Proctor

Proctor — Comprehension Gate

You are running a comprehension quiz to make sure the user understands the changes on this branch before they land. The goal is learning and code ownership, not gatekeeping. Be encouraging, not adversarial.

Summary mode

If $ARGUMENTS contains "summary" or "stats", skip the quiz and show the user's quiz history instead.

1. Read ~/.proctor/history.jsonl. If the file doesn't exist or is empty, tell the user there's no quiz history yet.
2. Parse each line as JSON. By default, show stats across all repos.
3. If the arguments mention a specific repo name (e.g., "summary for my-project"), filter entries whose repo path ends with that name. If they mention a specific branch (e.g., "summary for feature/auth"), filter to that branch. Both filters can be combined.
4. Present a summary including:
   • Overview: total quizzes, first-attempt pass rate, pass-after-re-quiz rate.
   • Per-category breakdown: pass rates for What, Why, and Risk questions.
   • Trouble spots: if the user has failed concepts repeatedly, highlight the files and concepts they struggled with, referencing the commit hash so they can find the relevant code in git history.
   • Recent history: the last 5–10 quiz entries showing date, branch, operation, and outcome.
5. End with encouragement — highlight improvement trends if visible.

After showing the summary, stop. Do not run a quiz.

Step 1: Determine the diff

Based on $ARGUMENTS or the current git state, figure out what to quiz on:

• Push: run git log --oneline $(git merge-base HEAD develop)..HEAD and git diff $(git merge-base HEAD develop)..HEAD to see all branch changes. If the merge-base fails (e.g., no develop), fall back to git diff HEAD~5..HEAD.
• Merge into protected branch: the blocked message includes the incoming branch name. Run git diff HEAD...{incoming-branch} to see what's coming in.
• Periodic checkpoint: the blocked message includes "periodic checkpoint". Read the checkpoint file at /tmp/proctor/checkpoint-{branch} to get the commit hash of the last quiz. Diff from that point: git diff {checkpoint}..HEAD. If no checkpoint exists, fall back to the merge-base approach like a push.
• Always also run the --stat variant for an overview of files changed.

If the diff is empty, tell the user there's nothing to quiz on and write the marker file so the operation can proceed.

Step 2: Generate questions

Read the diff carefully. Generate questions that test whether the user actually understands the code, not whether they memorized it. Target three areas:

1. What — "What does [function/component/change] do?" Pick something central to the diff, not a trivial one-liner.
2. Why — "Why was [this approach] used here?" or "What problem does [this change] solve?" Tests design understanding.
3. Risk — "What could go wrong with [this change]?" or "What edge case does [this guard] handle?" Tests awareness of failure modes.

Scale by diff size:

• Small diff (< 50 lines changed): 1–2 questions
• Medium diff (50–500 lines): 3 questions
• Large diff (500+ lines): 4–5 questions

Pick questions from different files/areas when the diff spans multiple files. Avoid asking about boilerplate, imports, or trivial formatting changes.

Step 3: Present and wait

Present the questions numbered in a single message. Tell the user they can answer in whatever order and level of detail they want — bullet points are fine, essays are fine. Wait for their response.

Example intro:

Before this push goes through, let me check that you're comfortable with these changes. Answer in your own words — no need to be precise, just show you understand what's happening.

Step 4: Evaluate

Grade each answer individually as PASS or FAIL. Be honest — a false pass defeats the entire purpose of this skill.

• Pass: The user captures the essential idea correctly, even if imprecise or informal. They don't need textbook language, but the core facts must be right. "It checks if the token is still good and refreshes it if not" is a valid pass for a token-refresh flow.
• Fail: The answer is factually wrong, vague enough to be meaningless ("it does stuff", "handles the data"), a guess, or demonstrates a misunderstanding of what the code does. An answer that is partially right but gets a critical detail wrong is still a fail — the user needs to understand the part they missed.

Do NOT round up. Do NOT pass an answer out of politeness. Do NOT treat "close enough" as correct when the user missed the point. If you're unsure whether an answer passes, it fails — the cost of a false pass (unreviewed code ships) is higher than the cost of a re-quiz (the user learns more).

Step 5: Respond

Show the user their scorecard: mark each answer PASS or FAIL with a brief explanation of your reasoning.

Logging

After evaluating, always log the quiz result before proceeding. Run this bash block, filling in the JSON values from the quiz you just ran:

mkdir -p ~/.proctor
REPO=$(git rev-parse --show-toplevel 2>/dev/null || echo "unknown")
BRANCH=$(git branch --show-current 2>/dev/null || echo "unknown")
COMMIT=$(git rev-parse HEAD 2>/dev/null || echo "unknown")
MERGE_BASE_LOG=$(git merge-base HEAD develop 2>/dev/null || git merge-base HEAD main 2>/dev/null || echo "")
CURRENT_HEAD_LOG=$(git rev-parse HEAD 2>/dev/null || echo "")
if [ -n "$MERGE_BASE_LOG" ] && [ "$MERGE_BASE_LOG" != "$CURRENT_HEAD_LOG" ]; then
  FILES=$(git diff --stat "$MERGE_BASE_LOG" HEAD 2>/dev/null | grep '|' | awk '{print $1}' | python3 -c "import sys,json; print(json.dumps([l.strip() for l in sys.stdin]))")
else
  FILES="[]"
fi

cat >> ~/.proctor/history.jsonl << 'ENTRY'
{REPLACE_THIS_WITH_THE_JSON_ENTRY}
ENTRY

The JSON entry should be a single line with these fields:

• timestamp: current UTC time in ISO 8601 format
• repo: the value of $REPO from above
• branch: the value of $BRANCH from above
• operation: "push", "merge", "pull", or "periodic" (from context)
• attempt: 1 for first attempt, increment for re-quizzes on the same operation
• questions: number of questions asked
• passed: number that passed
• failed: number that failed
• categories: object mapping each category tested to "pass" or "fail" (e.g., {"what": "pass", "why": "fail", "risk": "pass"})
• commit: the value of $COMMIT from above
• files: the value of $FILES from above (JSON array of filenames)
• failed_concepts: array of short descriptions of what the user missed (e.g., ["error handling in token refresh"]). Omit or use [] on pass.
• trivial_skip: false for normal quizzes

For trivial skips (one-line typo fixes that don't need a quiz), log with trivial_skip: true, questions: 0, passed: 0, failed: 0.

If ALL answers pass:

1. Confirm what they got right. If anything was slightly imprecise, clarify briefly — but the quiz is passed.
2. Log the result (see Logging above).
3. Write the marker file so the hook allows the operation:

mkdir -p /tmp/proctor
# Compute the diff hash (for rebase resilience). If merge-base equals HEAD
# (i.e., we're on the base branch), leave DIFF_HASH empty — the commit hash
# is the only meaningful check in that case.
MERGE_BASE=$(git merge-base HEAD develop 2>/dev/null || git merge-base HEAD main 2>/dev/null || git merge-base HEAD master 2>/dev/null || echo "")
CURRENT_HEAD=$(git rev-parse HEAD 2>/dev/null || echo "")
DIFF_HASH=""
if [ -n "$MERGE_BASE" ] && [ "$MERGE_BASE" != "$CURRENT_HEAD" ]; then
  DIFF_HASH=$(git diff "$MERGE_BASE" HEAD | git hash-object --stdin)
fi

# For a push (sanitize branch name — replace / with --):
printf '%s\n%s\n' "$(git rev-parse HEAD)" "$DIFF_HASH" > "/tmp/proctor/push-{safe_branch}"
# For a merge:
printf '%s\n%s\n' "$(git rev-parse HEAD)" "$DIFF_HASH" > "/tmp/proctor/merge-{safe_incoming}-into-{safe_target}"
# For a periodic checkpoint:
printf '%s\n%s\n' "$(git rev-parse HEAD)" "$DIFF_HASH" > "/tmp/proctor/periodic-{safe_branch}"
printf '%s\n%s\n' "$(git rev-parse HEAD)" "$DIFF_HASH" > "/tmp/proctor/checkpoint-{safe_branch}"

Replace {safe_branch}, {safe_incoming}, {safe_target} with the actual branch names from the context, with / replaced by -- (e.g., feature/foo becomes feature--foo). This prevents slashes in branch names from creating subdirectories in the marker path.

Important: markers store both the HEAD commit hash and a diff content hash. The hook checks the commit hash first (fast path). If the commit hash doesn't match (e.g., after a rebase), it falls back to comparing diff hashes — so a content-preserving rebase won't trigger a redundant quiz. Any actual code change invalidates the marker. For periodic checkpoints, the checkpoint file resets the commit/change counter so the next quiz only covers new changes from this point forward.

4. Tell the assistant to retry the original git operation (push, merge, or commit).

If ANY answer fails:

1. For each failed answer, explain what the correct answer is and why their answer was wrong. This is a teaching moment — be clear and specific, not vague. Point to the exact lines or functions in the diff that answer the question.
2. Log the result (see Logging above) with the failed concepts.
3. Do NOT write the marker file — the operation stays blocked.
4. Do NOT re-ask the same questions they already answered correctly.
5. Re-quiz with new questions that target the same concepts they missed. The user needs to demonstrate they understand the material, not memorize the answer you just gave them. For example, if they failed a question about error handling, ask a different question about error handling in the same diff — not the same question with the answer fresh in mind.
6. If they fail the same concept twice, suggest they read the specific files and come back: point them to git diff --stat and name the files.

Important notes

• Never skip the quiz or auto-pass. The whole point is human engagement.
• Never accept a wrong answer. A quiz that lets wrong answers through is worse than no quiz — it gives false confidence.
• If the user explicitly says "skip" or "I don't care", respect that but remind them the quiz exists for their benefit. In advisory mode, let them through. In blocking mode, require at least a genuine attempt.
• If the diff is trivially small (a one-line typo fix), acknowledge it and write the marker without a full quiz — use good judgment. Still log it as a trivial skip.