Answer: B. The agentic loop contract is managed by stop_reason : "tool_use" → execute and continue; "end_turn" → the model is finished. Parse helper text to detect completion and use an iteration limit as a stopping mechanism main are the two canonical anti-patterns (the limit remains useful as a secondary safeguard). A and C refine the anti-pattern instead of correcting it; D breaks the loop.

Answer: C. Logs show subagents who executed correctly what was asked of them — the problem is upstream: decomposition. “Health” has been reduced to just diagnosis. B is a desirable safety net but not the cause; Contradicted the newspapers; D is not supported by the situation. Examination reflex: partial coverage + correct subagents = suspect the contractor .

Answer: B. This is the most tested invariant of Domain 1: context isolation of subagents. “Based on the conclusions of phase 1” conveys nothing — the subagent has never seen these conclusions. The correction: include the complete outputs (or a structured summary) of the previous phases in the prompt, ideally with a structured format separating data and instructions. Invents a bug; C and D do not correspond to the symptom (generic report = lack of material, no tool or truncation).

Answer: C. The key word in the question is “ guarantee ". The prompt (A, B) and model choice (D) improve the probability compliance — they can reduce 7% to 1%, never to 0%. A critical business rule (financial operations) requires enforcement determinist : the software blocks the call, regardless of what the model “decides”. This is the structuring distinction of the entire examination: deterministic guarantees (hooks, preconditions) vs probabilistic conformity (prompts).

Answer: C. Tool descriptions are main mechanism model selection; here they are minimal and almost identical — this is the root cause. Enriching it is the correction with the least effort and the greatest impact. D can help Next (few-shots cost tokens and do not correct empty descriptions); A is over-engineering that adds a component to maintain; B degrades the design (catch-all tool with unclear contract).

Answer: A. The transient / business / validation taxonomy (+ permission) with isRetryable explicit is what allows the agent to decide intelligently : retry timeout, never retry a policy violation (it will always fail — need to notify or escalate), correct input then retry validation. B leaves the interpretation to chance; C causes the agent to loop on non-retryable errors — exactly the symptom observed; D hides the failure, worst anti-pattern: an empty “valid” result and a failure must be distinguishable .

Answer: B. Two rules combine: scope project = .mcp.json versioned (available to clone for the whole team, unlike ~/.claude.json which is individual and not shared); secrets = substitution of environment variables, never a clear token in a versioned file. A does not satisfy “available to clone” and duplicates the effort; C confuses instructions (CLAUDE.md) and server configuration; D versions a secret — eliminatory security fault.

Answer: C. Path-scoped rules (.claude/rules/ + glob patterns) load uniquely when editing matching files — exactly what is needed (conditional + context saving). Permanently dependent on conventions, quite the opposite; B is unmaintainable when the tests are co-located in dozens of directories (and this is the switching criterion: conventions by file type scattered → glob ; agreements by subtree → CLAUDE.md directory); D is manual and on demand — the situation calls for automatic.

Answer: B. -p (Or --print ) is the documented headless mode of Claude Code ⚠: non-interactive, suitable for pipelines, combinable with --output-format json And --json-schema for a structured output usable by the job. A and C are invented features — the most frequent family of distractors in this scenario; D is a Unix workaround that does not match the documented execution mode.

Answer: B. The three plan mode markers are combined: significant change (45 files), several viable approaches , architectural decisions (service boundaries). Plan mode allows safe exploration before any modification. A is the classic “reactive” distractor: when “things go wrong”, recovery is already costly; C avoids local disasters but does not solve the central problem (service boundaries are a decision global ) ; D destroys the inter-file consistency of the extraction.

Answer: B. This is THE distinction of Domain 4: strict JSON schemas eliminate errors of syntax , not the errors semantics (bad field, unreconciled totals). The solution is programmatic: consistency validations + retest with concrete error feedback (document + faulty extraction + specific errors). A is false — no type constraint prevents putting the right number in the wrong field; C regresses (reintroduces the syntax errors that tool_use eliminated); D increases variability, making the problem worse.

Answer: A. A field required on information that is sometimes missing structurally strong the model to produce something — the contrary instruction creates a contradictory injunction that the model wins. The fix is ​​in the schema design: nullable/optional gives "information does not exist" a legitimate output. Same family: enums with "other"/"unclear" + detail field. B has already shown its inefficiency (the diagram takes precedence); C tries again without changing the constraint that causes the invention; D does not detect invented dates plausible — that’s precisely the problem.

Answer: B. Two examination knowledge combines: (1) the explicit and categorical criteria surpass generic guidance — “be more conservative” (A) is precisely the documented anti-example; (2) false positives from a noisy category undermine confidence in reliable categories — hence the temporary deactivation of categories with a high rate of false positives. C (consensus by repetition) treats the noise without treating the cause and triples the cost; D reduces the variance between runs, not the over-reporting bias.

Answer: B. The documented risk of progressive summarization is exactly this: numeric values, dates and identifiers are diluted into vague formulations over successive summaries. The structural parade: a persistent block of facts out of the summary cycle. A and D only delay or attenuate the dilution (D remains probabilistic, and each summary recompresses); It's worse: the sliding window delete purely the 5th round, franchise included.

Answer: C. The aggregate metric mask weak segments — 97% overall is here compatible with 64% in a critical field for transfers. The architect's response is segmented: automate where performance is demonstrated by document type and field , keeping the human where it is not, with continuous stratified random sampling to detect deviations (new supplier formats). Exactly reproduced the trap; B over-corrects and destroys the value on 85% of the volume without justification; D perhaps improves a segment but renews the faulty criterion (global metric).