Core viva

Fast, adaptive, interview-style questioning on CS fundamentals. The format is the one used in placement technical rounds: short questions, short answers, the interviewer probing deeper whenever an answer is shaky.

0. Setup

1. Call the Skill tool with "cracked-profile".
2. Settle the scope in one line, with defaults: subjects (default: all five, OS · DBMS · CN · OOP · SQL), number of questions (default 15), level (default: adaptive). If the user names a subject or a topic ("only deadlocks", "normalization"), run just that.
3. Company mode: if the user names a company, call the Skill tool with "company-intel". Weight subjects and topics by what its brief reports for technical rounds (service and mass-recruiter archetypes lean heavily on OS/DBMS/OOP/SQL theory; product companies on DBMS, SQL and OOP design; networking-heavy companies on CN). If the brief has web-sourced, reported core questions, include them, labelled "reported for <Company>". With no company, don't ask; use the default spread.
4. Use the syllabus in SYLLABUS.md to pick topics. Spread questions across topics. Never ask two questions from the same topic in a row unless you're probing.

1. The loop

Ask one question at a time. Number them (Q4/15 · DBMS · medium).

After each answer, give instant feedback in at most four lines:

✅ Correct  /  🟡 Partial  /  ❌ Incorrect
<the ideal 1–3 sentence interview answer>
<if partial or wrong: the specific thing they missed or got wrong>

Then move on. Adaptive rules:

• Two correct in a row → raise difficulty one step.
• Wrong → ask one follow-up probe on the same topic at an easier level, to tell a gap from a slip.
• "I don't know" is a valid answer. Treat it as incorrect but say it's the right thing to say in a real interview, rather than bluffing.
• Mix question types: definition, "difference between X and Y", "what happens when…", scenario ("Your query is slow. What do you check?"), and small numericals (page faults with FIFO/LRU, subnetting, CPU scheduling waiting time).

2. SQL questions run for real

For SQL questions, give a tiny schema with sample rows and ask for a query. Say the dialect up front: "I'll run this in SQLite; MySQL/PostgreSQL syntax is fine where it's equivalent." Check sqlite3 --version once (window functions need 3.25+, FULL OUTER JOIN and RIGHT JOIN need 3.39+); if the installed version is older, use Python's sqlite3 module or say which features you can't run. When they answer:

1. Create a temp SQLite database (sqlite3 is preinstalled on macOS and most Linux systems; else use Python's sqlite3 module), load the schema and rows, and run their query.
2. Also run your reference query, and compare the results.
3. Show both result sets if they differ. Accept any query that returns the correct result, including equivalent MySQL/PostgreSQL syntax you had to translate to run it (say what you translated: TOP→LIMIT, ILIKE, :: casts…). Flag syntax that wouldn't run on MySQL or PostgreSQL, since interviews usually use one of those. Mention style issues separately.

Typical SQL targets: joins (inner/left, self join), GROUP BY + HAVING, Nth highest salary, duplicates, window functions (RANK, ROW_NUMBER), subqueries vs joins, NULL handling.

3. Results

After the last question:

1. A per-subject table: Subject | Asked | Correct | Partial | Score/10.
2. Gap list: the specific topics they got wrong, each with a one-line "revise this" pointer (e.g. "Deadlock: Coffman conditions and Banker's algorithm").
3. The scorecard, via the Skill tool with "cracked-profile", one dimension per subject asked. Weak-spot tags are the topic slugs (os-deadlock, dbms-normalization, cn-tcp-handshake, sql-window-functions).
4. Offer: "Run a deep-dive on <weakest topic>?" A deep-dive is 5 escalating questions on one topic.

Tone and accuracy

• Keep the pace brisk. No lectures between questions.
• Numericals are computed, not recalled. For page faults, CPU scheduling, subnetting and similar, compute the reference answer with a short Python snippet before grading, and say "reference checked by execution".
• Be precise. Fundamentals have right answers, and an ambiguous "ideal answer" teaches the wrong thing. If a question has textbook-dependent variants (e.g. number of OSI layers vs the TCP/IP model), say so.

Skill calls: if your agent has no Skill tool, open the named skill's SKILL.md in the cracked skills folder (core/, interview/, profile/, strategy/) and follow it instead.