shipwright-cli 1.7.1 → 1.9.0

package/scripts/sw-intelligence.sh

@@ -0,0 +1,1196 @@
+#!/usr/bin/env bash
+# ╔═══════════════════════════════════════════════════════════════════════════╗
+# ║  shipwright intelligence — AI-Powered Analysis & Decision Engine        ║
+# ║  Semantic issue analysis · Pipeline composition · Cost prediction       ║
+# ╚═══════════════════════════════════════════════════════════════════════════╝
+set -euo pipefail
+trap 'echo "ERROR: $BASH_SOURCE:$LINENO exited with status $?" >&2' ERR
+
+VERSION="1.9.0"
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_DIR="$(cd "$SCRIPT_DIR/.." && pwd)"
+
+# ─── Colors (matches Seth's tmux theme) ─────────────────────────────────────
+CYAN='\033[38;2;0;212;255m'     # #00d4ff — primary accent
+PURPLE='\033[38;2;124;58;237m'  # #7c3aed — secondary
+BLUE='\033[38;2;0;102;255m'     # #0066ff — tertiary
+GREEN='\033[38;2;74;222;128m'   # success
+YELLOW='\033[38;2;250;204;21m'  # warning
+RED='\033[38;2;248;113;113m'    # error
+DIM='\033[2m'
+BOLD='\033[1m'
+RESET='\033[0m'
+
+# ─── Cross-platform compatibility ──────────────────────────────────────────
+# shellcheck source=lib/compat.sh
+[[ -f "$SCRIPT_DIR/lib/compat.sh" ]] && source "$SCRIPT_DIR/lib/compat.sh"
+
+# ─── Output Helpers ─────────────────────────────────────────────────────────
+info()    { echo -e "${CYAN}${BOLD}▸${RESET} $*"; }
+success() { echo -e "${GREEN}${BOLD}✓${RESET} $*"; }
+warn()    { echo -e "${YELLOW}${BOLD}⚠${RESET} $*"; }
+error()   { echo -e "${RED}${BOLD}✗${RESET} $*" >&2; }
+
+now_iso() { date -u +"%Y-%m-%dT%H:%M:%SZ"; }
+now_epoch() { date +%s; }
+
+# ─── Structured Event Log ──────────────────────────────────────────────────
+EVENTS_FILE="${HOME}/.shipwright/events.jsonl"
+
+emit_event() {
+    local event_type="$1"
+    shift
+    local json_fields=""
+    for kv in "$@"; do
+        local key="${kv%%=*}"
+        local val="${kv#*=}"
+        if [[ "$val" =~ ^-?[0-9]+\.?[0-9]*$ ]]; then
+            json_fields="${json_fields},\"${key}\":${val}"
+        else
+            val="${val//\"/\\\"}"
+            json_fields="${json_fields},\"${key}\":\"${val}\""
+        fi
+    done
+    mkdir -p "${HOME}/.shipwright"
+    echo "{\"ts\":\"$(now_iso)\",\"ts_epoch\":$(now_epoch),\"type\":\"${event_type}\"${json_fields}}" >> "$EVENTS_FILE"
+}
+
+# ─── Intelligence Configuration ─────────────────────────────────────────────
+INTELLIGENCE_CACHE="${REPO_DIR}/.claude/intelligence-cache.json"
+INTELLIGENCE_CONFIG_DIR="${HOME}/.shipwright/optimization"
+CACHE_TTL_CONFIG="${INTELLIGENCE_CONFIG_DIR}/cache-ttl.json"
+CACHE_STATS_FILE="${INTELLIGENCE_CONFIG_DIR}/cache-stats.json"
+DEFAULT_CACHE_TTL=3600  # 1 hour (fallback)
+
+# Load adaptive cache TTL from config or use default
+_intelligence_get_cache_ttl() {
+    if [[ -f "$CACHE_TTL_CONFIG" ]]; then
+        local ttl
+        ttl=$(jq -r '.ttl // empty' "$CACHE_TTL_CONFIG" 2>/dev/null || true)
+        if [[ -n "$ttl" && "$ttl" != "null" && "$ttl" -gt 0 ]] 2>/dev/null; then
+            echo "$ttl"
+            return 0
+        fi
+    fi
+    echo "$DEFAULT_CACHE_TTL"
+}
+
+# Track cache hit/miss and adjust TTL
+_intelligence_track_cache_access() {
+    local hit_or_miss="${1:-miss}"  # "hit" or "miss"
+
+    mkdir -p "$INTELLIGENCE_CONFIG_DIR"
+    if [[ ! -f "$CACHE_STATS_FILE" ]]; then
+        echo '{"hits":0,"misses":0,"total":0}' > "$CACHE_STATS_FILE"
+    fi
+
+    local tmp_file
+    tmp_file=$(mktemp "${TMPDIR:-/tmp}/sw-cache-stats.XXXXXX")
+    if [[ "$hit_or_miss" == "hit" ]]; then
+        jq '.hits += 1 | .total += 1' "$CACHE_STATS_FILE" > "$tmp_file" && mv "$tmp_file" "$CACHE_STATS_FILE" || rm -f "$tmp_file"
+    else
+        jq '.misses += 1 | .total += 1' "$CACHE_STATS_FILE" > "$tmp_file" && mv "$tmp_file" "$CACHE_STATS_FILE" || rm -f "$tmp_file"
+    fi
+
+    # Every 20 accesses, evaluate whether to adjust TTL
+    local total
+    total=$(jq '.total // 0' "$CACHE_STATS_FILE" 2>/dev/null || echo "0")
+    if [[ "$((total % 20))" -eq 0 && "$total" -gt 0 ]]; then
+        _intelligence_adjust_cache_ttl
+    fi
+}
+
+# Adjust cache TTL based on hit/miss rates
+_intelligence_adjust_cache_ttl() {
+    [[ -f "$CACHE_STATS_FILE" ]] || return 0
+
+    local hits misses total
+    hits=$(jq '.hits // 0' "$CACHE_STATS_FILE" 2>/dev/null || echo "0")
+    misses=$(jq '.misses // 0' "$CACHE_STATS_FILE" 2>/dev/null || echo "0")
+    total=$(jq '.total // 0' "$CACHE_STATS_FILE" 2>/dev/null || echo "0")
+
+    [[ "$total" -lt 10 ]] && return 0
+
+    local miss_rate
+    miss_rate=$(awk -v m="$misses" -v t="$total" 'BEGIN { printf "%.0f", (m / t) * 100 }')
+
+    local current_ttl
+    current_ttl=$(_intelligence_get_cache_ttl)
+    local new_ttl="$current_ttl"
+
+    if [[ "$miss_rate" -gt 30 ]]; then
+        # High miss rate — reduce TTL (data getting stale too often)
+        new_ttl=$(awk -v ttl="$current_ttl" 'BEGIN { v = int(ttl * 0.75); if (v < 300) v = 300; print v }')
+    elif [[ "$miss_rate" -lt 5 ]]; then
+        # Very low miss rate — increase TTL (cache is very effective)
+        new_ttl=$(awk -v ttl="$current_ttl" 'BEGIN { v = int(ttl * 1.25); if (v > 14400) v = 14400; print v }')
+    fi
+
+    if [[ "$new_ttl" != "$current_ttl" ]]; then
+        local tmp_file
+        tmp_file=$(mktemp "${TMPDIR:-/tmp}/sw-cache-ttl.XXXXXX")
+        jq -n \
+            --argjson ttl "$new_ttl" \
+            --argjson miss_rate "$miss_rate" \
+            --arg updated "$(now_iso)" \
+            '{ttl: $ttl, miss_rate_pct: $miss_rate, updated: $updated}' \
+            > "$tmp_file" && mv "$tmp_file" "$CACHE_TTL_CONFIG" || rm -f "$tmp_file"
+
+        emit_event "intelligence.cache_ttl_adjusted" \
+            "old_ttl=$current_ttl" \
+            "new_ttl=$new_ttl" \
+            "miss_rate=$miss_rate"
+    fi
+
+    # Reset stats for next window
+    local tmp_reset
+    tmp_reset=$(mktemp "${TMPDIR:-/tmp}/sw-cache-reset.XXXXXX")
+    echo '{"hits":0,"misses":0,"total":0}' > "$tmp_reset" && mv "$tmp_reset" "$CACHE_STATS_FILE" || rm -f "$tmp_reset"
+}
+
+# ─── Feature Flag ───────────────────────────────────────────────────────────
+
+_intelligence_enabled() {
+    local config="${REPO_DIR}/.claude/daemon-config.json"
+    if [[ -f "$config" ]]; then
+        local enabled
+        enabled=$(jq -r '.intelligence.enabled // false' "$config" 2>/dev/null || echo "false")
+        [[ "$enabled" == "true" ]]
+    else
+        return 1
+    fi
+}
+
+# MD5 hashing: uses compute_md5 from lib/compat.sh
+
+# ─── Cache Operations ───────────────────────────────────────────────────────
+
+_intelligence_cache_init() {
+    local cache_dir
+    cache_dir="$(dirname "$INTELLIGENCE_CACHE")"
+    mkdir -p "$cache_dir"
+    if [[ ! -f "$INTELLIGENCE_CACHE" ]]; then
+        echo '{"entries":{}}' > "$INTELLIGENCE_CACHE"
+    fi
+}
+
+_intelligence_cache_get() {
+    local cache_key="$1"
+    local adaptive_ttl
+    adaptive_ttl=$(_intelligence_get_cache_ttl)
+    local ttl="${2:-$adaptive_ttl}"
+
+    _intelligence_cache_init
+
+    local hash
+    hash=$(compute_md5 --string "$cache_key")
+
+    local entry
+    entry=$(jq -r --arg h "$hash" '.entries[$h] // empty' "$INTELLIGENCE_CACHE" 2>/dev/null || true)
+
+    if [[ -z "$entry" ]]; then
+        _intelligence_track_cache_access "miss"
+        return 1
+    fi
+
+    local cached_ts
+    cached_ts=$(echo "$entry" | jq -r '.timestamp // 0')
+    local now
+    now=$(now_epoch)
+    local age=$(( now - cached_ts ))
+
+    if [[ "$age" -gt "$ttl" ]]; then
+        _intelligence_track_cache_access "miss"
+        return 1
+    fi
+
+    _intelligence_track_cache_access "hit"
+    echo "$entry" | jq -r '.result'
+    return 0
+}
+
+_intelligence_cache_set() {
+    local cache_key="$1"
+    local result="$2"
+    local ttl="${3:-$DEFAULT_CACHE_TTL}"
+
+    _intelligence_cache_init
+
+    local hash
+    hash=$(compute_md5 --string "$cache_key")
+    local now
+    now=$(now_epoch)
+
+    local tmp_file
+    tmp_file=$(mktemp "${TMPDIR:-/tmp}/sw-intel-cache.XXXXXX")
+    jq --arg h "$hash" \
+       --argjson result "$result" \
+       --argjson ts "$now" \
+       --argjson ttl "$ttl" \
+       '.entries[$h] = {"result": $result, "timestamp": $ts, "ttl": $ttl}' \
+       "$INTELLIGENCE_CACHE" > "$tmp_file" 2>/dev/null && mv "$tmp_file" "$INTELLIGENCE_CACHE" || rm -f "$tmp_file"
+}
+
+# ─── Core Claude Call ────────────────────────────────────────────────────────
+
+_intelligence_call_claude() {
+    local prompt="$1"
+    local cache_key="${2:-}"
+    local ttl="${3:-$DEFAULT_CACHE_TTL}"
+
+    # Check cache first
+    local cached
+    if cached=$(_intelligence_cache_get "$cache_key" "$ttl"); then
+        emit_event "intelligence.cache_hit" "cache_key=$cache_key"
+        echo "$cached"
+        return 0
+    fi
+
+    # Verify claude CLI is available
+    if ! command -v claude >/dev/null 2>&1; then
+        error "claude CLI not found"
+        echo '{"error":"claude_cli_not_found"}'
+        return 1
+    fi
+
+    # Call Claude (--print mode returns raw text response)
+    local response
+    if ! response=$(claude -p "$prompt" 2>/dev/null); then
+        error "Claude call failed"
+        echo '{"error":"claude_call_failed"}'
+        return 1
+    fi
+
+    # Extract JSON from the response
+    local result
+    # First try: raw response is valid JSON directly
+    if echo "$response" | jq '.' >/dev/null 2>&1; then
+        result=$(echo "$response" | jq '.')
+    else
+        # Second try: extract JSON from markdown code blocks (```json ... ```)
+        local extracted
+        extracted=$(echo "$response" | sed -n '/^```/,/^```$/p' | sed '1d;$d' || true)
+        if [[ -n "$extracted" ]] && echo "$extracted" | jq '.' >/dev/null 2>&1; then
+            result="$extracted"
+        else
+            # Third try: find first { to last } in response
+            local braced
+            braced=$(echo "$response" | sed -n '/{/,/}/p' || true)
+            if [[ -n "$braced" ]] && echo "$braced" | jq '.' >/dev/null 2>&1; then
+                result="$braced"
+            else
+                # Wrap raw text in a JSON object
+                result=$(jq -n --arg text "$response" '{"raw_response": $text}')
+            fi
+        fi
+    fi
+
+    # Cache the result
+    _intelligence_cache_set "$cache_key" "$result" "$ttl"
+
+    echo "$result"
+    return 0
+}
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# PUBLIC FUNCTIONS
+# ═══════════════════════════════════════════════════════════════════════════════
+
+# ─── Analyze Issue ───────────────────────────────────────────────────────────
+
+intelligence_analyze_issue() {
+    local issue_json="${1:-"{}"}"
+
+    if ! _intelligence_enabled; then
+        echo '{"error":"intelligence_disabled","complexity":5,"risk_level":"medium","success_probability":50,"recommended_template":"standard","key_risks":[],"implementation_hints":[]}'
+        return 0
+    fi
+
+    local title body labels
+    title=$(echo "$issue_json" | jq -r '.title // "untitled"' 2>/dev/null || echo "untitled")
+    body=$(echo "$issue_json" | jq -r '.body // ""' 2>/dev/null || echo "")
+    labels=$(echo "$issue_json" | jq -r '(.labels // []) | join(", ")' 2>/dev/null || echo "")
+
+    local prompt
+    prompt="Analyze this GitHub issue for a software project and return ONLY a JSON object (no markdown, no explanation).
+
+Issue title: ${title}
+Issue body: ${body}
+Labels: ${labels}
+
+Return JSON with exactly these fields:
+{
+  \"complexity\": <number 1-10>,
+  \"risk_level\": \"<low|medium|high|critical>\",
+  \"success_probability\": <number 0-100>,
+  \"recommended_template\": \"<fast|standard|full|hotfix|autonomous|enterprise|cost-aware>\",
+  \"key_risks\": [\"risk1\", \"risk2\"],
+  \"implementation_hints\": [\"hint1\", \"hint2\"]
+}"
+
+    local cache_key
+    cache_key="analyze_issue_$(compute_md5 --string "${title}${body}")"
+
+    local result
+    if result=$(_intelligence_call_claude "$prompt" "$cache_key"); then
+        # Validate schema: ensure required fields exist
+        local valid
+        valid=$(echo "$result" | jq 'has("complexity") and has("risk_level") and has("success_probability") and has("recommended_template")' 2>/dev/null || echo "false")
+
+        if [[ "$valid" != "true" ]]; then
+            warn "Intelligence response missing required fields, using fallback"
+            result='{"complexity":5,"risk_level":"medium","success_probability":50,"recommended_template":"standard","key_risks":["analysis_incomplete"],"implementation_hints":[]}'
+        fi
+
+        emit_event "intelligence.analysis" \
+            "complexity=$(echo "$result" | jq -r '.complexity')" \
+            "risk_level=$(echo "$result" | jq -r '.risk_level')" \
+            "success_probability=$(echo "$result" | jq -r '.success_probability')" \
+            "recommended_template=$(echo "$result" | jq -r '.recommended_template')"
+
+        # Enrich with GitHub data if available
+        result=$(intelligence_github_enrich "$result")
+
+        echo "$result"
+        return 0
+    else
+        echo '{"error":"analysis_failed","complexity":5,"risk_level":"medium","success_probability":50,"recommended_template":"standard","key_risks":["analysis_failed"],"implementation_hints":[]}'
+        return 0
+    fi
+}
+
+# ─── Compose Pipeline ───────────────────────────────────────────────────────
+
+intelligence_compose_pipeline() {
+    local issue_analysis="${1:-"{}"}"
+    local repo_context="${2:-"{}"}"
+    local budget="${3:-0}"
+
+    if ! _intelligence_enabled; then
+        echo '{"error":"intelligence_disabled","stages":[]}'
+        return 0
+    fi
+
+    local complexity risk_level
+    complexity=$(echo "$issue_analysis" | jq -r '.complexity // 5' 2>/dev/null || echo "5")
+    risk_level=$(echo "$issue_analysis" | jq -r '.risk_level // "medium"' 2>/dev/null || echo "medium")
+
+    local prompt
+    prompt="You are a CI/CD pipeline composer. Given the issue analysis and constraints, compose an optimal pipeline.
+
+Issue complexity: ${complexity}/10
+Risk level: ${risk_level}
+Budget remaining: \$${budget} USD
+Repository context: ${repo_context}
+
+Available stages: intake, plan, design, build, test, review, compound_quality, pr, merge, deploy, validate, monitor
+Available models: opus (most capable, expensive), sonnet (balanced), haiku (fast, cheap)
+
+Return ONLY a JSON object (no markdown):
+{
+  \"stages\": [
+    {\"id\": \"stage_name\", \"enabled\": true, \"model\": \"sonnet\", \"config\": {}}
+  ],
+  \"rationale\": \"brief explanation\"
+}"
+
+    local cache_key
+    cache_key="compose_pipeline_$(compute_md5 --string "${issue_analysis}${budget}")"
+
+    local result
+    if result=$(_intelligence_call_claude "$prompt" "$cache_key"); then
+        local has_stages
+        has_stages=$(echo "$result" | jq 'has("stages") and (.stages | type == "array")' 2>/dev/null || echo "false")
+
+        if [[ "$has_stages" != "true" ]]; then
+            warn "Pipeline composition missing stages array, using fallback"
+            result='{"stages":[{"id":"intake","enabled":true,"model":"sonnet","config":{}},{"id":"build","enabled":true,"model":"sonnet","config":{}},{"id":"test","enabled":true,"model":"sonnet","config":{}},{"id":"pr","enabled":true,"model":"sonnet","config":{}}],"rationale":"fallback pipeline"}'
+        fi
+
+        emit_event "intelligence.compose" \
+            "stage_count=$(echo "$result" | jq '.stages | length')" \
+            "complexity=$complexity"
+
+        echo "$result"
+        return 0
+    else
+        echo '{"error":"composition_failed","stages":[]}'
+        return 0
+    fi
+}
+
+# ─── Predict Cost ────────────────────────────────────────────────────────────
+
+intelligence_predict_cost() {
+    local issue_analysis="${1:-"{}"}"
+    local historical_data="${2:-"{}"}"
+
+    if ! _intelligence_enabled; then
+        echo '{"error":"intelligence_disabled","estimated_cost_usd":0,"estimated_iterations":0,"likely_failure_stage":"unknown"}'
+        return 0
+    fi
+
+    local complexity
+    complexity=$(echo "$issue_analysis" | jq -r '.complexity // 5' 2>/dev/null || echo "5")
+
+    local prompt
+    prompt="Estimate the cost and effort for a CI pipeline run. Return ONLY JSON (no markdown).
+
+Issue analysis: ${issue_analysis}
+Historical data from past runs: ${historical_data}
+
+Based on similar complexity (${complexity}/10) issues, estimate:
+{
+  \"estimated_cost_usd\": <number>,
+  \"estimated_iterations\": <number of build-test cycles>,
+  \"estimated_tokens\": <total token count>,
+  \"likely_failure_stage\": \"<stage name or 'none'>\",
+  \"confidence\": <0-100>
+}"
+
+    local cache_key
+    cache_key="predict_cost_$(compute_md5 --string "${issue_analysis}${historical_data}")"
+
+    local result
+    if result=$(_intelligence_call_claude "$prompt" "$cache_key"); then
+        local valid
+        valid=$(echo "$result" | jq 'has("estimated_cost_usd") and has("estimated_iterations")' 2>/dev/null || echo "false")
+
+        if [[ "$valid" != "true" ]]; then
+            warn "Cost prediction missing required fields, using fallback"
+            result='{"estimated_cost_usd":5.0,"estimated_iterations":3,"estimated_tokens":500000,"likely_failure_stage":"test","confidence":30}'
+        fi
+
+        emit_event "intelligence.prediction" \
+            "estimated_cost=$(echo "$result" | jq -r '.estimated_cost_usd')" \
+            "estimated_iterations=$(echo "$result" | jq -r '.estimated_iterations')" \
+            "complexity=$complexity"
+
+        echo "$result"
+        return 0
+    else
+        echo '{"error":"prediction_failed","estimated_cost_usd":0,"estimated_iterations":0,"likely_failure_stage":"unknown"}'
+        return 0
+    fi
+}
+
+# ─── Synthesize Findings ─────────────────────────────────────────────────────
+
+intelligence_synthesize_findings() {
+    local findings_json="${1:-"[]"}"
+
+    if ! _intelligence_enabled; then
+        echo '{"error":"intelligence_disabled","priority_fixes":[],"root_causes":[],"recommended_approach":""}'
+        return 0
+    fi
+
+    local prompt
+    prompt="Synthesize multiple signal sources (patrol findings, test results, review comments) into a unified fix strategy. Return ONLY JSON (no markdown).
+
+Findings: ${findings_json}
+
+Return:
+{
+  \"priority_fixes\": [\"fix1\", \"fix2\"],
+  \"root_causes\": [\"cause1\", \"cause2\"],
+  \"recommended_approach\": \"description of unified strategy\",
+  \"estimated_effort\": \"<low|medium|high>\"
+}"
+
+    local cache_key
+    cache_key="synthesize_$(compute_md5 --string "$findings_json")"
+
+    local result
+    if result=$(_intelligence_call_claude "$prompt" "$cache_key"); then
+        local valid
+        valid=$(echo "$result" | jq 'has("priority_fixes") and has("root_causes")' 2>/dev/null || echo "false")
+
+        if [[ "$valid" != "true" ]]; then
+            result='{"priority_fixes":[],"root_causes":["analysis_incomplete"],"recommended_approach":"manual review needed","estimated_effort":"medium"}'
+        fi
+
+        emit_event "intelligence.synthesize" \
+            "fix_count=$(echo "$result" | jq '.priority_fixes | length')" \
+            "cause_count=$(echo "$result" | jq '.root_causes | length')"
+
+        echo "$result"
+        return 0
+    else
+        echo '{"error":"synthesis_failed","priority_fixes":[],"root_causes":[],"recommended_approach":""}'
+        return 0
+    fi
+}
+
+# ─── Search Memory ───────────────────────────────────────────────────────────
+
+intelligence_search_memory() {
+    local context="${1:-""}"
+    local memory_dir="${2:-"${HOME}/.shipwright/memory"}"
+    local top_n="${3:-5}"
+
+    if ! _intelligence_enabled; then
+        echo '{"error":"intelligence_disabled","results":[]}'
+        return 0
+    fi
+
+    # Gather memory file contents
+    local memory_content=""
+    if [[ -d "$memory_dir" ]]; then
+        local file_list=""
+        local count=0
+        while IFS= read -r f; do
+            [[ -z "$f" ]] && continue
+            local fname
+            fname=$(basename "$f")
+            local content
+            content=$(head -100 "$f" 2>/dev/null || true)
+            if [[ -n "$content" ]]; then
+                file_list="${file_list}
+--- ${fname} ---
+${content}
+"
+                count=$((count + 1))
+            fi
+        done < <(find "$memory_dir" -name "*.json" -o -name "*.md" 2>/dev/null | head -20 || true)
+
+        if [[ "$count" -eq 0 ]]; then
+            echo '{"results":[],"message":"no memory files found"}'
+            return 0
+        fi
+        memory_content="$file_list"
+    else
+        echo '{"results":[],"message":"memory directory not found"}'
+        return 0
+    fi
+
+    local prompt
+    prompt="Rank the following memory entries by relevance to this context. Return ONLY JSON (no markdown).
+
+Context: ${context}
+
+Memory entries:
+${memory_content}
+
+Return the top ${top_n} most relevant entries:
+{
+  \"results\": [
+    {\"file\": \"filename\", \"relevance\": <0-100>, \"summary\": \"why this is relevant\"}
+  ]
+}"
+
+    local cache_key
+    cache_key="search_memory_$(compute_md5 --string "${context}${memory_dir}")"
+
+    local result
+    if result=$(_intelligence_call_claude "$prompt" "$cache_key" 1800); then
+        local valid
+        valid=$(echo "$result" | jq 'has("results") and (.results | type == "array")' 2>/dev/null || echo "false")
+
+        if [[ "$valid" != "true" ]]; then
+            result='{"results":[]}'
+        fi
+
+        echo "$result"
+        return 0
+    else
+        echo '{"error":"memory_search_failed","results":[]}'
+        return 0
+    fi
+}
+
+# ─── Estimate Iterations ─────────────────────────────────────────────────────
+
+intelligence_estimate_iterations() {
+    local issue_analysis="${1:-"{}"}"
+    local historical_data="${2:-""}"
+
+    local iteration_model="${HOME}/.shipwright/optimization/iteration-model.json"
+
+    # Extract complexity from issue analysis (numeric 1-10)
+    local complexity
+    complexity=$(echo "$issue_analysis" | jq -r '.complexity // 5' 2>/dev/null || echo "5")
+
+    # Map numeric complexity to bucket: 1-3=low, 4-6=medium, 7-10=high
+    local bucket
+    if [[ "$complexity" -le 3 ]]; then
+        bucket="low"
+    elif [[ "$complexity" -le 6 ]]; then
+        bucket="medium"
+    else
+        bucket="high"
+    fi
+
+    # Strategy 1: Intelligence-enabled — call Claude for fine-grained estimation
+    if _intelligence_enabled; then
+        local prompt
+        prompt="Estimate the number of build-test iterations needed for this issue. Return ONLY a JSON object.
+
+Issue analysis: ${issue_analysis}
+Historical iteration data: ${historical_data:-none}
+
+Consider:
+- Complexity: ${complexity}/10 (${bucket})
+- Higher complexity usually needs more iterations
+- Well-understood patterns need fewer iterations
+
+Return JSON: {\"estimated_iterations\": <integer 1-50>}"
+
+        local cache_key
+        cache_key="estimate_iterations_$(compute_md5 --string "${issue_analysis}")"
+
+        local result
+        if result=$(_intelligence_call_claude "$prompt" "$cache_key" 1800); then
+            # Extract number from result
+            local estimate
+            estimate=$(echo "$result" | jq -r '.estimated_iterations // .iterations // .estimate // empty' 2>/dev/null || true)
+            if [[ -z "$estimate" ]]; then
+                # Try raw number from response
+                estimate=$(echo "$result" | jq -r '.raw_response // empty' 2>/dev/null | tr -dc '0-9' | head -c 3 || true)
+            fi
+            if [[ -n "$estimate" ]] && [[ "$estimate" =~ ^[0-9]+$ ]] && \
+               [[ "$estimate" -ge 1 ]] && [[ "$estimate" -le 50 ]]; then
+                emit_event "intelligence.estimate_iterations" \
+                    "estimate=$estimate" \
+                    "complexity=$complexity" \
+                    "source=claude"
+                echo "$estimate"
+                return 0
+            fi
+        fi
+        # Fall through to historical data if Claude call fails
+    fi
+
+    # Strategy 2: Use historical averages from iteration model
+    if [[ -f "$iteration_model" ]]; then
+        local mean samples
+        mean=$(jq -r --arg b "$bucket" '.[$b].mean // 0' "$iteration_model" 2>/dev/null || echo "0")
+        samples=$(jq -r --arg b "$bucket" '.[$b].samples // 0' "$iteration_model" 2>/dev/null || echo "0")
+
+        if [[ "$samples" -gt 0 ]] && [[ "$mean" != "0" ]]; then
+            # Round to nearest integer, clamp 1-50
+            local estimate
+            estimate=$(awk "BEGIN{v=int($mean + 0.5); if(v<1)v=1; if(v>50)v=50; print v}")
+            emit_event "intelligence.estimate_iterations" \
+                "estimate=$estimate" \
+                "complexity=$complexity" \
+                "source=historical" \
+                "samples=$samples"
+            echo "$estimate"
+            return 0
+        fi
+    fi
+
+    # Strategy 3: Heuristic fallback based on numeric complexity
+    local estimate
+    if [[ "$complexity" -le 2 ]]; then
+        estimate=5
+    elif [[ "$complexity" -le 4 ]]; then
+        estimate=10
+    elif [[ "$complexity" -le 6 ]]; then
+        estimate=15
+    elif [[ "$complexity" -le 8 ]]; then
+        estimate=25
+    else
+        estimate=35
+    fi
+
+    emit_event "intelligence.estimate_iterations" \
+        "estimate=$estimate" \
+        "complexity=$complexity" \
+        "source=heuristic"
+
+    echo "$estimate"
+    return 0
+}
+
+# ─── Recommend Model ─────────────────────────────────────────────────────────
+
+intelligence_recommend_model() {
+    local stage="${1:-"build"}"
+    local complexity="${2:-5}"
+    local budget_remaining="${3:-100}"
+
+    local model="sonnet"
+    local reason="default balanced choice"
+
+    # Strategy 1: Check historical model routing data
+    local routing_file="${HOME}/.shipwright/optimization/model-routing.json"
+    if [[ -f "$routing_file" ]]; then
+        local stage_data
+        stage_data=$(jq -r --arg s "$stage" '.[$s] // empty' "$routing_file" 2>/dev/null || true)
+
+        if [[ -n "$stage_data" && "$stage_data" != "null" ]]; then
+            local recommended sonnet_rate sonnet_samples opus_rate opus_samples
+            recommended=$(echo "$stage_data" | jq -r '.recommended // empty' 2>/dev/null || true)
+            sonnet_rate=$(echo "$stage_data" | jq -r '.sonnet_rate // 0' 2>/dev/null || echo "0")
+            sonnet_samples=$(echo "$stage_data" | jq -r '.sonnet_samples // 0' 2>/dev/null || echo "0")
+            opus_rate=$(echo "$stage_data" | jq -r '.opus_rate // 0' 2>/dev/null || echo "0")
+            opus_samples=$(echo "$stage_data" | jq -r '.opus_samples // 0' 2>/dev/null || echo "0")
+
+            # Load adaptive routing thresholds from config or use defaults
+            local routing_config="${HOME}/.shipwright/optimization/model-routing-thresholds.json"
+            local min_samples=3
+            local success_threshold=90
+            local complexity_upgrade=8
+
+            if [[ -f "$routing_config" ]]; then
+                local cfg_min cfg_success cfg_complexity
+                cfg_min=$(jq -r '.min_samples // empty' "$routing_config" 2>/dev/null || true)
+                cfg_success=$(jq -r '.success_threshold // empty' "$routing_config" 2>/dev/null || true)
+                cfg_complexity=$(jq -r '.complexity_upgrade // empty' "$routing_config" 2>/dev/null || true)
+                [[ -n "$cfg_min" && "$cfg_min" != "null" ]] && min_samples="$cfg_min"
+                [[ -n "$cfg_success" && "$cfg_success" != "null" ]] && success_threshold="$cfg_success"
+                [[ -n "$cfg_complexity" && "$cfg_complexity" != "null" ]] && complexity_upgrade="$cfg_complexity"
+            fi
+
+            # SPRT-inspired evidence check: if enough data, use log-likelihood ratio
+            local use_sonnet=false
+            local total_samples=$((sonnet_samples + opus_samples))
+
+            if [[ "$total_samples" -ge 10 ]]; then
+                # With 10+ data points, use evidence ratio
+                # Log-likelihood ratio: ln(P(data|sonnet_good) / P(data|sonnet_bad))
+                # Simplified: if sonnet_rate / opus_rate > 0.95, switch to sonnet
+                local rate_ratio
+                if awk -v or="$opus_rate" 'BEGIN { exit !(or > 0) }' 2>/dev/null; then
+                    rate_ratio=$(awk -v sr="$sonnet_rate" -v or="$opus_rate" 'BEGIN { printf "%.3f", sr / or }')
+                else
+                    rate_ratio="1.0"
+                fi
+
+                if [[ "$sonnet_samples" -ge "$min_samples" ]] && \
+                   awk -v sr="$sonnet_rate" -v st="$success_threshold" 'BEGIN { exit !(sr >= st) }' 2>/dev/null && \
+                   awk -v rr="$rate_ratio" 'BEGIN { exit !(rr >= 0.95) }' 2>/dev/null; then
+                    use_sonnet=true
+                fi
+            elif [[ "$sonnet_samples" -ge "$min_samples" ]] && \
+                 awk -v sr="$sonnet_rate" -v st="$success_threshold" 'BEGIN { exit !(sr >= st) }' 2>/dev/null; then
+                # Fewer data points — fall back to simple threshold check
+                use_sonnet=true
+            fi
+
+            if [[ "$use_sonnet" == "true" ]]; then
+                if [[ "$budget_remaining" != "" ]] && [[ "$(echo "$budget_remaining < 5" | bc 2>/dev/null || echo "0")" == "1" ]]; then
+                    model="haiku"
+                    reason="sonnet viable (${sonnet_rate}% success) but budget constrained"
+                else
+                    model="sonnet"
+                    reason="evidence-based: ${sonnet_rate}% success on ${stage} (${sonnet_samples} samples, SPRT)"
+                fi
+            elif [[ -n "$recommended" && "$recommended" != "null" ]]; then
+                model="$recommended"
+                reason="historical routing recommendation for ${stage}"
+            fi
+
+            # Override: high complexity + critical stage → upgrade to opus if budget allows
+            if [[ "$complexity" -ge "$complexity_upgrade" ]]; then
+                case "$stage" in
+                    plan|design|review|compound_quality)
+                        if [[ "$model" != "opus" ]]; then
+                            if [[ "$budget_remaining" == "" ]] || [[ "$(echo "$budget_remaining >= 10" | bc 2>/dev/null || echo "1")" == "1" ]]; then
+                                model="opus"
+                                reason="high complexity (${complexity}/10) overrides historical for critical stage (${stage})"
+                            fi
+                        fi
+                        ;;
+                esac
+            fi
+
+            local result
+            result=$(jq -n --arg model "$model" --arg reason "$reason" --arg stage "$stage" --argjson complexity "$complexity" \
+                '{"model": $model, "reason": $reason, "stage": $stage, "complexity": $complexity, "source": "historical"}')
+
+            emit_event "intelligence.model" \
+                "stage=$stage" \
+                "complexity=$complexity" \
+                "model=$model" \
+                "source=historical"
+
+            echo "$result"
+            return 0
+        fi
+    fi
+
+    # Strategy 2: Heuristic fallback (no historical data available)
+    # Budget-constrained: use haiku
+    if [[ "$budget_remaining" != "" ]] && [[ "$(echo "$budget_remaining < 5" | bc 2>/dev/null || echo "0")" == "1" ]]; then
+        model="haiku"
+        reason="budget constrained (< \$5 remaining)"
+    # High complexity + critical stages: use opus
+    elif [[ "$complexity" -ge 8 ]]; then
+        case "$stage" in
+            plan|design|review|compound_quality)
+                model="opus"
+                reason="high complexity (${complexity}/10) + critical stage (${stage})"
+                ;;
+            build|test)
+                model="sonnet"
+                reason="high complexity but execution stage — sonnet is sufficient"
+                ;;
+            *)
+                model="sonnet"
+                reason="high complexity, non-critical stage"
+                ;;
+        esac
+    # Low complexity: use haiku for simple stages
+    elif [[ "$complexity" -le 3 ]]; then
+        case "$stage" in
+            intake|pr|merge)
+                model="haiku"
+                reason="low complexity (${complexity}/10), simple stage (${stage})"
+                ;;
+            build|test)
+                model="sonnet"
+                reason="low complexity but code execution stage"
+                ;;
+            *)
+                model="haiku"
+                reason="low complexity, standard stage"
+                ;;
+        esac
+    fi
+
+    local result
+    result=$(jq -n --arg model "$model" --arg reason "$reason" --arg stage "$stage" --argjson complexity "$complexity" \
+        '{"model": $model, "reason": $reason, "stage": $stage, "complexity": $complexity, "source": "heuristic"}')
+
+    emit_event "intelligence.model" \
+        "stage=$stage" \
+        "complexity=$complexity" \
+        "model=$model" \
+        "source=heuristic"
+
+    echo "$result"
+    return 0
+}
+
+# ─── Prediction Validation ─────────────────────────────────────────────────
+
+# intelligence_validate_prediction <issue_id> <predicted_complexity> <actual_iterations> <actual_success>
+# Compares predicted complexity to actual outcome for feedback learning.
+intelligence_validate_prediction() {
+    local issue_id="${1:-}"
+    local predicted_complexity="${2:-0}"
+    local actual_iterations="${3:-0}"
+    local actual_success="${4:-false}"
+
+    if [[ -z "$issue_id" ]]; then
+        error "Usage: intelligence_validate_prediction <issue_id> <predicted> <actual_iterations> <actual_success>"
+        return 1
+    fi
+
+    # Infer actual complexity from iterations (heuristic: map iterations to 1-10 scale)
+    local actual_complexity
+    if [[ "$actual_iterations" -le 5 ]]; then
+        actual_complexity=2
+    elif [[ "$actual_iterations" -le 10 ]]; then
+        actual_complexity=4
+    elif [[ "$actual_iterations" -le 15 ]]; then
+        actual_complexity=6
+    elif [[ "$actual_iterations" -le 25 ]]; then
+        actual_complexity=8
+    else
+        actual_complexity=10
+    fi
+
+    # Calculate prediction error (signed delta)
+    local delta=$(( predicted_complexity - actual_complexity ))
+    local abs_delta="${delta#-}"
+
+    # Emit prediction error event
+    emit_event "intelligence.prediction_error" \
+        "issue_id=$issue_id" \
+        "predicted=$predicted_complexity" \
+        "actual=$actual_complexity" \
+        "delta=$delta" \
+        "actual_iterations=$actual_iterations" \
+        "actual_success=$actual_success"
+
+    # Warn if prediction was significantly off
+    if [[ "$abs_delta" -gt 3 ]]; then
+        warn "Prediction error for issue #${issue_id}: predicted complexity=${predicted_complexity}, actual~=${actual_complexity} (delta=${delta})"
+    fi
+
+    # Update cache entry with actual outcome for future learning
+    _intelligence_cache_init
+
+    local validation_file="${HOME}/.shipwright/optimization/prediction-validation.jsonl"
+    mkdir -p "${HOME}/.shipwright/optimization"
+
+    local record
+    record=$(jq -c -n \
+        --arg ts "$(now_iso)" \
+        --arg issue "$issue_id" \
+        --argjson predicted "$predicted_complexity" \
+        --argjson actual "$actual_complexity" \
+        --argjson delta "$delta" \
+        --argjson iterations "$actual_iterations" \
+        --arg success "$actual_success" \
+        '{
+            ts: $ts,
+            issue: $issue,
+            predicted_complexity: $predicted,
+            actual_complexity: $actual,
+            delta: $delta,
+            actual_iterations: $iterations,
+            actual_success: $success
+        }')
+
+    echo "$record" >> "$validation_file"
+
+    # Output summary
+    local accuracy_label="good"
+    if [[ "$abs_delta" -gt 3 ]]; then
+        accuracy_label="poor"
+    elif [[ "$abs_delta" -gt 1 ]]; then
+        accuracy_label="fair"
+    fi
+
+    jq -n \
+        --arg issue "$issue_id" \
+        --argjson predicted "$predicted_complexity" \
+        --argjson actual "$actual_complexity" \
+        --argjson delta "$delta" \
+        --arg accuracy "$accuracy_label" \
+        --arg success "$actual_success" \
+        '{
+            issue: $issue,
+            predicted_complexity: $predicted,
+            actual_complexity: $actual,
+            delta: $delta,
+            accuracy: $accuracy,
+            actual_success: $success
+        }'
+}
+
+# ─── GitHub Enrichment ─────────────────────────────────────────────────────
+
+intelligence_github_enrich() {
+    local analysis_json="$1"
+
+    # Skip if GraphQL not available
+    type _gh_detect_repo &>/dev/null 2>&1 || { echo "$analysis_json"; return 0; }
+    _gh_detect_repo 2>/dev/null || { echo "$analysis_json"; return 0; }
+
+    local owner="${GH_OWNER:-}" repo="${GH_REPO:-}"
+    [[ -z "$owner" || -z "$repo" ]] && { echo "$analysis_json"; return 0; }
+
+    # Get repo context
+    local repo_context="{}"
+    if type gh_repo_context &>/dev/null 2>&1; then
+        repo_context=$(gh_repo_context "$owner" "$repo" 2>/dev/null || echo "{}")
+    fi
+
+    # Get security alerts count
+    local security_count=0
+    if type gh_security_alerts &>/dev/null 2>&1; then
+        local alerts
+        alerts=$(gh_security_alerts "$owner" "$repo" 2>/dev/null || echo "[]")
+        security_count=$(echo "$alerts" | jq 'length' 2>/dev/null || echo "0")
+    fi
+
+    # Get dependabot alerts count
+    local dependabot_count=0
+    if type gh_dependabot_alerts &>/dev/null 2>&1; then
+        local deps
+        deps=$(gh_dependabot_alerts "$owner" "$repo" 2>/dev/null || echo "[]")
+        dependabot_count=$(echo "$deps" | jq 'length' 2>/dev/null || echo "0")
+    fi
+
+    # Merge GitHub context into analysis
+    echo "$analysis_json" | jq --arg ctx "$repo_context" \
+        --argjson sec "${security_count:-0}" \
+        --argjson dep "${dependabot_count:-0}" \
+        '. + {
+            github_context: ($ctx | fromjson? // {}),
+            security_alert_count: $sec,
+            dependabot_alert_count: $dep
+        }' 2>/dev/null || echo "$analysis_json"
+}
+
+intelligence_file_risk_score() {
+    local file_path="$1"
+    local risk_score=0
+
+    type _gh_detect_repo &>/dev/null 2>&1 || { echo "0"; return 0; }
+    _gh_detect_repo 2>/dev/null || { echo "0"; return 0; }
+
+    local owner="${GH_OWNER:-}" repo="${GH_REPO:-}"
+    [[ -z "$owner" || -z "$repo" ]] && { echo "0"; return 0; }
+
+    # Factor 1: File churn (high change frequency = higher risk)
+    local changes=0
+    if type gh_file_change_frequency &>/dev/null 2>&1; then
+        changes=$(gh_file_change_frequency "$owner" "$repo" "$file_path" 30 2>/dev/null || echo "0")
+    fi
+    if [[ "${changes:-0}" -gt 20 ]]; then
+        risk_score=$((risk_score + 30))
+    elif [[ "${changes:-0}" -gt 10 ]]; then
+        risk_score=$((risk_score + 15))
+    elif [[ "${changes:-0}" -gt 5 ]]; then
+        risk_score=$((risk_score + 5))
+    fi
+
+    # Factor 2: Security alerts on this file
+    if type gh_security_alerts &>/dev/null 2>&1; then
+        local file_alerts
+        file_alerts=$(gh_security_alerts "$owner" "$repo" 2>/dev/null | \
+            jq --arg path "$file_path" '[.[] | select(.most_recent_instance.location.path == $path)] | length' 2>/dev/null || echo "0")
+        [[ "${file_alerts:-0}" -gt 0 ]] && risk_score=$((risk_score + 40))
+    fi
+
+    # Factor 3: Many contributors = higher coordination risk
+    if type gh_blame_data &>/dev/null 2>&1; then
+        local author_count
+        author_count=$(gh_blame_data "$owner" "$repo" "$file_path" 2>/dev/null | jq 'length' 2>/dev/null || echo "0")
+        [[ "${author_count:-0}" -gt 5 ]] && risk_score=$((risk_score + 10))
+    fi
+
+    # Cap at 100
+    [[ "$risk_score" -gt 100 ]] && risk_score=100
+    echo "$risk_score"
+}
+
+intelligence_contributor_expertise() {
+    local file_path="$1"
+
+    type _gh_detect_repo &>/dev/null 2>&1 || { echo "[]"; return 0; }
+    _gh_detect_repo 2>/dev/null || { echo "[]"; return 0; }
+
+    local owner="${GH_OWNER:-}" repo="${GH_REPO:-}"
+    [[ -z "$owner" || -z "$repo" ]] && { echo "[]"; return 0; }
+
+    if type gh_blame_data &>/dev/null 2>&1; then
+        gh_blame_data "$owner" "$repo" "$file_path" 2>/dev/null || echo "[]"
+    else
+        echo "[]"
+    fi
+}
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# CLI INTERFACE
+# ═══════════════════════════════════════════════════════════════════════════════
+
+show_help() {
+    echo ""
+    echo -e "${CYAN}${BOLD}shipwright intelligence${RESET} — AI-Powered Analysis & Decision Engine"
+    echo ""
+    echo -e "${BOLD}USAGE${RESET}"
+    echo -e "  shipwright intelligence <command> [options]"
+    echo ""
+    echo -e "${BOLD}COMMANDS${RESET}"
+    echo -e "  ${CYAN}analyze${RESET} <issue_json>          Analyze an issue semantically"
+    echo -e "  ${CYAN}compose${RESET} <analysis> [ctx] [$$]  Compose optimal pipeline"
+    echo -e "  ${CYAN}predict${RESET} <analysis> [history]   Predict cost and effort"
+    echo -e "  ${CYAN}synthesize${RESET} <findings_json>     Synthesize findings into strategy"
+    echo -e "  ${CYAN}search-memory${RESET} <context> [dir]  Search memory by relevance"
+    echo -e "  ${CYAN}estimate-iterations${RESET} <analysis>  Estimate build iterations"
+    echo -e "  ${CYAN}recommend-model${RESET} <stage> [cplx] Recommend model for stage"
+    echo -e "  ${CYAN}cache-stats${RESET}                    Show cache statistics"
+    echo -e "  ${CYAN}validate-prediction${RESET} <id> <pred> <iters> <success>  Validate prediction accuracy"
+    echo -e "  ${CYAN}cache-clear${RESET}                    Clear intelligence cache"
+    echo -e "  ${CYAN}help${RESET}                           Show this help"
+    echo ""
+    echo -e "${BOLD}CONFIGURATION${RESET}"
+    echo -e "  Enable in ${DIM}.claude/daemon-config.json${RESET}:"
+    echo -e "    ${DIM}{\"intelligence\": {\"enabled\": true}}${RESET}"
+    echo ""
+    echo -e "${DIM}Version ${VERSION}${RESET}"
+}
+
+cmd_cache_stats() {
+    _intelligence_cache_init
+
+    local entry_count
+    entry_count=$(jq '.entries | length' "$INTELLIGENCE_CACHE" 2>/dev/null || echo "0")
+    local cache_size
+    cache_size=$(wc -c < "$INTELLIGENCE_CACHE" 2>/dev/null | tr -d ' ' || echo "0")
+
+    echo ""
+    echo -e "${BOLD}Intelligence Cache${RESET}"
+    echo -e "  Entries:  ${CYAN}${entry_count}${RESET}"
+    echo -e "  Size:     ${DIM}${cache_size} bytes${RESET}"
+    echo -e "  Location: ${DIM}${INTELLIGENCE_CACHE}${RESET}"
+
+    if [[ "$entry_count" -gt 0 ]]; then
+        local now
+        now=$(now_epoch)
+        local expired=0
+        local active=0
+        while IFS= read -r key; do
+            [[ -z "$key" ]] && continue
+            local ts ttl
+            ts=$(jq -r --arg k "$key" '.entries[$k].timestamp // 0' "$INTELLIGENCE_CACHE" 2>/dev/null || echo "0")
+            ttl=$(jq -r --arg k "$key" '.entries[$k].ttl // 3600' "$INTELLIGENCE_CACHE" 2>/dev/null || echo "3600")
+            local age=$(( now - ts ))
+            if [[ "$age" -gt "$ttl" ]]; then
+                expired=$((expired + 1))
+            else
+                active=$((active + 1))
+            fi
+        done < <(jq -r '.entries | keys[]' "$INTELLIGENCE_CACHE" 2>/dev/null || true)
+        echo -e "  Active:   ${GREEN}${active}${RESET}"
+        echo -e "  Expired:  ${DIM}${expired}${RESET}"
+    fi
+    echo ""
+}
+
+cmd_cache_clear() {
+    if [[ -f "$INTELLIGENCE_CACHE" ]]; then
+        echo '{"entries":{}}' > "$INTELLIGENCE_CACHE"
+        success "Intelligence cache cleared"
+    else
+        info "No cache file found"
+    fi
+}
+
+main() {
+    local cmd="${1:-help}"
+    shift 2>/dev/null || true
+
+    case "$cmd" in
+        analyze)
+            intelligence_analyze_issue "$@"
+            ;;
+        compose)
+            intelligence_compose_pipeline "$@"
+            ;;
+        predict)
+            intelligence_predict_cost "$@"
+            ;;
+        synthesize)
+            intelligence_synthesize_findings "$@"
+            ;;
+        search-memory)
+            intelligence_search_memory "$@"
+            ;;
+        estimate-iterations)
+            intelligence_estimate_iterations "$@"
+            ;;
+        recommend-model)
+            intelligence_recommend_model "$@"
+            ;;
+        validate-prediction)
+            intelligence_validate_prediction "$@"
+            ;;
+        cache-stats)
+            cmd_cache_stats
+            ;;
+        cache-clear)
+            cmd_cache_clear
+            ;;
+        help|--help|-h)
+            show_help
+            ;;
+        *)
+            error "Unknown command: $cmd"
+            show_help
+            exit 1
+            ;;
+    esac
+}
+
+# Only run main if executed directly (not sourced)
+if [[ "${BASH_SOURCE[0]}" == "$0" ]]; then
+    main "$@"
+fi