GH-300 - Understand GitHub Copilot data and architecture (13% of the exam) - Section 3.2
Describe the proxy filtering pipeline, post-processing, and the code suggestion lifecycle.
Trace a suggestion from the editor through the GitHub Copilot proxy, where input processing, filtering for toxic language and insecure patterns, and post-processing happen, and back to the editor. Describe the code suggestion lifecycle and the role of the duplication-detection filter for matching public code.
lock_openFree sampleUnderstand GitHub Copilot data and architecturemedium
A developer is mapping the GitHub Copilot inline completion lifecycle from keystroke to displayed ghost text. They want the correct order of the main stages. Which sequence reflects how a completion is produced?
AThe editor displays a draft candidate, the proxy filters it after the developer reads it, the model regenerates from the edit, then local context is gathered for the next keystroke entirely.
BThe model generates a candidate before any context is gathered, the editor displays it immediately, and the proxy filters only suggestions that the developer has already accepted into the file.
CThe proxy generates the candidate text itself, the model only filters toxic content, local context is gathered last, and the editor stores the result without showing ghost text.
DLocal context is gathered, the proxy assembles and forwards the prompt, the model generates a candidate, the proxy post-processes and filters it, then the editor displays any survivor.check_circle Correct
Trace the inline completion lifecycle as context gathering, proxy prompt assembly, model generation, proxy post-processing and filtering, then editor display. In the standard flow the client gathers surrounding context, the proxy assembles the prompt and forwards it, the model generates a candidate, and the proxy post-processes and filters that candidate before the editor renders any survivor as ghost text. Generation belongs to the model and filtering belongs to the proxy.
Why A is wrong: Tempting because filtering and display both happen, but the editor never shows an unfiltered draft first and the model does not regenerate from a displayed candidate in this way.
Why B is wrong: Plausible to those who think filtering is post-acceptance, but context precedes generation and the proxy filters candidates before display, not after acceptance.
Why C is wrong: Reverses the roles: the model generates and the proxy filters, not the other way around, and context is gathered first rather than last.
Why D is correct: Context gathering feeds prompt assembly at the proxy, the model produces a candidate, and the proxy then post-processes and filters before the editor shows what remains, matching the documented flow.
Exam traps in Understand GitHub Copilot data and architecture
Answers that look right on this material and are not. Each one is a distractor from a different question in the GH-300 bank for this domain.
It scans the developer's open files before the prompt is built and removes any lines that resemble licensed code from the context window.
errorWhy it is wrong: Duplication detection acts on generated suggestions, not by stripping lines from the input context before the prompt is built, so this misplaces where in the flow the control operates.
Inside the editor extension on the developer's machine, which holds a local copy of public repositories and matches each candidate against them before the proxy is contacted.
errorWhy it is wrong: Tempting because the extension is local, but it does not carry a public-code corpus, and duplication detection runs in the proxy rather than on the client.
The editor uploads the developer's entire local repository to GitHub on every keystroke so the model can index the whole project before replying.
errorWhy it is wrong: Copilot does not upload the whole repository on each keystroke; it assembles a bounded prompt of surrounding context, so describing a full-repo upload per keystroke overstates what is transmitted.