirements to make the prompt shorter. Run the same giant prompt ten times and take the shortest answer. Break the review into focused scopes and add an explicit integration check. Simply double the context window again. Explanation Correct Answer C:Attention allocation, not just token capacity, is the observed bottleneck. Incorrect Answers A:Dropping necessary validation creates new quality defects. B:Repetition without structured coverage may reproduce the same omissions. D:A larger window does not guarantee equivalent attention to all details.
Question 151 A dual-review design aims to avoid an erroneous release of a regulated report. Which TWO mechanisms help most? A defined escalation process for unresolved material reviewer disagreements. Treat a successful JSON parse as sufficient factual approval. Trust any majority vote even if the source is absent. Independent source-linked verification of key figures by the second reviewer. Force all reviewers to accept the drafting agent’s assumptions. Explanation Correct Answers A:Risky unresolved discrepancies need adjudication rather than silent publication. D:Separate evidence checks can catch hallucinated or miscopied values. Incorrect Answers B:Schema validity cannot validate a report’s claims. C:Votes cannot substitute for source evidence. E:Anchoring defeats independent review.
Question 152 A repository has an existing function and the agent must replace one carefully scoped section while preserving the remainder. Which built-in action fits? Grep to print the function and stop. Glob to find filenames only. Write the entire file from a guessed reconstruction. Edit with a precise old-to-new replacement. Explanation Correct Answer D:A targeted edit minimizes the chance of clobbering unrelated content. Incorrect Answers A:Content search alone does not modify the function. B:File discovery does not apply the required code change. C:A full overwrite can lose unaffected code and comments.
Question 153 A CI pipeline invokes Claude Code interactively and hangs waiting for a prompt. What invocation style should be used? Delete pipeline timeouts globally. Open an interactive TTY and wait indefinitely for someone to respond. Replace the CI script with a screenshot of a successful local run. Use a noninteractive print mode with explicit input and a bounded, machine-readable result. Explanation Correct Answer D:Headless execution does not depend on a human terminal session. Incorrect Answers A:Unbounded hangs reduce reliability. B:Unattended CI jobs must not require manual prompts to finish. C:A screenshot cannot execute current automated checks.
Question 154 A model explores a codebase, then loses context before applying the approved edit. What should its checkpoint include? Only a note that many files were read. A new global refactor plan unrelated to the identified bug. An assertion that tests passed without having run them. Selected files and symbols, known constraints, planned change and baseline test outcome. Explanation Correct Answer D:A small source-linked plan supports accurate continuation. Incorrect Answers A:File count does not identify actionable code. B:That restarts scope rather than continuing it. C:Unverified status can hide regressions.
Question 155 A run has used all permitted tool calls while the agent still lacks a verified answer. Which result should the client return? An explicit incomplete outcome with remaining uncertainty and recovery information. A permanent success flag because the call budget was fully used. A confident answer assembled from unverified guesses. A silent empty string with no status. Explanation Correct Answer A:Resource exhaustion is a controlled stopping condition, not evidence of task completion. Incorrect Answers B:Budget consumption is unrelated to correctness. C:An unresolved evidence gap must not be hidden from downstream users. D:Callers need to know that the process ended without fulfillment.
Question 156 An assistant copies a sample file path into every generated edit command. What is the likely few-shot trap? The example did not distinguish placeholder paths from real runtime arguments. The model needs examples of more unrelated file formats. The application should disable path validation. All example paths should be valid production targets. Explanation Correct Answer A:Examples should teach parameter binding, not literal reuse of demo artifacts. Incorrect Answers B:The immediate problem is copied literal path data. C:Unsafe literals should be caught, not accepted. D:That would make accidental reuse more dangerous.
Question 157 A project’s MCP tool is registered but fails for all teammates after one developer commits a personal API key in .mcp.json. What should be repaired? Grant the compromised key unrestricted production permissions. Disable authentication permanently for the server. Keep shareable server definitions separate from per-user secret credentials. Publish the private key in a documentation page to simplify setup. Explanation Correct Answer C:Project configuration should avoid distributing one person’s private token. Incorrect Answers A:More privilege makes a leaked credential more dangerous. B:Removing access controls does not solve safe credential distribution. D:That expands the secret exposure.
Question 158 A code agent stores a checkpoint before handing a bug fix to another session. Which entry is most useful? A claim that all tests passed when one failed. Only the phrase “almost finished”. The changed commit or file paths, verified tests, remaining failing case and versioned repository state. A summary of unrelated programming conventions. Explanation Correct Answer C:Specific artifacts and test results support continuation without restarting. Incorrect Answers A:False completion state can conceal a defect. B:Progress language alone is not actionable. D:That omits the current code and known blocker.
Question 159 A coordinator issues a delegation with a vague request “Investigate the issue.” The worker returns irrelevant information. What should the invocation contain? The entire customer database as background context. An instruction to invent missing logs if none are found. A concrete objective, scope, relevant prior evidence, and desired output schema. Only a worker name with no task context. Explanation Correct Answer C:Clear handoff context lets the specialist target the unresolved question. Incorrect Answers A:Unnecessary sensitive data increases risk and distracts from the task. B:Unavailability of evidence must be reported, not fabricated. D:Names do not supply the problem statement.
Question 160 A project CLAUDE.md grows to 8,000 lines of detailed API examples, increasing context load. What should its owner do? Keep concise universal conventions in CLAUDE.md and move specialized references into scoped rules or skills. Rely on the model to infer the team’s style from filenames. Add every historical incident transcript to the top of CLAUDE.md. Delete all project-specific instructions. Explanation Correct Answer A:Modular guidance reduces irrelevant context while preserving discoverability. Incorrect Answers B:Implicit inference is not a dependable substitute for explicit policy. C:More unrelated text increases attention overhead. D:Critical build and safety conventions would be lost.
