#!/usr/bin/env bash
# continuous-improve.sh — Continuous enrichment and quality loop
#
# Each cycle: processes every entity (resolve reviews, enrich from outside
# sources), then runs a deep cross-document validation scan.
#
# Usage: .automate/continuous-improve.sh [options]
#   --priority reviews|stale|random   Queue ordering (default: reviews)
#   --cycle-delay N                   Seconds between entities (default: 5)
#   --model MODEL                     LLM model (default: claude-sonnet-4.6)
#   --start N                         Skip first N entities in queue (resume)
#   --skip-unchanged                  Skip entities unchanged since last pass

set -euo pipefail

# ─── Parse arguments ───
PRIORITY="reviews"
CYCLE_DELAY=5
MODEL="claude-sonnet-4.6"
START_AT=0
SKIP_UNCHANGED=true
while [[ $# -gt 0 ]]; do
    case "$1" in
        --priority) PRIORITY="$2"; shift 2 ;;
        --cycle-delay) CYCLE_DELAY="$2"; shift 2 ;;
        --model) MODEL="$2"; shift 2 ;;
        --start) START_AT="$2"; shift 2 ;;
        --skip-unchanged) SKIP_UNCHANGED=true; shift ;;
        *) echo "Usage: $0 [--priority reviews|stale|random] [--cycle-delay N] [--model MODEL] [--start N] [--skip-unchanged]"; exit 1 ;;
    esac
done

SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
DOCS_DIR="$SCRIPT_DIR/.."
STATE_FILE="$SCRIPT_DIR/.improve-state.tsv"
ACTION_LOG="$SCRIPT_DIR/improve-history.log"

exec > >(tee -a "$SCRIPT_DIR/continuous-improve.log") 2>&1

log() { echo "[$(date '+%Y-%m-%d %H:%M:%S')] $*"; }

do_commit() {
    local msg="$1"
    cd "$DOCS_DIR"
    git add -A
    if ! git diff --cached --quiet; then
        git commit -m "$msg"
        for attempt in 1 2 3; do
            if git push 2>/dev/null; then
                log "✅ Committed: $msg"
                return 0
            fi
            log "⚠️  Push attempt $attempt failed, rebasing..."
            git pull --rebase
        done
        log "❌ Push failed after 3 attempts"
    fi
}

build_queue() {
    cd "$DOCS_DIR"
    local tmpfile
    tmpfile=$(mktemp)

    find . -name '*.md' \
        -not -path './.git/*' \
        -not -path './.automate/*' \
        -not -path './.kiro/*' \
        -not -path './.factbase/*' \
        -not -path './_orphans.md' \
        -print0 | while IFS= read -r -d '' file; do

        local fb_id
        fb_id=$(grep -oP '(?<=factbase:)[a-f0-9]{6}' "$file" 2>/dev/null | head -1)
        [[ -z "$fb_id" ]] && continue

        local review_count
        review_count=$(grep -c '^\- \[ \] `@q\[' "$file" 2>/dev/null) || true

        local mtime
        mtime=$(stat -c %Y "$file")

        local last_processed=0
        if [[ -f "$STATE_FILE" ]]; then
            last_processed=$(grep "^${fb_id}	" "$STATE_FILE" 2>/dev/null | tail -1 | cut -f2) || true
            [[ -z "$last_processed" ]] && last_processed=0
        fi

        local garbage_count
        garbage_count=$(grep -ciP '^\[\^.*\b(not a conflict|sequential|boundary overlap|not simultaneous|malformed tag|garbled|artifact|remove)\b' "$file" 2>/dev/null) || true

        # Flag ruler docs with incomplete names (single word, alias, no space)
        local incomplete_name=0
        local parent_dir
        parent_dir=$(echo "$file" | sed 's|^\./||' | rev | cut -d/ -f2 | rev)
        if [[ "$parent_dir" == "rulers" ]]; then
            local doc_title
            doc_title=$(grep '^# ' "$file" 2>/dev/null | head -1 | sed 's/^# //' | sed 's/ @t\[.*//;s/ \[\^.*//')
            if [[ -n "$doc_title" ]] && ! echo "$doc_title" | grep -q ' '; then
                incomplete_name=1
            fi
        fi

        echo -e "${fb_id}\t${file}\t${review_count}\t${mtime}\t${last_processed}\t${garbage_count}\t${incomplete_name}"
    done > "$tmpfile"

    case "$PRIORITY" in
        reviews) sort -t$'\t' -k7,7rn -k3,3rn -k6,6rn -k5,5n "$tmpfile" ;;
        stale)   sort -t$'\t' -k7,7rn -k5,5n -k3,3rn "$tmpfile" ;;
        random)  shuf "$tmpfile" ;;
    esac

    rm -f "$tmpfile"
}

mark_processed() {
    echo -e "$1\t$(date +%s)" >> "$STATE_FILE"
}

# ─── Bash-based mechanical cleanup (no agent needed) ───
bash_cleanup() {
    local file="$1"
    local changed=false

    # Fix corrupted title: strip @t[...] and [^N] suffixes
    if grep -qP '^# .+(\s+@t\[|\s+\[\^)' "$file" 2>/dev/null; then
        sed -i -E 's/^(# .+?)\s+(@t\[.*|\[\^.*)$/\1/' "$file"
        changed=true
    fi

    # Delete garbage footnotes (review answers dumped as source citations)
    if grep -qiP '^\[\^\d+\]:.*\b(not a conflict|sequential role|boundary overlap|not simultaneous|malformed tag|garbled|artifact from previous|remove garbled)\b' "$file" 2>/dev/null; then
        # Get line numbers of garbage footnotes, delete them
        local lines
        lines=$(grep -niP '^\[\^\d+\]:.*\b(not a conflict|sequential role|boundary overlap|not simultaneous|malformed tag|garbled|artifact from previous|remove garbled)\b' "$file" | cut -d: -f1 | sort -rn)
        if [[ -n "$lines" ]]; then
            for ln in $lines; do
                sed -i "${ln}d" "$file"
            done
            changed=true
        fi
    fi

    # Remove review answer artifact lines in body
    if grep -qP '^- Artifact from previous review application' "$file" 2>/dev/null; then
        sed -i '/^- Artifact from previous review application/d' "$file"
        changed=true
    fi

    # Remove empty Review Queue sections and factbase:review markers
    if grep -qP '^## Review Queue|<!-- factbase:review -->' "$file" 2>/dev/null; then
        # Only remove if the review queue has no actual unanswered questions
        local has_open
        has_open=$(grep -c '^\- \[ \] `@q\[' "$file" 2>/dev/null) || true
        if [[ "$has_open" -eq 0 ]]; then
            sed -i '/^## Review Queue$/d; /^<!-- factbase:review -->$/d' "$file"
            changed=true
        fi
    fi

    # Remove answered review Q&A lines (- [x] `@q[...]` and their > answer lines)
    if grep -qP '^\- \[x\] `@q\[' "$file" 2>/dev/null; then
        local has_open
        has_open=$(grep -c '^\- \[ \] `@q\[' "$file" 2>/dev/null) || true
        if [[ "$has_open" -eq 0 ]]; then
            # Remove - [x] `@q[...] lines and the > answer line that follows each
            awk '
                /^\- \[x\] `@q\[/ { skip=1; next }
                skip && /^>/ { skip=0; next }
                { skip=0; print }
            ' "$file" > "${file}.tmp" && mv "${file}.tmp" "$file"
            changed=true
        fi
    fi

    # Remove duplicate H1 headings (keep first, remove subsequent)
    local h1_count
    h1_count=$(grep -c '^# ' "$file" 2>/dev/null) || true
    if [[ "$h1_count" -gt 1 ]]; then
        # Keep the first H1, delete all others
        awk '/^# / { if (++count > 1) next } { print }' "$file" > "${file}.tmp" && mv "${file}.tmp" "$file"
        changed=true
    fi

    # Clean up trailing blank lines
    sed -i -e :a -e '/^\n*$/{$d;N;ba' -e '}' "$file"

    [[ "$changed" == true ]]
}

get_related_context() {
    local file="$1"
    local title
    title=$(grep '^# ' "$file" | head -1 | sed 's/^# //' | sed 's/ @t\[.*//;s/ \[\^.*//')
    [[ -z "$title" ]] && return

    cd "$DOCS_DIR"
    local related=""
    while IFS= read -r match_file; do
        [[ "$match_file" == "$file" ]] && continue
        local match_title
        match_title=$(grep '^# ' "$match_file" 2>/dev/null | head -1 | sed 's/^# //')
        local excerpt
        excerpt=$(grep -i "$title" "$match_file" 2>/dev/null | head -3)
        if [[ -n "$excerpt" ]]; then
            related+="--- From: ${match_title} (${match_file}) ---
${excerpt}

"
        fi
    done < <(grep -rl "$title" . --include='*.md' \
        | grep -v '.git\|.automate\|.kiro\|.factbase\|_orphans' \
        | head -5)

    echo "$related"
}

# ─── Fetch existing Vikunja bug titles (for dedup) ───
get_existing_bug_titles() {
    curl -s "https://vikunja.home.everyonce.com/api/v1/projects/2/tasks" \
        -H "Authorization: Bearer tk_ff251f3d3512775c71913bc2f8ec0dabbf5016a8" \
        2>/dev/null | grep -oP '"title":"[^"]*"' | sed 's/"title":"//;s/"$//' || true
}

# ─── Agent-based processing (review questions + enrichment) ───
process_entity_agent() {
    local fb_id="$1"
    local file="$2"
    local title="$3"
    local review_count="${4:-0}"

    cd "$DOCS_DIR"

    local content
    content=$(cat "$file")

    local related
    related=$(get_related_context "$file")

    local existing_bugs
    existing_bugs=$(get_existing_bug_titles)

    local prompt
    prompt="You are continuously improving a factbase knowledge base, one entity at a time.
Use factbase MCP tools — call get_authoring_guide if you need formatting rules.

ENTITY FILE: $file
ENTITY ID: $fb_id

=== CURRENT DOCUMENT CONTENT ===
$content
=== END DOCUMENT CONTENT ==="

    if [[ -n "$related" ]]; then
        prompt+="

=== MENTIONS IN OTHER DOCUMENTS ===
$related
=== END MENTIONS ==="
    fi

    if [[ -n "$existing_bugs" ]]; then
        prompt+="

=== EXISTING VIKUNJA BUG REPORTS (do NOT file duplicates) ===
$existing_bugs
=== END EXISTING BUGS ==="
    fi

    # Build review step conditionally
    local review_step=""
    if [[ "$review_count" -gt 0 ]]; then
        review_step='1. RESOLVE REVIEW QUESTIONS:
   Call get_review_queue(doc_id='"'"''"$fb_id"''"'"') — if there are open (unanswered) questions, answer them.

   Patterns learned from resolving thousands of these:
   - CONFLICT (chronological overlaps): Boundary-year overlaps in sequential reigns or periods
     are NOT conflicts (date granularity artifact). Concurrent roles (e.g., ruler + military
     commander, pharaoh + high priest) are both true simultaneously. Approximate dates that
     overlap by a few years reflect scholarly uncertainty, not contradiction.
   - AMBIGUOUS (terms): Expand the term AND create or update a definitions/ file so the
     term is not flagged again. Do NOT just answer inline — the definitions file is what
     prevents recurrence. Check existing definitions files first with search_content.
     Common: BCE, CE, polis, satrapy, pharaoh, consul, tribune, cuneiform, stele, ziggurat.
   - TEMPORAL/STALE: BCE dates are written in text, not temporal tags. CE date ranges
     (@t[...]) that are closed are historical, not stale. For open @t[~...] tags, search
     for newer scholarship.
   - MISSING: Search with search_knowledge and search_content. If not found, defer.

   IMPORTANT: Only call apply_review_answers if get_review_queue shows unanswered > 0.
   After applying, re-read the file with get_entity and verify apply did not corrupt it
   (garbage footnotes, mangled title). If it did, fix with update_document.'
    else
        review_step='1. REVIEW QUESTIONS: SKIP — the wrapper already confirmed 0 open review questions for this entity.'
    fi

    prompt+='

STEPS — work through in order, skip any that do not apply:

'"$review_step"'

2. IDENTITY & ORGANIZATION:
   For ruler documents: if the title is a single name, alias, or epithet (not a full name),
   prioritize finding their full or commonly known name. Search local sources, check mentions
   in other documents, and cross-reference with civilization docs.
   If you find the full name, update the document title with update_document.

   For any document: if the title or file location could be improved (e.g. a ruler doc is in
   the wrong civilization folder, or the filename does not match the title), use the organize MCP
   tool to rename/move it. Use organize(action='"'"'move'"'"', doc_id=..., to=...) to relocate
   or update_document(id=..., title=...) to fix the title.

3. ENRICH FROM OUTSIDE SOURCES:
   This is the most important step. Use web_search to find high-quality information about
   this entity from scholarly and encyclopedic sources. Search for:
   - The entity name + "archaeology" or "ancient history"
   - Key events, dates, or relationships mentioned in the document
   - Recent archaeological discoveries or revised scholarly consensus
   Prefer peer-reviewed sources, university publications, museum databases, and established
   encyclopedias. Add any new facts not already present, with source citations, following
   factbase authoring conventions. Do NOT add speculative or poorly-sourced claims.

4. FACTBASE TOOL BUGS & FEATURE REQUESTS:
   A key purpose of this improvement loop is to make factbase itself better. As you work
   through steps 1-3, pay close attention to how the factbase tools behave. File a Vikunja
   task for EACH of the following you encounter:

   BUGS — any MCP tool error, unexpected result, silent failure, data corruption, or behavior
   that contradicts the tool description. Use title prefix "[factbase]".

   FEATURES — any friction, missing capability, awkward multi-step workaround, or gap that
   slowed you down or limited what you could accomplish. Use title prefix "[factbase][feature]".

   CRITICAL: Before filing ANY bug or feature request, check the EXISTING VIKUNJA BUG REPORTS
   section above. If a similar issue is already filed, do NOT file a duplicate. Only file if
   the issue is genuinely new.

   CRITICAL: Factbase is domain-agnostic — it is used for many different knowledge bases, not
   just this one. Your bug reports and feature requests MUST describe the issue in terms of
   factbase tool behavior, not in terms of ancient history content. For example:
   - GOOD: "apply_review_answers corrupts footnote references when document has >5 citations"
   - BAD: "apply_review_answers broke the Ramesses II document"
   - GOOD: "[feature] check_repository should detect duplicate entities across different folders"
   - BAD: "[feature] factbase should know that Babylon appears in both civilizations/ and cities/"

   Include in every task description:
   - Which MCP tool was called (or which tool is missing/needed)
   - What you expected to happen vs what actually happened
   - Specific parameters or conditions that triggered the issue
   - For features: the general use case and how it would help ANY factbase, not just this one

   curl -s -X PUT "https://vikunja.home.everyonce.com/api/v1/projects/2/tasks" \
     -H "Authorization: Bearer tk_ff251f3d3512775c71913bc2f8ec0dabbf5016a8" \
     -H "Content-Type: application/json" \
     -d '"'"'{"title":"[factbase] <summary>","description":"<detailed description>","priority":3}'"'"'
   Priority: 1-2 for features/minor issues, 3 for normal bugs, 4-5 for data loss or blocking failures.

RULES:
- The document content is already provided above — do NOT call get_entity to re-read it
  unless you just called apply_review_answers and need to verify the result
- Use update_document to edit — be surgical, change only what needs changing
- When calling update_document, do NOT include the <!-- factbase:XXXXXX --> comment or the
  # Title heading in the content — factbase adds those automatically. Start content with
  the first section (e.g. ## Overview). Including them causes duplicate headings.
- If the document has answered review questions (- [x] `@q[...] lines with > answer lines),
  ALWAYS remove them from the content — they are stale artifacts, not part of the document.
- Do NOT run git add, git commit, or git push — the wrapper script handles all git operations
- If nothing needs changing, say so and move on

IMPORTANT: When finished, output exactly one line:
<action_summary>status: STATUS | '"$title"' | changes: DESCRIPTION</action_summary>
Status values: UPDATED (made changes), NO_CHANGE (nothing to do), ERROR (something failed)'

    local output
    output=$(kiro-cli chat --trust-all-tools --no-interactive --model "$MODEL" \
        "$prompt" 2>&1) || {
        log "❌ kiro-cli error for $title, continuing..."
        return 1
    }

    echo "$output"

    local summary
    summary=$(echo "$output" | grep -oP '(?<=<action_summary>).*(?=</action_summary>)' | tail -1)
    [[ -z "$summary" ]] && summary="status: UNKNOWN | $title | no summary returned"
    echo "$summary"
}

# ─── Main entity processing dispatcher ───
process_entity() {
    local fb_id="$1"
    local file="$2"
    local review_count="$3"
    local mtime="$4"
    local last_processed="$5"
    local garbage_count="$6"
    local incomplete_name="$7"

    cd "$DOCS_DIR"

    local title
    title=$(grep '^# ' "$file" | head -1 | sed 's/^# //' | sed 's/ @t\[.*//;s/ \[\^.*//')

    log "━━━ [$title] ($fb_id) reviews=$review_count garbage=$garbage_count ━━━"

    local start_time
    start_time=$(date +%s)
    local status="NO_CHANGE"
    local summary=""

    # Phase 1: Bash cleanup (milliseconds, no agent)
    if bash_cleanup "$file"; then
        log "  🧹 Bash cleanup applied"
        status="UPDATED"
        # Recount after cleanup
        review_count=$(grep -c '^\- \[ \] `@q\[' "$file" 2>/dev/null) || true
        garbage_count=0
    fi

    # Phase 2: Decide if agent is needed
    local needs_agent=true
    if [[ "$SKIP_UNCHANGED" == true && "$last_processed" -gt 0 && "$mtime" -le "$last_processed" ]]; then
        needs_agent=false
        log "  ⏭️  Not modified since last pass → skipping (--skip-unchanged)"
    elif [[ "${incomplete_name:-0}" -eq 1 ]]; then
        log "  👤 Incomplete name (ruler doc) → agent needed to resolve identity"
    elif [[ "$review_count" -gt 0 ]]; then
        log "  📋 $review_count review questions → agent needed"
    elif [[ "$SKIP_UNCHANGED" == true && "$last_processed" -gt 0 && "$mtime" -le "$last_processed" ]]; then
        needs_agent=false
        log "  ⏭️  No questions, not modified since last pass → skipping (--skip-unchanged)"
    else
        log "  🔍 Enrichment + review pass"
    fi

    if [[ "$needs_agent" == true ]]; then
        local agent_output
        agent_output=$(process_entity_agent "$fb_id" "$file" "$title" "$review_count")
        echo "$agent_output"

        summary=$(echo "$agent_output" | grep -oP '(?<=<action_summary>).*(?=</action_summary>)' | tail -1)
        local agent_status
        agent_status=$(echo "$summary" | grep -oP '^status: \K[A-Z_]+' || echo "UNKNOWN")
        if [[ "$agent_status" == "UPDATED" ]]; then
            status="UPDATED"
        fi
    fi

    local end_time
    end_time=$(date +%s)
    local duration=$((end_time - start_time))

    if [[ "$status" == "UPDATED" ]]; then
        do_commit "improve: $title"
    fi

    [[ -z "$summary" ]] && summary="status: $status | $title | bash-only pass"

    {
        echo "[$(date -Iseconds)] $fb_id | $title"
        echo "  $summary"
        echo "  duration: ${duration}s"
    } >> "$ACTION_LOG"

    mark_processed "$fb_id"
    log "  Done (${duration}s) — $status"

    [[ "$status" == "UPDATED" ]] && return 0 || return 1
}

# ═══════════════════════════════════════════
# DEEP CROSS-DOCUMENT SCAN (once per cycle)
# ═══════════════════════════════════════════
run_deep_scan() {
    log "🔬 Running deep cross-document validation scan..."
    local output
    output=$(kiro-cli chat --trust-all-tools --no-interactive --model "$MODEL" \
        "Run check_repository with deep_check=true. Review any new issues found — answer what you can, defer what you cannot. If any factbase tool behaves unexpectedly during this process, file a bug to Vikunja (project 2, Authorization: Bearer tk_ff251f3d3512775c71913bc2f8ec0dabbf5016a8). Describe issues in domain-agnostic terms — factbase is used for many knowledge bases, not just this one." 2>&1) || {
        log "❌ Deep scan agent failed, continuing..."
        return 1
    }
    echo "$output"
    do_commit "deep scan: cross-document validation"
    log "✅ Deep scan complete"
}

# ═══════════════════════════════════════════
# MAIN LOOP
# ═══════════════════════════════════════════
log "🚀 Starting continuous improvement loop (priority=$PRIORITY, model=$MODEL, start=$START_AT, skip_unchanged=$SKIP_UNCHANGED)"
log "Docs dir: $DOCS_DIR"
log "State file: $STATE_FILE"
log "Press Ctrl+C to stop"

PASS=0
while true; do
    PASS=$((PASS + 1))
    log ""
    log "═══════════════════════════════════════════"
    log "  PASS $PASS — $(TZ='America/Chicago' date '+%Y-%m-%d %r') — priority=$PRIORITY"
    log "═══════════════════════════════════════════"

    QUEUE=$(build_queue)
    TOTAL=$(echo "$QUEUE" | grep -c . || echo 0)
    log "Queue: $TOTAL entities (starting at $((START_AT + 1)))"

    PROCESSED=0
    UPDATED=0
    SKIPPED=0
    POSITION=0

    while IFS=$'\t' read -r fb_id file review_count mtime last_processed garbage_count incomplete_name; do
        [[ -z "$fb_id" ]] && continue
        POSITION=$((POSITION + 1))

        if [[ $POSITION -le $START_AT ]]; then
            continue
        fi

        PROCESSED=$((PROCESSED + 1))
        log "[$POSITION/$TOTAL] Next up..."

        if process_entity "$fb_id" "$file" "$review_count" "$mtime" "$last_processed" "$garbage_count" "$incomplete_name"; then
            UPDATED=$((UPDATED + 1))
        fi

        sleep "$CYCLE_DELAY"
    done <<< "$QUEUE"

    log ""
    log "═══ Pass $PASS complete: $PROCESSED processed, $UPDATED updated ═══"

    run_deep_scan

    START_AT=0
    log "Looping back to start..."
    sleep "$CYCLE_DELAY"
done