fluent-db-updater - SKILL.md Agent Skill

name: fluent-db-updater description: Atomically update all 6 Fluent learner databases (learner-profile, progress, mistakes, mastery, spaced-repetition, session-log) at session end by calling .claude/hooks/update-db.py with a single JSON payload. Use at the end of every practice session — fluent-writing, fluent-vocab, fluent-speaking, fluent-reading, fluent-review, fluent-learn — to persist the session's errors, review results, new vocabulary, and session metadata.

DB Updater

Overview

Every practice skill ends with a DB update. Instead of hand-editing 6 JSON files (error-prone, racy, easy to desync), pipe one JSON report to update-db.py. The script runs pre-write backups, validates the payload, applies all changes atomically via .tmp + fsync + rename, and rebuilds the spaced-repetition queue.

When to Use

Load this skill whenever the tutor:

Finishes a practice session and needs to persist results.
Needs to add new vocabulary to the spaced-repetition queue.
Needs to record new errors, review results, or mastery changes.
Needs to bump total_sessions, current_streak_days, or total_study_minutes.

Skip this skill for read-only operations (use the fluent-progress skill or read-db.py directly) and during session setup (use fluent-setup skill instead — update-db.py is for session deltas, not bootstrap).

Instructions

1. Call the script

Run from the repo root:

python3 "${CLAUDE_PLUGIN_ROOT:-${CLAUDE_PROJECT_DIR:-.}}/.claude/hooks/update-db.py" <<'EOF'
{ ...payload... }
EOF

Exit codes: 0 success, 1 validation error, 2 I/O error.

2. Fill the payload

Required fields

session_id — string, convention session-NNN. Use computed.next_session_id from read-db.py.
date — YYYY-MM-DD.

Optional fields — omit to skip. Full canonical example (copy-paste this and fill in):

${CLAUDE_PLUGIN_ROOT:-${CLAUDE_PROJECT_DIR:-.}}/.claude/references/db-updater-payload.example.json

Key blocks the example covers: skill_scores, errors[], new_vocabulary[], review_results[], topics_covered, breakthroughs, focus_next_session, session_notes, achievements_earned, milestones.

3. Field notes

errors[] — one entry per distinct mistake this session. Collapse duplicates (same pattern_id) before sending; frequency is bumped by the script.
new_vocabulary[] — items the learner met for the first time. Fill every field; incomplete entries yield incomplete spaced-repetition records.
review_results[] — items already in the queue that were reviewed. The script runs SM-2 on each. See the fluent-sm2-calculator skill. Mapping: quality = floor(score / 2).
skill_scores[].correct counts correct exercises, not a percentage. Accuracy is derived.
confidence in learner-profile.skills is 0–100 integer; accuracy in progress-db is 0.0–1.0 float. The script handles the conversion.
milestones[] — each entry is a bare string OR an object { "milestone": <required non-empty string>, "date": <optional YYYY-MM-DD, defaults to the session date> }. Don't set a nested session_id; the script stamps the authoritative top-level one. A malformed entry (neither string nor object, or an object missing/empty milestone) exits 1 with no files written. Each milestone becomes both a session-log.milestones[] record and a learner-profile.achievements[] entry.

4. Read before writing

Always call read-db.py at session start to get current state + next_session_id. Don't read each JSON file separately:

python3 "${CLAUDE_PLUGIN_ROOT:-${CLAUDE_PROJECT_DIR:-.}}/.claude/hooks/read-db.py"

Returns all 6 databases plus computed fields (due_reviews_count, next_session_id, streak_active, days_since_last_session).

Examples

Example 1 — /fluent-review session with 5 items

python3 "${CLAUDE_PLUGIN_ROOT:-${CLAUDE_PROJECT_DIR:-.}}/.claude/hooks/update-db.py" <<'EOF'
{
  "session_id": "session-012",
  "date": "2026-04-24",
  "duration_minutes": 12,
  "command_used": "/fluent-review",
  "skills_practiced": ["vocabulary", "grammar"],
  "skill_scores": {
    "vocabulary": { "exercises": 3, "correct": 3, "time_minutes": 7 },
    "grammar":    { "exercises": 2, "correct": 1, "time_minutes": 5 }
  },
  "review_results": [
    { "item_id": "vocab_huis", "quality": 5 },
    { "item_id": "vocab_deur", "quality": 4 },
    { "item_id": "vocab_raam", "quality": 5 },
    { "item_id": "grammar_omdat_word_order", "quality": 2 },
    { "item_id": "grammar_past_tense", "quality": 4 }
  ],
  "errors": [
    {
      "pattern_id": "grammar_omdat_word_order",
      "category": "grammar",
      "your_answer": "omdat ik ben moe",
      "correct_answer": "omdat ik moe ben",
      "context": "subordinate clause word order",
      "severity": "critical"
    }
  ],
  "focus_next_session": ["Drill 'omdat' word order"]
}
EOF

Example 2 — /fluent-vocab session with a new word

python3 "${CLAUDE_PLUGIN_ROOT:-${CLAUDE_PROJECT_DIR:-.}}/.claude/hooks/update-db.py" <<'EOF'
{
  "session_id": "session-013",
  "date": "2026-04-25",
  "command_used": "/fluent-vocab",
  "skills_practiced": ["vocabulary"],
  "new_vocabulary": [
    {
      "item_id": "vocab_keuken",
      "item_type": "vocabulary",
      "content": "de keuken",
      "answer": "the kitchen",
      "category": "household_rooms",
      "difficulty": "A1",
      "initial_quality": 4,
      "priority": "medium"
    }
  ]
}
EOF

Critical Rules

Call once per session, at the end. The script rebuilds the review queue each run — partial updates risk inconsistency.
Never hand-edit spaced-repetition.review_queue. It's regenerated from scratch on every run.
Same session_id replaces. Sending the same ID twice overwrites the first call. Useful for corrections, dangerous if unintentional.
Backups are automatic. Written to .backups/pre-update-<session_id>/ before any change. Check there to roll back.
Exit code 1 means validation failed, no files touched. Fix the payload and retry.
Exit code 2 means I/O failure, no files touched. Check disk space, permissions, then retry.

Why This Matters

Six interdependent JSON files must agree: a new session-log entry, a bumped total_sessions, updated SM-2 params, new mistake patterns, recalculated accuracy, refreshed streak. Hand-editing causes silent desync — streak says 7 days but session-log has 6 entries, mastery says 4 stars but accuracy says 45%. The script is the single source of truth.