Free GH-600 practice test: 30 questions with answers

An allow list enforces the approved set in configuration rather than relying on developer discipline. A README note is advisory, open discovery is unsafe, and removing tools eliminates the capability the agent needs.

Repository-scoped configuration keeps the agent's actions inside the intended codebase. Organization-wide access, runtime repository choice, and shared tokens exceed the scope the task needs.

A clear failure with no partial changes keeps the repository consistent and surfaces the configuration problem. Placeholder secrets, exposing values in logs, and infinite retries each create a different failure or a security problem.

Defining outcomes, constraints, and measurable signals in advance allows the result to be evaluated objectively. Names, icons, and vague impressions do not make success testable.

Mismatched arguments to a tool schema are tool misuse, which can be addressed by clearer tool descriptions or validation. Infrastructure outages and unrelated factors do not explain schema violations.

Automated scans produce repeatable signals about defects and policy violations. Self-assessment by the agent is unreliable, partial reading misses problems, and commit counts do not measure quality.

A sequential pipeline with explicit handoffs is a common orchestration pattern for specialized agents. Random or isolated work and uncoordinated simultaneous edits do not produce a coherent feature.

Isolated workspaces or branches keep parallel work separate, and conflicts are resolved in a controlled merge step. Simultaneous writes, silent deletion, and disabling reviews all cause or hide conflicts.

Running the new agent in parallel, shifting work gradually, and retiring the old one with records preserved avoids disruption and keeps auditability. Restarts, deleted history, and mid-task replacement each break continuity or traceability.

Least privilege limits the agent to the folder and repository it needs, which contains mistakes and misuse. Maximum privilege, shared admin credentials, and unlimited access all expand the blast radius.

Irreversible, compliance-sensitive changes require explicit authorization through controlled channels. Acting first, having another agent approve, or falsely reporting success violate the policy.

Classifying by operational, security, and compliance risk right-sizes human review while keeping delivery fast. Name length, time of day, and avatar color are not risk factors.

Expiration and pruning rules remove stale entries systematically, and reset rules handle major changes. Keeping everything, copying all notes, and blanket deletion either preserve stale data or lose needed context.

Durable progress artifacts let the agent resume from recorded state without repeating steps or contradicting earlier decisions. Restarting wastes effort, and relying only on session memory fails once it is lost.

Short-term memory suits the current run, while persistent conventions belong in long-term or external storage. Storing temporary details permanently or restating by hand each run mismatches the need.

Removing all checkpoints lets errors compound without review, which is a common anti-pattern for agents. Structured plans, repository scoping, and artifacts are controls that reduce this risk.

Separating planning from execution, with an approval gate, prevents action before the plan is checked. Combining the steps or reviewing afterward allows changes to happen before validation.

Explicit inputs, outputs, and tools make each step checkable, and a parseable format lets validation be automated. Vague steps, discarded plans, and silent changes make validation impossible or unreliable.

Shared, versioned definitions give every repository the same reviewed behavior and make changes traceable. Unreviewed or personal instructions and undocumented setups produce drift between teams.

Without defined inputs, outputs, and success criteria, there is no shared test for acceptance. Model size, branch counts, and hardware do not determine whether a fix is acceptable.

Limiting outputs to comments and withholding commit or merge permission enforce the boundary. Write access, merge rights, and discretionary edits all extend the agent beyond commenting.

Bounded attempts followed by escalation with details give a human what they need to decide. Infinite loops, silent closure, and ignoring results each hide the failure or waste effort.

Recording the acting identity and the triggering input and decision makes each action attributable and explainable. Personal preferences, office equipment, and avatars do not establish accountability.

A controlled revert through a reviewed change restores the last good state with an audit trail. Direct production edits, deleting history, and waiting all bypass control or fail to fix the problem.

A checkpoint before destructive steps enforces validation and approval, so the plan is checked before action. Removing planning, self-approval, and repeated execution all increase the risk.

Listing tools lets reviewers and validators check permissions and scope before execution. Omitting them makes the plan unverifiable, which is why the claim that it has no effect is wrong.

A missing progress artifact over an expected interval and repeated identical calls both show the execution has stopped advancing. A new commit or a passing summary indicates progress, and avatar changes say nothing about execution.

Clear ownership and explicit handoff contracts prevent confusion about who does what and in what format. Unlocked shared edits, no coordinator, and removing timeouts all make failures more likely or harder to recover from.

Rolling back to the last known good state restores consistency after a failed change. Continuing from failure, rebuilding from memory, and committing broken state all leave the repository inconsistent.

Timestamped handoff logs and decision records reconstruct how the workflow unfolded and why. The final output alone hides the path, and screenshots and display names do not explain the behavior.