Skip to content

patch2.md

Latest commit

 

History

History
174 lines (147 loc) · 8.96 KB

patch2.md

File metadata and controls

174 lines (147 loc) · 8.96 KB

MockFlow-AI Patch 2 — Technical Coding Track

Overview

Patch 2 wires up the Technical Coding track stubbed in Patch 1. Adds a Monaco editor split-view, LLM-generated coding problems, a code evaluator (separate LLM call), per-problem timers, retry logic, and coding-specific feedback.

Architecture Decisions

Decision Choice
Code submission transport HTTP POST to /api/coding/submit (reliable for large code payloads)
Problem delivery to frontend LiveKit data channel {"type": "coding_problem", "problem": {...}}
Problem generation LLM-generated at WARM_UP stage, stored in CodingInterviewState
Monaco editor CDN via RequireJS, lazy-loaded only for coding track
Split layout .interview-main.coding-mode CSS class enables 3-column layout
Timer Pure JS countdown, auto-submits on expire
Retry limit 3 attempts per problem
Agent during coding Silent observer — speaks only on submission, user question, or timer expiry
Code evaluation Separate openai.chat.completions.create() call (not voice context)

Files to Create

  • supabase-backend/patch2_migration.sql — coding_submissions table + RLS

Files to Modify

File Change
fsm.py CodingInterviewState, CODING_STAGE_TIME_LIMITS, CODING_STAGE_MIN_QUESTIONS
tracks/technical_coding.py Real time limits, is_available=True
prompts.py CODING_GREETING/SELF_INTRO/WARM_UP/PROBLEM_STAGE/CLOSING, CODE_EVALUATOR, extend QUESTION_GENERATION
agent_worker.py Coding dispatch branches, evaluate_code_submission tool, skip_coding_problem tool, code_submitted data handler
app.py /api/coding/submit endpoint
supabase_client.py save_coding_submission()
templates/form.html Enable coding card (remove disabled/Coming Soon), add language/difficulty/count fields
templates/interview.html Monaco editor panel, split layout, submit/skip buttons, timer, retry counter
static/interview.css .coding-mode 3-column layout, editor panel styles

Ordered Task List

Step 1 — CodingInterviewState + Constants (fsm.py)

CODING_STAGE_TIME_LIMITS = {
    CodingStage.GREETING: 30,
    CodingStage.SELF_INTRO: 120,
    CodingStage.WARM_UP: 300,
    CodingStage.CODING_PROBLEM_1: 900,
    CodingStage.CODING_PROBLEM_2: 900,
    CodingStage.CLOSING: 45,
}

@dataclass
class CodingInterviewState(InterviewState):
    track_type: str = "coding"
    preferred_language: str = "python"
    active_problem_count: int = 2
    generated_problems: List[dict] = field(default_factory=list)
    current_problem_index: int = 0
    submissions: List[dict] = field(default_factory=list)
    submissions_per_problem: dict = field(default_factory=dict)
    problem_start_times: dict = field(default_factory=dict)
    coding_stage_active: bool = False

Methods: get_active_stages(), get_next_coding_stage(), get_attempts_for_problem(idx), override get_stage_time_limit(), override get_document_context().

Step 2 — Enable Track Config (tracks/technical_coding.py)

Real time limits from CODING_STAGE_TIME_LIMITS, is_available=True.

Step 3 — Coding Prompts (prompts.py)

  • CODING_GREETING — brief greeting, confirm readiness
  • CODING_SELF_INTRO — background, preferred language
  • CODING_WARM_UP — past coding experience, confirm language
  • CODING_PROBLEM_STAGE — reads problem aloud, goes silent; template vars: {problem_title}, {problem_description}, {examples}, {constraints}, {time_limit_minutes}, {attempt_number}, {max_attempts}
  • CODING_CLOSING — wrap up
  • CODE_EVALUATOR.system — returns structured JSON: {correctness, approach_quality, edge_cases_handled, edge_cases_missed, time_complexity, space_complexity, code_quality_notes, suggestions, brief_verbal_feedback}
  • QUESTION_GENERATION.coding_system — generates {problems: [{title, description, examples, constraints, difficulty, time_limit_minutes, hints}]}
  • Update build_stage_instructions(), get_transition_ack(), get_fallback_ack() for CodingStage

Step 4 — Agent Worker Coding Branches (agent_worker.py)

  • State init: elif track_type == 'coding' → instantiate CodingInterviewState, read preferred_language
  • Stage transition: add coding dispatch branch
  • generate_interview_questions: add elif track_type == 'coding' → generate problems, set active_problem_count
  • get_current_question: add coding branch → emit {"type": "coding_problem", ...} to UI
  • New tool evaluate_code_submission(problem_index, code, language) → separate OpenAI call → POST to /api/coding/submit → return brief_verbal_feedback
  • New tool skip_coding_problem() → marks problem skipped → transitions stage
  • Data channel: add code_submitted handler → validate attempt count ≤ 3 → call evaluate_code_submission
  • finalize_and_disconnect: include problems_generated, preferred_language, submissions in track_config

Step 5 — HTTP Endpoint (app.py + supabase_client.py)

POST /api/coding/submit → stores to coding_submissions table; save_coding_submission() in supabase_client.

Step 6 — Database Migration (supabase-backend/patch2_migration.sql)

CREATE TABLE IF NOT EXISTS coding_submissions (
    id UUID PRIMARY KEY DEFAULT uuid_generate_v4(),
    interview_id UUID NOT NULL REFERENCES interviews(id) ON DELETE CASCADE,
    user_id UUID NOT NULL REFERENCES auth.users(id) ON DELETE CASCADE,
    problem_title TEXT NOT NULL,
    problem_description TEXT,
    language VARCHAR NOT NULL,
    code_submitted TEXT NOT NULL,
    attempt_number INTEGER NOT NULL DEFAULT 1,
    evaluation_result JSONB,
    time_spent_seconds INTEGER,
    created_at TIMESTAMP WITH TIME ZONE DEFAULT NOW()
);
CREATE INDEX IF NOT EXISTS idx_coding_submissions_interview ON coding_submissions(interview_id);
ALTER TABLE coding_submissions ENABLE ROW LEVEL SECURITY;
CREATE POLICY "Users can view own submissions" ON coding_submissions FOR SELECT USING (auth.uid() = user_id);
CREATE POLICY "Service can insert submissions" ON coding_submissions FOR INSERT WITH CHECK (auth.uid() = user_id);

Step 7 — Form Page (templates/form.html)

  • Remove disabled class/Coming Soon from coding card
  • Add coding section: language dropdown (Python/JS/Java/C++/Go), problem count (1/2), difficulty (Easy/Medium/Hard)
  • Pass preferred_language, problem_count, problem_difficulty to /api/token

Step 8 — Interview Room (templates/interview.html + static/interview.css)

HTML additions:

  • <div class="code-editor-panel" id="codeEditorPanel"> with: problem display area, language select, problem timer, attempt counter, Monaco container id="monacoEditorContainer", Submit + Skip buttons
  • Monaco CDN script (lazy-loaded for coding track only)

JS additions:

  • initMonaco() — RequireJS loads Monaco, creates editor instance
  • setEditorLanguage(lang), getEditorCode(), setEditorCode(code)
  • handleCodeSubmit() — POST to /api/coding/submit + signal agent via data channel
  • startProblemTimer(seconds) — countdown, auto-submit at 0
  • updateAttemptCounter(attempt, max)
  • Data channel: handle coding_problem → display problem, start timer; handle evaluation_result → show feedback

CSS additions:

.interview-main.coding-mode { grid-template-columns: 1fr 2fr 1fr; }
.code-editor-panel { display: flex; flex-direction: column; gap: 0.75rem; }
.problem-display { overflow-y: auto; max-height: 200px; font-size: 0.875rem; }
.problem-timer { font-variant-numeric: tabular-nums; font-weight: 700; font-size: 1.25rem; }
@media (max-width: 1024px) { .interview-main.coding-mode { grid-template-columns: 1fr; } }

Step 9 — Coding Feedback (app.py + prompts.py)

  • In feedback endpoints: if track == 'coding', fetch submissions from coding_submissions, pass to LLM
  • TRACK_FEEDBACK.coding_track_specific_schema: per-problem code quality grade, approach grade, edge cases, time management

Step 10 — End-to-End Verification

  1. Form → coding track → language/difficulty fields appear → submit → interview starts
  2. Agent greets → self intro → warm up discussion → generates 2 problems at WARM_UP
  3. First problem delivered via data channel → Monaco appears → timer starts 15:00
  4. User writes code → Submit → evaluation returned → agent speaks brief_verbal_feedback
  5. Retry (attempt 2/3) → evaluation again → counter updates
  6. Timer expires → auto-submit → agent evaluates
  7. Skip problem → transitions to CODING_PROBLEM_2 or CLOSING
  8. After closing → coding_submissions table has rows with evaluation_result JSON
  9. Feedback page → shows per-problem grades

Verification Command

python -c "
from fsm import CodingStage, CodingInterviewState
s = CodingInterviewState()
print(s.stage.value, [x.value for x in s.get_active_stages()])
from tracks import get_track_config
c = get_track_config('coding')
print(c.is_available, c.time_limits[CodingStage.CODING_PROBLEM_1])
"
# Expected: greeting [...] / True 900