Question 220
A team accidentally committed an API token inside a project MCP configuration file. What change should they make?
- Reference credentials through environment or a secret manager, then rotate the leaked token
- Re-establish or reload the affected server and verify its active configuration
- Review tool exposure, permission filters and server-side authorization
- Select supported authenticated transport and credential renewal rather than embedding a fixed secret
Explanation
Correct Answer
A:Configuration may be shared and reviewed; raw secrets do not belong in it. External secret injection reduces exposure, while rotation addresses the already-compromised credential.
Incorrect Answers
B:Running sessions may retain existing transport state; this contributes to applying modified MCP server settings without meeting the need for removing leaked secrets from MCP settings.
C:Connectivity alone does not grant every capability; that serves diagnosing a connected but unavailable MCP tool, whereas the case requires removing leaked secrets from MCP settings.
D:Because remote MCP services need a secure authentication lifecycle, the measure supports maintaining authenticated MCP transport rather than the needed removing leaked secrets from MCP settings.
Question 221
An international research summary must retain source spelling of names and clearly distinguish source-backed conclusions from unknowns. Which TWO prompt requirements achieve these goals? Choose TWO.
- Separate cited research findings from openly labeled uncertainty or missing evidence
- Furnish a clear output structure with explicit labels for evidence and risk rationale
- Retain the sources' spelling of international names, normalizing only approved output fields
- Link decisions and proposals to named sources, preserving unresolved disagreement
Explanation
Correct Answers
A:This choice addresses separating sourced conclusions from unknowns: requiring evidence mapping and explicit uncertainty discourages treating unverified inference as fact; the reader can then distinguish grounded conclusions from gaps.
C:Field-level instructions distinguish transcription from interpretation. Unnecessary translation can break matching against source systems that store original forms.
Incorrect Answers
B:Combining an evaluation rubric with an output contract makes responses comparable; this may support standardizing the structure of risk assessments but cannot alone meet the requirement for preserving exact source spelling where required.
D:Attribution is part of factual correctness; a method suited to preserving attribution in generated meeting summaries does not meet the need for preserving exact source spelling where required.
Question 222
A coordinator creates tiny subagents for every sentence of a short report. What design correction reduces overhead?
- Group closely related low-cost tasks into a meaningful worker assignment
- Introduce concurrency limits and a work queue to the task scheduler
- Keep agents for uncertain reasoning and use tools for well-defined operations
- Produce parallel prerequisite branches with a join gate before completion
- Distinguish read-only verification from the authorized state-changing step
Explanation
Correct Answer
A:Excessive decomposition multiplies setup and context cost. Workers should own coherent deliverables rather than fragments requiring more coordination than reasoning.
Incorrect Answers
B:Capacity-aware scheduling controls rate limits and tool contention; that relates to scheduling within available execution capacity, not the necessary avoiding overfragmentation of delegated work.
C:Although deterministic database queries are better expressed as typed tool actions, the choice addresses choosing tools rather than unnecessary agents and not the requested avoiding overfragmentation of delegated work.
D:A fan-out and join pattern preserves concurrency while enforcing that no required verification branch is skipped; the deciding condition is avoiding overfragmentation of delegated work, whereas this action concerns joining mandatory parallel checks.
E:Distinct stages expose eligibility decisions and permit human approval before mutations; the option would serve separating validation from side-effecting tasks, not avoiding overfragmentation of delegated work as specified in the stem.
Question 223
A planner repeatedly creates duplicate tasks because workers return unconstrained prose. Which output requirement helps?
- Call for evidence of which documents and versions the worker actually consulted
- Supply condensed relevant findings and source handles rather than full unrelated transcripts
- Require structured task results with IDs, status and completion evidence
- Standardize worker field names but omit task IDs, status values, and duplicate-detection criteria
Explanation
Correct Answer
C:Stable identifiers and machine-readable outcomes let the parent distinguish completed assignments from new work. Free-form narratives make reconciliation fragile.
Incorrect Answers
A:Source and version metadata distinguish inspected facts from memory-based guesses; the unsatisfied requirement is normalizing delegated task results, not validating the evidentiary basis of a subagent answer.
B:Selecting task-relevant evidence controls token growth without erasing provenance; in this scenario, controlling delegated context size is secondary to normalizing delegated task results.
D:Names alone cannot establish whether two worker messages describe the same planned task. A machine-readable task result needs stable IDs and completion status to support deduplication.
Question 224
A coordinator suppresses a citation-validator failure because the report writer succeeded. Why is that unsafe?
- Log a partial workflow outcome and identify the unavailable branch explicitly
- Fail the publication gate whenever mandatory validation fails
- Confirm the worker's claimed completion against authoritative tool status
- Show completed, failed and pending work as separate outcome categories
Explanation
Correct Answer
B:Successful generation is distinct from successful verification. Suppressing the validator's outcome lets unsupported claims reach the reader.
Incorrect Answers
A:Partial success is not whole-task success; this concerns preserving partial-failure status across worker aggregation, leaving the requirement for honoring validation failures during workflow completion outstanding.
C:Subagent prose should not override an explicit failed execution record; the action supports verifying delegated completion against real error state but falls short of honoring validation failures during workflow completion in this case.
D:Progress counters must reflect task outcomes rather than merely finished attempts; this can serve distinguishing task completion from task success, but leaves the requirement for honoring validation failures during workflow completion unmet.
Question 225
A vendor integration needs the model to call a real lookup tool rather than merely output data resembling a tool call. What should the design enforce?
- Require an actual authorized tool invocation and process its result through the dispatcher
- Flag nonstandard stop or refusal outcomes before treating content as a schema result
- Constrain the schema to supported constructs and validate remaining rules in application code
- Require schema-constrained JSON output through the supported structured-output interface
Explanation
Correct Answer
A:A typed JSON response is not a network operation. The host must execute the selected tool and distinguish performed actions from generated descriptions.
Incorrect Answers
B:Because schema-oriented APIs may have special refusal or termination paths, the measure supports handling non-result terminations in structured workflows rather than the needed distinguishing actual tool execution from JSON text.
C:Constrained decoding has a documented supported JSON Schema subset; its benefits for respecting structured-output schema limitations do not resolve the need for distinguishing actual tool execution from JSON text.
D:Constrained JSON generation can enforce a declared schema for supported responses, avoiding brittle scraping of prose or fences; supporting producing machine-parseable schema-conformant JSON cannot replace the required distinguishing actual tool execution from JSON text.
Question 226
A user asks to undo an earlier analysis in a retained session but preserve the audit record. What strategy fits?
- Require renewed authentication or authorization before sensitive operations resume
- Create a new fork from the last accepted checkpoint for revised analysis
- Queue state-changing updates or use optimistic concurrency on the session
- Retain message order and essential tool-result linkage along with session metadata
- Initiate a separate session with its own customer and authorization context
Explanation
Correct Answer
B:Branching preserves what was previously considered while enabling a corrected trajectory. Deleting historical messages would erase provenance needed for review.
Incorrect Answers
A:A stored conversation is not a perpetual permission grant; the stated problem requires revising an analysis without erasing history rather than revalidating authorization on session resume.
C:For preventing concurrent session-state conflicts, concurrent writes can produce lost updates and inconsistent tool state; however, this requirement concerns revising an analysis without erasing history.
D:Resumable state includes structured execution history, not merely the final natural-language summary; that may serve checkpointing structured agent progress, but it is no substitute for revising an analysis without erasing history.
E:Separate sessions prevent accidental disclosure and policy contamination between customers; this may support isolating unrelated customer conversations but cannot alone meet the requirement for revising an analysis without erasing history.
Question 227
A finance agent should not send emails directly, although it can draft response text. Which tool decision meets that constraint?
- Differentiate or rename tools to identify their owning integration and purpose
- Control publish authority on validated workflow state instead of mere tool availability
- Make explicit each tool's selection criteria and record type in descriptions
- Disallow the unauthorized call and report a permission denial to orchestration
- Remove the email-sending tool while keeping read and draft capabilities
Explanation
Correct Answer
E:Tool distribution should distinguish generating a proposed communication from transmitting it. Without the send operation, a prompt cannot accidentally deliver an unapproved message.
Incorrect Answers
A:Although unique, descriptive identifiers help the agent and dispatcher target the intended backend, the choice addresses resolving tool name collisions and not the requested separating drafting from outbound delivery permissions.
B:Tool existence is not business authorization for every stage; the deciding condition is separating drafting from outbound delivery permissions, whereas this action concerns making a sensitive tool contingent on workflow state.
C:Tool routing quality depends on distinct metadata; the option would serve disambiguating adjacent retrieval tools, not separating drafting from outbound delivery permissions as specified in the stem.
D:A capability boundary must be enforced at execution time; this addresses enforcing a denied capability at runtime, while the question is about separating drafting from outbound delivery permissions.
Question 228
A configuration file is thousands of lines long. Claude Code already knows its path and the specific region containing a suspicious setting. Which read-only built-in-tool approach best limits unnecessary context?
- Run Grep across the entire repository despite knowing the file and the relevant lines.
- Apply Edit to the suspected value before obtaining approval to modify the configuration.
- Use Read on the known file with a bounded line offset and limit around the suspect setting.
- Rewrite the configuration through Write and inspect the generated output afterward.
Explanation
Correct Answer
C:Read can retrieve a targeted portion of a large local file while preserving its contents. Supplying a known offset and limit avoids filling the context with thousands of unrelated lines.
Incorrect Answers
A:Grep is helpful when the location of text is unknown. Here the path and the target region are already established, so searching broadly can add irrelevant output instead of inspecting the desired lines.
B:Edit is a mutation tool, not an inspection mechanism. Altering configuration before reviewing the relevant content violates the read-only objective and could introduce an unverified change.
D:Write would replace file contents rather than perform an inspection. That introduces an unnecessary and potentially destructive side effect, whereas a bounded Read serves the stated goal.
Question 229
A compliance assistant produces findings but never tests how many false positives its rubric creates. What evaluation asset is needed?
- Build a labeled test set including true positives, negatives and borderline cases
- Instruct the reviewer to recommend changes and require authorized execution afterwards
- Check claims against sources before a separate editorial clarity pass
- Cross-check each candidate value to document evidence and escalate unresolved disagreement
Explanation
Correct Answer
A:Calibration requires examples with known outcomes, not only successful demonstrations. Precision and recall measurements can expose overzealous review behavior.
Incorrect Answers
B:Reviewer and implementer responsibilities differ; this can serve separating review conclusions from edit authority, but leaves the requirement for measuring review precision against ground truth unmet.
C:Fact checking and presentation optimization use different evidence and criteria; that addresses separating factual checking from writing polish, whereas this case calls for measuring review precision against ground truth.
D:Model disagreement does not supply a factual tie-breaker; this concerns resolving extraction disagreements with evidence, leaving the requirement for measuring review precision against ground truth outstanding.
Question 230
An unsafe call must be blocked even if Claude confidently explains why it is acceptable. What safety architecture meets that requirement?
- Introduce a PreToolUse hook that denies matching unsafe invocations
- Enforce the boundary in executable hook or application policy code
- Limit the hook matcher to the intended tool name or invocation pattern
- Examine result payload semantics in a post-tool validation hook
Explanation
Correct Answer
B:Model reasoning can propose an operation but should not override deterministic authorization. Executable policy separates permissions from probabilistic language generation.
Incorrect Answers
A:PreToolUse runs before a tool call and can participate in the decision to block it; its benefits for blocking disallowed commands before tool execution do not resolve the need for keeping permission enforcement independent of model judgment.
C:Matchers determine which hook executions apply; this contributes to targeting a hook to relevant operations without meeting the need for keeping permission enforcement independent of model judgment.
D:Transport success does not necessarily mean business success; that serves detecting business failure despite transport success, whereas the case requires keeping permission enforcement independent of model judgment.