shipwright-cli 3.2.0 → 3.3.0

package/docs/plans/2026-03-01-compound-audit-cascade-plan.md

@@ -0,0 +1,921 @@
+# Compound Audit Cascade — Implementation Plan
+
+> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** Replace the one-shot compound_quality stage with an adaptive multi-agent cascade that iteratively probes for bugs until confidence is high.
+
+**Architecture:** New library `compound-audit.sh` with four functions (run_cycle, dedup, escalate, converged). Integrates into existing `stage_compound_quality()` in `pipeline-intelligence.sh`. Agents run as parallel `claude -p --model haiku` calls. Dedup uses structural matching + haiku LLM judge. Convergence stops on no new critical/high OR >98% dup rate OR max_cycles.
+
+**Tech Stack:** Bash 3.2, `claude -p`, `jq`, audit-trail.sh JSONL events
+
+---
+
+### Task 1: Compound Audit Library — Core Scaffolding + Tests
+
+**Files:**
+- Create: `scripts/lib/compound-audit.sh`
+- Create: `scripts/sw-lib-compound-audit-test.sh`
+
+**Step 1: Write the test file scaffold**
+
+```bash
+#!/usr/bin/env bash
+# ╔═══════════════════════════════════════════════════════════════════════════╗
+# ║  compound-audit test suite                                              ║
+# ║  Tests adaptive cascade audit: agents, dedup, escalation, convergence   ║
+# ╚═══════════════════════════════════════════════════════════════════════════╝
+set -euo pipefail
+trap 'echo "ERROR: $BASH_SOURCE:$LINENO exited with status $?" >&2' ERR
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "$SCRIPT_DIR/lib/test-helpers.sh"
+
+print_test_header "Lib: compound-audit Tests"
+
+setup_test_env "sw-lib-compound-audit-test"
+trap cleanup_test_env EXIT
+
+mock_claude
+
+# Source dependencies
+export ARTIFACTS_DIR="$TEST_TEMP_DIR/artifacts"
+mkdir -p "$ARTIFACTS_DIR"
+export LOG_DIR="$TEST_TEMP_DIR/logs"
+mkdir -p "$LOG_DIR"
+
+_AUDIT_TRAIL_LOADED=""
+source "$SCRIPT_DIR/lib/audit-trail.sh"
+audit_init --issue 99 --goal "Test compound audit"
+
+_COMPOUND_AUDIT_LOADED=""
+source "$SCRIPT_DIR/lib/compound-audit.sh"
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# compound_audit_build_prompt
+# ═══════════════════════════════════════════════════════════════════════════════
+print_test_section "compound_audit_build_prompt"
+
+# Test: logic auditor prompt contains specialization
+result=$(compound_audit_build_prompt "logic" "diff --git a/foo.sh" "plan summary" "[]")
+assert_contains "Logic auditor prompt mentions bugs" "$result" "logic error"
+
+# Test: integration auditor prompt contains specialization
+result=$(compound_audit_build_prompt "integration" "diff --git a/foo.sh" "plan summary" "[]")
+assert_contains "Integration auditor prompt mentions wiring" "$result" "wiring"
+
+# Test: completeness auditor prompt mentions spec
+result=$(compound_audit_build_prompt "completeness" "diff --git a/foo.sh" "plan summary" "[]")
+assert_contains "Completeness auditor prompt mentions spec" "$result" "spec"
+
+# Test: prompt includes previous findings when provided
+result=$(compound_audit_build_prompt "logic" "diff" "plan" '[{"description":"known bug"}]')
+assert_contains "Prompt includes previous findings" "$result" "known bug"
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# compound_audit_parse_findings
+# ═══════════════════════════════════════════════════════════════════════════════
+print_test_section "compound_audit_parse_findings"
+
+# Test: parses valid JSON findings
+agent_output='{"findings":[{"severity":"high","category":"logic","file":"foo.sh","line":10,"description":"Off-by-one","evidence":"i < n","suggestion":"Use i <= n"}]}'
+result=$(compound_audit_parse_findings "$agent_output")
+count=$(echo "$result" | jq 'length')
+assert_eq "Parses one finding" "1" "$count"
+
+# Test: extracts severity correctly
+sev=$(echo "$result" | jq -r '.[0].severity')
+assert_eq "Severity is high" "high" "$sev"
+
+# Test: handles empty findings array
+result=$(compound_audit_parse_findings '{"findings":[]}')
+count=$(echo "$result" | jq 'length')
+assert_eq "Empty findings returns empty array" "0" "$count"
+
+# Test: handles malformed output gracefully
+result=$(compound_audit_parse_findings "This is not JSON at all")
+count=$(echo "$result" | jq 'length' 2>/dev/null || echo "0")
+assert_eq "Malformed output returns empty" "0" "$count"
+
+# Test: handles output with markdown wrapping
+result=$(compound_audit_parse_findings '```json
+{"findings":[{"severity":"low","category":"completeness","file":"bar.sh","line":5,"description":"Missing test","evidence":"no test file","suggestion":"Add test"}]}
+```')
+count=$(echo "$result" | jq 'length')
+assert_eq "Markdown-wrapped JSON parsed" "1" "$count"
+
+print_summary
+```
+
+**Step 2: Write the library scaffold with prompt builder and parser**
+
+```bash
+#!/usr/bin/env bash
+# ╔═══════════════════════════════════════════════════════════════════════════╗
+# ║  compound-audit — Adaptive multi-agent audit cascade                    ║
+# ║                                                                         ║
+# ║  Runs specialized audit agents in parallel, deduplicates findings,      ║
+# ║  escalates to specialists when needed, and converges when confidence    ║
+# ║  is high. All functions fail-open with || return 0.                     ║
+# ╚═══════════════════════════════════════════════════════════════════════════╝
+
+[[ -n "${_COMPOUND_AUDIT_LOADED:-}" ]] && return 0
+_COMPOUND_AUDIT_LOADED=1
+
+# ─── Agent prompt templates ────────────────────────────────────────────────
+# Each agent gets the same context but a specialized lens.
+
+_COMPOUND_AGENT_PROMPTS_logic="You are a Logic Auditor. Focus ONLY on:
+- Control flow bugs, off-by-one errors, wrong conditions
+- Algorithm errors, incorrect logic, null/undefined paths
+- Race conditions, state management bugs
+- Edge cases in arithmetic or string operations
+Do NOT report style issues, missing features, or integration problems."
+
+_COMPOUND_AGENT_PROMPTS_integration="You are an Integration Auditor. Focus ONLY on:
+- Missing imports, broken call chains, unconnected components
+- Mismatched interfaces between modules
+- Functions called with wrong arguments or missing arguments
+- Wiring gaps where new code isn't connected to existing code
+Do NOT report logic bugs, style issues, or missing features."
+
+_COMPOUND_AGENT_PROMPTS_completeness="You are a Completeness Auditor. Focus ONLY on:
+- Spec vs. implementation gaps (does the code do what the plan says?)
+- Missing test coverage for new functionality
+- TODO/FIXME/placeholder code left behind
+- Partial implementations (feature started but not finished)
+Do NOT report logic bugs, style issues, or integration problems."
+
+_COMPOUND_AGENT_PROMPTS_security="You are a Security Auditor. Focus ONLY on:
+- Command injection, path traversal, input validation gaps
+- Credential/secret exposure in code or logs
+- Authentication/authorization bypass paths
+- OWASP top 10 vulnerability patterns
+Do NOT report non-security issues."
+
+_COMPOUND_AGENT_PROMPTS_error_handling="You are an Error Handling Auditor. Focus ONLY on:
+- Silent error swallowing (empty catch blocks, ignored return codes)
+- Missing error paths (what happens when X fails?)
+- Inconsistent error handling patterns
+- Unchecked return values from external commands
+Do NOT report non-error-handling issues."
+
+_COMPOUND_AGENT_PROMPTS_performance="You are a Performance Auditor. Focus ONLY on:
+- O(n^2) or worse patterns in loops
+- Unbounded memory allocation or file reads
+- Missing pagination or streaming for large data
+- Repeated expensive operations that could be cached
+Do NOT report non-performance issues."
+
+_COMPOUND_AGENT_PROMPTS_edge_case="You are an Edge Case Auditor. Focus ONLY on:
+- Zero-length inputs, empty strings, empty arrays
+- Maximum/minimum boundary values
+- Unicode, special characters, newlines in data
+- Concurrent access, timing-dependent behavior
+Do NOT report non-edge-case issues."
+
+# ─── compound_audit_build_prompt ───────────────────────────────────────────
+# Builds the full prompt for a specific agent type.
+#
+# Usage: compound_audit_build_prompt "logic" "$diff" "$plan" "$prev_findings_json"
+compound_audit_build_prompt() {
+    local agent_type="$1"
+    local diff="$2"
+    local plan_summary="$3"
+    local prev_findings="$4"
+
+    # Get agent-specific instructions
+    local varname="_COMPOUND_AGENT_PROMPTS_${agent_type}"
+    local specialization="${!varname:-"You are a code auditor. Review the changes for issues."}"
+
+    cat <<EOF
+${specialization}
+
+## Code Changes (cumulative diff)
+\`\`\`
+${diff}
+\`\`\`
+
+## Implementation Plan/Spec
+${plan_summary}
+
+## Previously Found Issues (do NOT repeat these)
+${prev_findings}
+
+## Output Format
+Return ONLY valid JSON (no markdown, no explanation):
+{"findings":[{"severity":"critical|high|medium|low","category":"${agent_type}","file":"path/to/file","line":0,"description":"One sentence","evidence":"The specific code","suggestion":"How to fix"}]}
+
+If no issues found, return: {"findings":[]}
+EOF
+}
+
+# ─── compound_audit_parse_findings ─────────────────────────────────────────
+# Parses agent output into a findings array. Handles malformed output.
+#
+# Usage: compound_audit_parse_findings "$agent_stdout"
+# Output: JSON array of findings (or empty array on failure)
+compound_audit_parse_findings() {
+    local raw_output="$1"
+
+    # Strip markdown code fences if present
+    local cleaned
+    cleaned=$(echo "$raw_output" | sed 's/^```json//;s/^```//;s/```$//' | tr -d '\r')
+
+    # Try to extract findings array
+    local findings
+    findings=$(echo "$cleaned" | jq -r '.findings // []' 2>/dev/null) || findings="[]"
+
+    # Validate it's actually an array
+    if echo "$findings" | jq -e 'type == "array"' >/dev/null 2>&1; then
+        echo "$findings"
+    else
+        echo "[]"
+    fi
+}
+```
+
+**Step 3: Run tests to verify they pass**
+
+Run: `bash scripts/sw-lib-compound-audit-test.sh`
+Expected: All tests pass (prompt builder returns correct content, parser handles valid/invalid/wrapped JSON)
+
+**Step 4: Commit**
+
+```bash
+git add scripts/lib/compound-audit.sh scripts/sw-lib-compound-audit-test.sh
+git commit -m "feat: compound audit library scaffold with prompt builder and parser"
+```
+
+---
+
+### Task 2: Deduplication — Structural + LLM Judge
+
+**Files:**
+- Modify: `scripts/lib/compound-audit.sh`
+- Modify: `scripts/sw-lib-compound-audit-test.sh`
+
+**Step 1: Add dedup tests to test file**
+
+Append to `scripts/sw-lib-compound-audit-test.sh` (before `print_summary`):
+
+```bash
+# ═══════════════════════════════════════════════════════════════════════════════
+# compound_audit_dedup_structural
+# ═══════════════════════════════════════════════════════════════════════════════
+print_test_section "compound_audit_dedup_structural"
+
+# Test: same file + category + nearby lines = duplicate
+findings='[
+  {"severity":"high","category":"logic","file":"foo.sh","line":10,"description":"Bug A","evidence":"code","suggestion":"fix"},
+  {"severity":"high","category":"logic","file":"foo.sh","line":12,"description":"Bug B","evidence":"code","suggestion":"fix"},
+  {"severity":"medium","category":"integration","file":"bar.sh","line":50,"description":"Wiring gap","evidence":"code","suggestion":"fix"}
+]'
+result=$(compound_audit_dedup_structural "$findings")
+count=$(echo "$result" | jq 'length')
+assert_eq "Structural dedup merges nearby same-file same-category" "2" "$count"
+
+# Test: different files are NOT deduped
+findings='[
+  {"severity":"high","category":"logic","file":"foo.sh","line":10,"description":"Bug A","evidence":"code","suggestion":"fix"},
+  {"severity":"high","category":"logic","file":"bar.sh","line":10,"description":"Bug B","evidence":"code","suggestion":"fix"}
+]'
+result=$(compound_audit_dedup_structural "$findings")
+count=$(echo "$result" | jq 'length')
+assert_eq "Different files not deduped" "2" "$count"
+
+# Test: different categories are NOT deduped
+findings='[
+  {"severity":"high","category":"logic","file":"foo.sh","line":10,"description":"Bug A","evidence":"code","suggestion":"fix"},
+  {"severity":"high","category":"security","file":"foo.sh","line":10,"description":"Bug B","evidence":"code","suggestion":"fix"}
+]'
+result=$(compound_audit_dedup_structural "$findings")
+count=$(echo "$result" | jq 'length')
+assert_eq "Different categories not deduped" "2" "$count"
+
+# Test: empty input returns empty
+result=$(compound_audit_dedup_structural "[]")
+count=$(echo "$result" | jq 'length')
+assert_eq "Empty input returns empty" "0" "$count"
+```
+
+**Step 2: Run tests to verify they fail**
+
+Run: `bash scripts/sw-lib-compound-audit-test.sh`
+Expected: FAIL — `compound_audit_dedup_structural: command not found`
+
+**Step 3: Implement structural dedup**
+
+Add to `scripts/lib/compound-audit.sh`:
+
+```bash
+# ─── compound_audit_dedup_structural ───────────────────────────────────────
+# Tier 1 dedup: same file + same category + lines within 5 = duplicate.
+# Keeps the first (highest severity) finding in each group.
+#
+# Usage: compound_audit_dedup_structural "$findings_json_array"
+# Output: Deduplicated JSON array
+compound_audit_dedup_structural() {
+    local findings="$1"
+
+    [[ -z "$findings" || "$findings" == "[]" ]] && { echo "[]"; return 0; }
+
+    # Use jq to group by file+category, then within each group merge findings
+    # whose lines are within 5 of each other
+    echo "$findings" | jq '
+      # Sort by severity priority (critical first) then by line
+      def sev_order: if . == "critical" then 0 elif . == "high" then 1
+        elif . == "medium" then 2 else 3 end;
+
+      sort_by([(.severity | sev_order), .line]) |
+
+      # Group by file + category
+      group_by([.file, .category]) |
+
+      # Within each group, merge findings with lines within 5
+      map(
+        reduce .[] as $item ([];
+          if length == 0 then [$item]
+          elif (. | last | .line) and $item.line and
+               (($item.line - (. | last | .line)) | fabs) <= 5
+          then .  # Skip duplicate (nearby line, same file+category)
+          else . + [$item]
+          end
+        )
+      ) | flatten
+    ' 2>/dev/null || echo "$findings"
+}
+```
+
+**Step 4: Run tests to verify they pass**
+
+Run: `bash scripts/sw-lib-compound-audit-test.sh`
+Expected: All structural dedup tests pass
+
+**Step 5: Commit**
+
+```bash
+git add scripts/lib/compound-audit.sh scripts/sw-lib-compound-audit-test.sh
+git commit -m "feat: structural dedup for compound audit findings"
+```
+
+---
+
+### Task 3: Escalation Logic
+
+**Files:**
+- Modify: `scripts/lib/compound-audit.sh`
+- Modify: `scripts/sw-lib-compound-audit-test.sh`
+
+**Step 1: Add escalation tests**
+
+Append to test file (before `print_summary`):
+
+```bash
+# ═══════════════════════════════════════════════════════════════════════════════
+# compound_audit_escalate
+# ═══════════════════════════════════════════════════════════════════════════════
+print_test_section "compound_audit_escalate"
+
+# Test: security keyword triggers security specialist
+findings='[{"severity":"high","category":"logic","file":"auth.sh","line":10,"description":"Missing input validation allows injection","evidence":"eval $input","suggestion":"Sanitize"}]'
+result=$(compound_audit_escalate "$findings")
+assert_contains "Injection triggers security" "$result" "security"
+
+# Test: error handling keyword triggers error_handling specialist
+findings='[{"severity":"high","category":"integration","file":"foo.sh","line":10,"description":"Empty catch block silently swallows errors","evidence":".catch(() => {})","suggestion":"Log error"}]'
+result=$(compound_audit_escalate "$findings")
+assert_contains "Silent catch triggers error_handling" "$result" "error_handling"
+
+# Test: no trigger keywords returns empty
+findings='[{"severity":"low","category":"completeness","file":"readme.md","line":1,"description":"Missing section in docs","evidence":"no API section","suggestion":"Add it"}]'
+result=$(compound_audit_escalate "$findings")
+assert_eq "No triggers returns empty" "" "$result"
+
+# Test: multiple triggers return unique list
+findings='[
+  {"severity":"high","category":"logic","file":"auth.sh","line":10,"description":"SQL injection in query","evidence":"code","suggestion":"fix"},
+  {"severity":"high","category":"logic","file":"api.sh","line":20,"description":"Missing auth check allows bypass","evidence":"code","suggestion":"fix"}
+]'
+result=$(compound_audit_escalate "$findings")
+# Should contain security but only once
+count=$(echo "$result" | tr ' ' '\n' | grep -c "security" || true)
+assert_eq "Security appears once even with multiple triggers" "1" "$count"
+```
+
+**Step 2: Run tests to verify they fail**
+
+Run: `bash scripts/sw-lib-compound-audit-test.sh`
+Expected: FAIL — `compound_audit_escalate: command not found`
+
+**Step 3: Implement escalation**
+
+Add to `scripts/lib/compound-audit.sh`:
+
+```bash
+# ─── Escalation trigger keywords ──────────────────────────────────────────
+_COMPOUND_TRIGGERS_security="injection|auth|secret|credential|permission|bypass|xss|csrf|traversal|sanitiz"
+_COMPOUND_TRIGGERS_error_handling="catch|swallow|silent|ignore.*error|missing.*error|unchecked|unhandled"
+_COMPOUND_TRIGGERS_performance="O(n|loop.*loop|unbounded|pagination|cache|memory.*leak|quadratic"
+_COMPOUND_TRIGGERS_edge_case="boundary|empty.*input|null.*check|zero.*length|unicode|concurrent|race"
+
+# ─── compound_audit_escalate ──────────────────────────────────────────────
+# Scans findings for trigger keywords, returns space-separated specialist list.
+#
+# Usage: compound_audit_escalate "$findings_json_array"
+# Output: Space-separated specialist names (e.g., "security error_handling")
+compound_audit_escalate() {
+    local findings="$1"
+
+    [[ -z "$findings" || "$findings" == "[]" ]] && return 0
+
+    # Flatten all finding text for keyword scanning
+    local all_text
+    all_text=$(echo "$findings" | jq -r '.[].description + " " + .[].evidence + " " + .[].file' 2>/dev/null | tr '[:upper:]' '[:lower:]') || return 0
+
+    local specialists=""
+    local spec
+    for spec in security error_handling performance edge_case; do
+        local varname="_COMPOUND_TRIGGERS_${spec}"
+        local pattern="${!varname:-}"
+        if [[ -n "$pattern" ]] && echo "$all_text" | grep -qEi "$pattern" 2>/dev/null; then
+            specialists="${specialists:+${specialists} }${spec}"
+        fi
+    done
+
+    echo "$specialists"
+}
+```
+
+**Step 4: Run tests to verify they pass**
+
+Run: `bash scripts/sw-lib-compound-audit-test.sh`
+Expected: All escalation tests pass
+
+**Step 5: Commit**
+
+```bash
+git add scripts/lib/compound-audit.sh scripts/sw-lib-compound-audit-test.sh
+git commit -m "feat: trigger-based escalation for compound audit"
+```
+
+---
+
+### Task 4: Convergence Detection
+
+**Files:**
+- Modify: `scripts/lib/compound-audit.sh`
+- Modify: `scripts/sw-lib-compound-audit-test.sh`
+
+**Step 1: Add convergence tests**
+
+Append to test file (before `print_summary`):
+
+```bash
+# ═══════════════════════════════════════════════════════════════════════════════
+# compound_audit_converged
+# ═══════════════════════════════════════════════════════════════════════════════
+print_test_section "compound_audit_converged"
+
+# Test: no critical/high findings = converged
+result=$(compound_audit_converged "[]" "[]" 1 3)
+assert_eq "No findings = converged" "no_criticals" "$result"
+
+# Test: only low/medium findings = converged
+new='[{"severity":"low","category":"logic","file":"foo.sh","line":1,"description":"Minor","evidence":"x","suggestion":"y"}]'
+result=$(compound_audit_converged "$new" "[]" 1 3)
+assert_eq "Only low findings = converged" "no_criticals" "$result"
+
+# Test: critical finding = NOT converged
+new='[{"severity":"critical","category":"logic","file":"foo.sh","line":1,"description":"Major bug","evidence":"x","suggestion":"y"}]'
+result=$(compound_audit_converged "$new" "[]" 1 3)
+assert_eq "Critical finding = not converged" "" "$result"
+
+# Test: max cycles reached = converged regardless
+new='[{"severity":"critical","category":"logic","file":"foo.sh","line":1,"description":"Major bug","evidence":"x","suggestion":"y"}]'
+result=$(compound_audit_converged "$new" "[]" 3 3)
+assert_eq "Max cycles = converged" "max_cycles" "$result"
+
+# Test: all findings are duplicates of previous = converged (dup rate)
+prev='[{"severity":"high","category":"logic","file":"foo.sh","line":10,"description":"Bug","evidence":"x","suggestion":"y"}]'
+new='[{"severity":"high","category":"logic","file":"foo.sh","line":11,"description":"Same bug","evidence":"x","suggestion":"y"}]'
+result=$(compound_audit_converged "$new" "$prev" 1 3)
+assert_eq "All dupes = converged" "dup_rate" "$result"
+```
+
+**Step 2: Run tests to verify they fail**
+
+Run: `bash scripts/sw-lib-compound-audit-test.sh`
+Expected: FAIL — `compound_audit_converged: command not found`
+
+**Step 3: Implement convergence**
+
+Add to `scripts/lib/compound-audit.sh`:
+
+```bash
+# ─── compound_audit_converged ─────────────────────────────────────────────
+# Checks stop conditions for the cascade loop.
+#
+# Usage: compound_audit_converged "$new_findings" "$all_prev_findings" $cycle $max_cycles
+# Output: Reason string if converged ("no_criticals", "dup_rate", "max_cycles"), empty if not
+compound_audit_converged() {
+    local new_findings="$1"
+    local prev_findings="$2"
+    local cycle="$3"
+    local max_cycles="$4"
+
+    # Hard cap: max cycles reached
+    if [[ "$cycle" -ge "$max_cycles" ]]; then
+        echo "max_cycles"
+        return 0
+    fi
+
+    # No findings at all = converged
+    local new_count
+    new_count=$(echo "$new_findings" | jq 'length' 2>/dev/null || echo "0")
+    if [[ "$new_count" -eq 0 ]]; then
+        echo "no_criticals"
+        return 0
+    fi
+
+    # Check for critical/high in new findings
+    local crit_high_count
+    crit_high_count=$(echo "$new_findings" | jq '[.[] | select(.severity == "critical" or .severity == "high")] | length' 2>/dev/null || echo "0")
+
+    # If previous findings exist, check duplicate rate via structural match
+    local prev_count
+    prev_count=$(echo "$prev_findings" | jq 'length' 2>/dev/null || echo "0")
+    if [[ "$prev_count" -gt 0 && "$new_count" -gt 0 ]]; then
+        # Count how many new findings structurally match previous ones
+        local dup_count=0
+        local i=0
+        while [[ "$i" -lt "$new_count" ]]; do
+            local nf nc nl
+            nf=$(echo "$new_findings" | jq -r ".[$i].file // \"\"" 2>/dev/null)
+            nc=$(echo "$new_findings" | jq -r ".[$i].category // \"\"" 2>/dev/null)
+            nl=$(echo "$new_findings" | jq -r ".[$i].line // 0" 2>/dev/null)
+
+            # Check if any previous finding matches file+category+nearby line
+            local match
+            match=$(echo "$prev_findings" | jq --arg f "$nf" --arg c "$nc" --argjson l "$nl" \
+                '[.[] | select(.file == $f and .category == $c and ((.line // 0) - $l | fabs) <= 5)] | length' 2>/dev/null || echo "0")
+            [[ "$match" -gt 0 ]] && dup_count=$((dup_count + 1))
+            i=$((i + 1))
+        done
+
+        # If all findings are duplicates, converged
+        if [[ "$dup_count" -eq "$new_count" ]]; then
+            echo "dup_rate"
+            return 0
+        fi
+    fi
+
+    # No critical/high = converged
+    if [[ "$crit_high_count" -eq 0 ]]; then
+        echo "no_criticals"
+        return 0
+    fi
+
+    # Not converged
+    echo ""
+    return 0
+}
+```
+
+**Step 4: Run tests to verify they pass**
+
+Run: `bash scripts/sw-lib-compound-audit-test.sh`
+Expected: All convergence tests pass
+
+**Step 5: Commit**
+
+```bash
+git add scripts/lib/compound-audit.sh scripts/sw-lib-compound-audit-test.sh
+git commit -m "feat: convergence detection for compound audit cascade"
+```
+
+---
+
+### Task 5: Run Cycle — Parallel Agent Execution
+
+**Files:**
+- Modify: `scripts/lib/compound-audit.sh`
+- Modify: `scripts/sw-lib-compound-audit-test.sh`
+
+**Step 1: Add run_cycle tests**
+
+Append to test file (before `print_summary`):
+
+```bash
+# ═══════════════════════════════════════════════════════════════════════════════
+# compound_audit_run_cycle
+# ═══════════════════════════════════════════════════════════════════════════════
+print_test_section "compound_audit_run_cycle"
+
+# Mock claude to return predictable findings
+cat > "$TEST_TEMP_DIR/bin/claude" <<'MOCK'
+#!/usr/bin/env bash
+# Return findings based on prompt content
+if echo "$*" | grep -q "Logic Auditor"; then
+    echo '{"findings":[{"severity":"high","category":"logic","file":"foo.sh","line":10,"description":"Off by one","evidence":"i < n","suggestion":"Use <="}]}'
+elif echo "$*" | grep -q "Integration Auditor"; then
+    echo '{"findings":[]}'
+elif echo "$*" | grep -q "Completeness Auditor"; then
+    echo '{"findings":[{"severity":"low","category":"completeness","file":"bar.sh","line":5,"description":"Missing test","evidence":"no test","suggestion":"Add test"}]}'
+else
+    echo '{"findings":[]}'
+fi
+MOCK
+chmod +x "$TEST_TEMP_DIR/bin/claude"
+
+# Test: run_cycle with core agents returns merged findings
+export COMPOUND_AUDIT_MODEL="haiku"
+result=$(compound_audit_run_cycle "logic integration completeness" "diff content" "plan summary" "[]" 1)
+count=$(echo "$result" | jq 'length')
+assert_eq "Core cycle returns 2 findings" "2" "$count"
+
+# Test: findings include correct categories
+cats=$(echo "$result" | jq -r '.[].category' | sort | tr '\n' ' ' | sed 's/ $//')
+assert_eq "Categories are logic and completeness" "completeness logic" "$cats"
+```
+
+**Step 2: Run tests to verify they fail**
+
+Run: `bash scripts/sw-lib-compound-audit-test.sh`
+Expected: FAIL — `compound_audit_run_cycle: command not found`
+
+**Step 3: Implement run_cycle**
+
+Add to `scripts/lib/compound-audit.sh`:
+
+```bash
+# ─── compound_audit_run_cycle ─────────────────────────────────────────────
+# Runs multiple agents in parallel and collects their findings.
+#
+# Usage: compound_audit_run_cycle "logic integration completeness" "$diff" "$plan" "$prev_findings" $cycle
+# Output: Merged JSON array of all findings
+compound_audit_run_cycle() {
+    local agents="$1"
+    local diff="$2"
+    local plan_summary="$3"
+    local prev_findings="$4"
+    local cycle="$5"
+
+    local model="${COMPOUND_AUDIT_MODEL:-haiku}"
+    local temp_dir
+    temp_dir=$(mktemp -d) || return 0
+
+    # Emit cycle start event
+    type audit_emit >/dev/null 2>&1 && \
+        audit_emit "compound.cycle_start" "cycle=$cycle" "agents=$agents" || true
+
+    # Launch agents in parallel
+    local pids=()
+    local agent
+    for agent in $agents; do
+        local prompt
+        prompt=$(compound_audit_build_prompt "$agent" "$diff" "$plan_summary" "$prev_findings")
+
+        (
+            local output
+            output=$(echo "$prompt" | claude -p --model "$model" 2>/dev/null) || output='{"findings":[]}'
+            echo "$output" > "$temp_dir/${agent}.json"
+        ) &
+        pids+=($!)
+    done
+
+    # Wait for all agents
+    local pid
+    for pid in "${pids[@]}"; do
+        wait "$pid" 2>/dev/null || true
+    done
+
+    # Merge findings from all agents
+    local all_findings="[]"
+    for agent in $agents; do
+        local agent_file="$temp_dir/${agent}.json"
+        if [[ -f "$agent_file" ]]; then
+            local agent_findings
+            agent_findings=$(compound_audit_parse_findings "$(cat "$agent_file")")
+
+            # Emit individual findings as audit events
+            local i=0
+            local fc
+            fc=$(echo "$agent_findings" | jq 'length' 2>/dev/null || echo "0")
+            while [[ "$i" -lt "$fc" ]]; do
+                local sev desc file line
+                sev=$(echo "$agent_findings" | jq -r ".[$i].severity" 2>/dev/null)
+                desc=$(echo "$agent_findings" | jq -r ".[$i].description" 2>/dev/null)
+                file=$(echo "$agent_findings" | jq -r ".[$i].file" 2>/dev/null)
+                line=$(echo "$agent_findings" | jq -r ".[$i].line" 2>/dev/null)
+                type audit_emit >/dev/null 2>&1 && \
+                    audit_emit "compound.finding" "cycle=$cycle" "agent=$agent" \
+                        "severity=$sev" "file=$file" "line=$line" "description=$desc" || true
+                i=$((i + 1))
+            done
+
+            # Merge into all_findings
+            all_findings=$(echo "$all_findings" "$agent_findings" | jq -s '.[0] + .[1]' 2>/dev/null || echo "$all_findings")
+        fi
+    done
+
+    # Cleanup
+    rm -rf "$temp_dir" 2>/dev/null || true
+
+    echo "$all_findings"
+}
+```
+
+**Step 4: Run tests to verify they pass**
+
+Run: `bash scripts/sw-lib-compound-audit-test.sh`
+Expected: All run_cycle tests pass
+
+**Step 5: Commit**
+
+```bash
+git add scripts/lib/compound-audit.sh scripts/sw-lib-compound-audit-test.sh
+git commit -m "feat: parallel agent execution for compound audit cycles"
+```
+
+---
+
+### Task 6: Pipeline Integration — Replace compound_quality Body
+
+**Files:**
+- Modify: `scripts/lib/pipeline-intelligence.sh` (~line 1310, inside the cycle loop)
+- Modify: `scripts/lib/pipeline-intelligence.sh` (~line 1148, source compound-audit.sh)
+
+**Step 1: Source compound-audit.sh in pipeline-intelligence.sh**
+
+Find the imports section at the top of `pipeline-intelligence.sh` and add:
+
+```bash
+# Source compound audit cascade library
+if [[ -f "$SCRIPT_DIR/lib/compound-audit.sh" ]]; then
+    source "$SCRIPT_DIR/lib/compound-audit.sh"
+fi
+```
+
+**Step 2: Replace the cycle body (lines ~1310-1365)**
+
+Inside `stage_compound_quality()`, after the hardened quality gates and before the summary, replace the existing cycle loop with the cascade:
+
+```bash
+    # ── ADAPTIVE CASCADE AUDIT ──
+    local all_findings="[]"
+    local active_agents="logic integration completeness"  # Core 3
+
+    # Gather context for agents
+    local _cascade_diff
+    _cascade_diff=$(git diff "${BASE_BRANCH:-main}...HEAD" 2>/dev/null | head -5000) || _cascade_diff=""
+    local _cascade_plan=""
+    if [[ -f "$ARTIFACTS_DIR/plan.md" ]]; then
+        _cascade_plan=$(head -200 "$ARTIFACTS_DIR/plan.md" 2>/dev/null) || true
+    fi
+
+    local cycle=0
+    while [[ "$cycle" -lt "$max_cycles" ]]; do
+        cycle=$((cycle + 1))
+
+        echo ""
+        echo -e "${PURPLE}${BOLD}━━━ Compound Audit — Cycle ${cycle}/${max_cycles} ━━━${RESET}"
+        info "Agents: $active_agents"
+
+        # Run agents in parallel
+        local cycle_findings
+        cycle_findings=$(compound_audit_run_cycle "$active_agents" "$_cascade_diff" "$_cascade_plan" "$all_findings" "$cycle") || cycle_findings="[]"
+
+        # Dedup within this cycle
+        cycle_findings=$(compound_audit_dedup_structural "$cycle_findings") || cycle_findings="[]"
+
+        local cycle_count
+        cycle_count=$(echo "$cycle_findings" | jq 'length' 2>/dev/null || echo "0")
+        local cycle_crit
+        cycle_crit=$(echo "$cycle_findings" | jq '[.[] | select(.severity == "critical" or .severity == "high")] | length' 2>/dev/null || echo "0")
+
+        # Report findings
+        if [[ "$cycle_count" -gt 0 ]]; then
+            warn "Cycle ${cycle}: ${cycle_count} findings (${cycle_crit} critical/high)"
+            # Count for pipeline scoring
+            total_critical=$((total_critical + $(echo "$cycle_findings" | jq '[.[] | select(.severity == "critical")] | length' 2>/dev/null || echo "0")))
+            total_major=$((total_major + $(echo "$cycle_findings" | jq '[.[] | select(.severity == "high")] | length' 2>/dev/null || echo "0")))
+            total_minor=$((total_minor + $(echo "$cycle_findings" | jq '[.[] | select(.severity == "medium" or .severity == "low")] | length' 2>/dev/null || echo "0")))
+        else
+            success "Cycle ${cycle}: no findings"
+        fi
+
+        # Check convergence
+        local converge_reason
+        converge_reason=$(compound_audit_converged "$cycle_findings" "$all_findings" "$cycle" "$max_cycles") || converge_reason=""
+
+        # Emit cycle complete event
+        type audit_emit >/dev/null 2>&1 && \
+            audit_emit "compound.cycle_complete" "cycle=$cycle" "findings=$cycle_count" \
+                "critical_high=$cycle_crit" "converged=$converge_reason" || true
+
+        if [[ -n "$converge_reason" ]]; then
+            success "Converged: $converge_reason"
+            type audit_emit >/dev/null 2>&1 && \
+                audit_emit "compound.converged" "reason=$converge_reason" "total_cycles=$cycle" || true
+            break
+        fi
+
+        # Merge findings for next cycle's context
+        all_findings=$(echo "$all_findings" "$cycle_findings" | jq -s '.[0] + .[1]' 2>/dev/null || echo "$all_findings")
+
+        # Escalation: trigger specialists for next cycle
+        if type compound_audit_escalate >/dev/null 2>&1; then
+            local specialists
+            specialists=$(compound_audit_escalate "$cycle_findings") || specialists=""
+            if [[ -n "$specialists" ]]; then
+                info "Escalating: adding $specialists"
+                active_agents="logic integration completeness $specialists"
+            fi
+        fi
+    done
+
+    # Save all findings to artifact
+    echo "$all_findings" > "$ARTIFACTS_DIR/compound-audit-findings.json" 2>/dev/null || true
+```
+
+**Step 3: Validate syntax**
+
+Run: `bash -n scripts/lib/pipeline-intelligence.sh`
+Expected: No errors
+
+**Step 4: Run existing tests to verify no regression**
+
+Run: `bash scripts/sw-lib-compound-audit-test.sh && bash scripts/sw-lib-audit-trail-test.sh`
+Expected: Both pass (28/28 audit trail, all compound audit tests)
+
+**Step 5: Commit**
+
+```bash
+git add scripts/lib/pipeline-intelligence.sh scripts/lib/compound-audit.sh
+git commit -m "feat: integrate compound audit cascade into compound_quality stage"
+```
+
+---
+
+### Task 7: Audit Trail Finalize — Include Compound Findings in Reports
+
+**Files:**
+- Modify: `scripts/lib/audit-trail.sh` (~line 188, `_audit_build_markdown`)
+
+**Step 1: Add compound findings section to markdown report**
+
+In `_audit_build_markdown()`, after the "Build Loop" section (line ~243), add:
+
+```bash
+  # Compound audit findings section
+  local compound_events
+  compound_events=$(grep '"type":"compound.finding"' "$jsonl_file" 2>/dev/null || true)
+  if [[ -n "$compound_events" ]]; then
+    cat <<'EOF'
+
+## Compound Audit Findings
+
+EOF
+    echo "$compound_events" | while IFS= read -r line; do
+      local sev file desc
+      sev=$(echo "$line" | grep -o '"severity":"[^"]*' | cut -d'"' -f4)
+      file=$(echo "$line" | grep -o '"file":"[^"]*' | cut -d'"' -f4)
+      desc=$(echo "$line" | grep -o '"description":"[^"]*' | cut -d'"' -f4)
+      echo "- **[$sev]** \`$file\`: $desc"
+    done
+
+    # Convergence summary
+    local converge_line
+    converge_line=$(grep '"type":"compound.converged"' "$jsonl_file" 2>/dev/null | tail -1 || true)
+    if [[ -n "$converge_line" ]]; then
+      local reason cycles
+      reason=$(echo "$converge_line" | grep -o '"reason":"[^"]*' | cut -d'"' -f4)
+      cycles=$(echo "$converge_line" | grep -o '"total_cycles":"[^"]*' | cut -d'"' -f4)
+      echo ""
+      echo "**Converged** after ${cycles} cycle(s): ${reason}"
+    fi
+  fi
+```
+
+**Step 2: Run audit trail tests**
+
+Run: `bash scripts/sw-lib-audit-trail-test.sh`
+Expected: 28/28 pass (existing tests unaffected)
+
+**Step 3: Commit**
+
+```bash
+git add scripts/lib/audit-trail.sh
+git commit -m "feat: include compound audit findings in markdown report"
+```
+
+---
+
+## Verification Checklist
+
+1. **Unit tests**: `bash scripts/sw-lib-compound-audit-test.sh` — all tests pass
+2. **Audit trail tests**: `bash scripts/sw-lib-audit-trail-test.sh` — 28/28 pass
+3. **Loop tests**: `bash scripts/sw-loop-test.sh` — 56/61 (5 pre-existing)
+4. **Detection tests**: `bash scripts/sw-lib-pipeline-detection-test.sh` — 70/70
+5. **Syntax check**: `bash -n scripts/lib/compound-audit.sh && bash -n scripts/lib/pipeline-intelligence.sh`
+6. **E2E**: Run pipeline with compound_quality enabled, verify `compound.cycle_start`, `compound.finding`, and `compound.converged` events in JSONL