Question 130
A coordinator times out a worker but starts a replacement while the original can still commit writes. What risk must be controlled?
- Both attempts may perform the same consequential action.
- The replacement cannot read any data.
- The replacement is guaranteed to cancel the original merely by starting.
- The coordinator can infer which attempt wrote from response arrival order alone.
Explanation
Correct Answer
A:Timeout does not guarantee termination. Cancellation, fencing tokens, or idempotent execution can prevent an obsolete worker from committing after replacement.
Incorrect Answers
B:Read access depends on configuration, not timeout alone.
C:Launching another worker does not automatically revoke the original worker's ability to act.
D:A late response can describe an earlier write, so arrival order cannot reliably identify the acting attempt.
Question 131
A distributed job lease moves from worker A to worker B, but A resumes after a pause. What should the storage service check before accepting writes?
- Only the timestamp supplied by A in its own write request.
- A current fencing or generation token associated with the active lease.
- Only the worker's original start time.
- Only whether A's process is still alive.
Explanation
Correct Answer
B:The service can reject writes from an obsolete executor even if that process resumes. Ownership must be enforced at the commit boundary.
Incorrect Answers
A:A stale worker can submit a recent timestamp; the storage service needs proof of current lease generation.
C:An earlier start does not preserve authority after ownership transfers.
D:Liveness does not prove current ownership of the lease.
Question 132
A worker catches every exception and returns an empty success object. Why is this harmful to orchestration?
- An empty object is invalid in every structured result format.
- The coordinator must retry every legitimately empty result indefinitely.
- The coordinator cannot distinguish successful emptiness from failed execution.
- Exception handling cannot be used in a worker that returns structured data.
Explanation
Correct Answer
C:Error information guides retry, escalation, and completeness decisions. Suppressing it can cause the workflow to advance on missing evidence.
Incorrect Answers
A:An empty object can be valid data under a suitable schema. The problem is that the same success-shaped value masks execution failures.
B:Infinite retry is not a necessary consequence or a sound remedy. The interface should first distinguish successful emptiness from a failure that may warrant recovery.
D:A structured result can carry an explicit failure status and actionable error category. Catching exceptions is compatible with clear reporting when their meaning is preserved.
Question 133
A worker produces a large dataset that another worker must analyze. What handoff best preserves fidelity without filling the coordinator's context?
- Pass a file path that exists only inside the producer's disposable container, without arranging access for the consumer.
- Return only the first hundred rows and describe the subset as the complete dataset.
- Store the artifact with a stable reference and pass its schema, scope, and access location.
- Paste a lossy prose summary and discard the dataset.
Explanation
Correct Answer
C:The next worker can load exact data through the reference. The coordinator retains compact metadata and can verify that the correct artifact is being used.
Incorrect Answers
A:A path is useful only if the next worker can resolve it. A producer-local location may disappear or be inaccessible at the handoff.
B:A sample does not preserve the exact input required for full analysis. The consumer could produce incomplete results without knowing that rows were omitted.
D:The next worker may need exact rows and values that the summary omits.
Question 134
A coordinator decomposes a migration into code changes and tests but omits data conversion. What planning check would catch this omission?
- Assign more workers to the existing code tasks.
- Map every user requirement and dependency to an owned subtask or explicit exclusion.
- Count how many subtasks were created.
- Check that every planned worker has received a prompt.
Explanation
Correct Answer
B:Coverage mapping exposes requirements that no worker is responsible for. It also makes intentional exclusions reviewable rather than accidental.
Incorrect Answers
A:Additional effort on covered work does not address the missing conversion requirement.
C:A large task count can still miss a required outcome.
D:Dispatch coverage shows that existing subtasks were assigned, not that the decomposition includes every required migration outcome.
Question 135
A coordinator changes a worker's expected output schema while the worker is running. What should it do?
- Version the contract and coordinate a restart, migration, or compatible adapter.
- Ignore the change when integrating results.
- Parse the old result as the new schema without checking.
- Accept either schema without recording which version produced the result.
Explanation
Correct Answer
A:The receiving side must know which schema produced the result. Explicit version handling prevents silent reinterpretation of old output.
Incorrect Answers
B:This leaves the coordinator and worker using incompatible contracts.
C:A schema change can alter field meanings or required values.
D:Unversioned acceptance leaves downstream consumers unable to interpret the fields reliably.
Question 136
A production failure cannot be reproduced because prompts, tool versions, and model identifiers were not recorded. What metadata should future runs retain?
- Only the final recommendation and an aggregate success flag.
- Only tool names, without implementation versions or relevant request parameters.
- Every unrelated user conversation.
- The relevant configuration versions and execution inputs, with sensitive data appropriately protected.
Explanation
Correct Answer
D:Reproduction requires knowing the conditions that shaped behavior. Versioned prompts and tool contracts help distinguish code changes from model or configuration changes.
Incorrect Answers
A:The outcome alone omits the inputs and configuration needed to reproduce the sequence that failed.
B:A named tool can change behavior, and the same operation can behave differently on different inputs.
C:Unrelated history adds privacy exposure without explaining this run.
Question 137
An investigation uses a tool session whose temporary database exists only on worker A. A load balancer sends the next dependent step to worker B. What must the orchestrator address?
- Assume all workers share temporary tool sessions because they use the same model.
- Drop the query's database requirement and claim the investigation finished.
- Keep dependent steps on the state-owning session or externalize and explicitly transfer the needed state.
- Retry the dependent query on any healthy worker without transferring the temporary database.
Explanation
Correct Answer
C:A stateful dependency cannot be recovered merely by dispatching the same command to another worker.
Incorrect Answers
A:Model identity does not imply shared tool-process memory or temporary databases.
B:Removing a required dependency would leave the analytical task unperformed.
D:Worker health does not supply the session-local state that the query requires.
Question 138
A streaming response contains only part of a tool's JSON arguments. When should the application execute the consequential tool?
- After the complete invocation has been assembled, parsed, and validated.
- As soon as the tool name appears.
- After guessing the missing fields from prior requests.
- Whenever the first argument resembles a valid ID.
Explanation
Correct Answer
A:Streaming fragments are not a validated command. The executor needs the finalized structured input before acting.
Incorrect Answers
B:The arguments may still be incomplete or change as the stream continues.
C:Guesses can target the wrong resource or amount.
D:One valid field does not establish a complete or authorized operation.
Question 139
A workflow should never publish a report without source verification. What production signal best detects violations of this invariant?
- A publish worker's uncorrelated statement that verification probably completed.
- Whether a report contains a references section, without checking the verification record.
- A check linking each publish event to the required successful verification record.
- The existence of a source-verification function somewhere in the codebase.
Explanation
Correct Answer
C:An invariant concerns the relationship between events and state. Correlated monitoring can identify a publication that bypassed its prerequisite.
Incorrect Answers
A:An unsupported assertion cannot establish the required link between the published artifact and its successful gate.
B:Citation formatting can be present even when the required source-validation step never ran.
D:An implemented checker does not prove that a particular published report passed it.
Question 140
An assistant can read a public catalog and submit a costly purchase. How should confirmation requirements differ?
- Let the model choose confirmation rules ad hoc without policy.
- Never confirm purchases if a catalog lookup succeeded.
- Require identical confirmation for every internal reasoning step.
- Apply explicit authorization rules based on the action's consequence and existing user approval.
Explanation
Correct Answer
D:The workflow should distinguish information gathering from commitments. It should also reuse valid prior authorization rather than asking redundantly.
Incorrect Answers
A:Inconsistent rules can miss high-impact actions or burden harmless ones.
B:Successful reading does not authorize a financial commitment.
C:Excessive confirmation adds friction without distinguishing consequential actions.