| name | quality-audit |
| description | Iterative codebase quality audit with multi-agent validation and escalating-depth SEEK/VALIDATE/FIX/RECURSE cycle. Enforces fix-all-per-cycle rule — every confirmed finding must be fixed before moving to the next cycle. Use for quality audit, code audit, codebase review, technical debt audit, refactoring opportunities, module quality check, or architecture review. |
| metadata | {"version":"4.0","author":"amplihack"} |
Quality Audit Workflow
Workflow Graph
flowchart TD
START[Start Audit] --> SEEK
subgraph CYCLE["Audit Cycle (min 3, max 6)"]
SEEK[SEEK: Scan for issues<br/>reviewer agent<br/>escalating depth per cycle] --> VAL_CHECK{Findings found?}
VAL_CHECK -->|no| DECISION
VAL_CHECK -->|yes| V1[VALIDATE Agent 1<br/>analyzer] & V2[VALIDATE Agent 2<br/>reviewer] & V3[VALIDATE Agent 3<br/>architect]
V1 & V2 & V3 --> MERGE[Merge Validations<br/>require >=2/3 agreement]
MERGE --> FIX_CHECK{Confirmed findings?}
FIX_CHECK -->|no| DECISION
FIX_CHECK -->|yes| FIX[FIX ALL confirmed<br/>fix-agent → DEFAULT_WORKFLOW<br/>fix-all-per-cycle rule]
FIX --> VERIFY[Verify all fixes applied]
VERIFY --> ACCUM[Accumulate cycle history]
ACCUM --> DECISION{Recurse decision}
end
DECISION -->|"CONTINUE<br/>(cycle < min OR<br/>NEW high/critical found OR<br/>>3 NEW medium found)"| SEEK
DECISION -->|"STOP<br/>(thresholds met OR<br/>max cycles reached)"| SUMMARY
SUMMARY[Final Summary<br/>architect agent] --> SELF[Self-Improvement Review<br/>architect agent]
SELF --> DONE[Audit Complete]
style CYCLE fill:#f9f9f9,stroke:#333,stroke-width:2px
Purpose
Orchestrates a systematic, parallel quality audit of any codebase with automated remediation through PR generation and PM-prioritized recommendations.
When I Activate
I automatically load when you mention:
- "quality audit" or "code audit"
- "codebase review" or "full code review"
- "refactoring opportunities" or "technical debt audit"
- "module quality check" or "architecture review"
- "parallel analysis" with multiple agents
What I Do
Execute a 7-phase workflow that:
- Familiarizes with the project (investigation phase)
- Audits using parallel agents across codebase divisions
- Creates GitHub issues for each discovered problem
3.5. Validates against recent PRs (prevents false positives)
- Generates PRs in parallel worktrees per remaining issues
- Reviews PRs with PM architect for prioritization
- Reports consolidated recommendations in master issue
Quick Start
User: "Run a quality audit on this codebase"
Skill: *activates automatically*
"Beginning quality audit workflow..."
The 7 Phases
Phase 1: Project Familiarization
- Run investigation workflow on project structure
- Map modules, dependencies, and entry points
- Understand existing patterns and architecture
Phase 2: Parallel Quality Audit
- Divide codebase into logical sections
- Deploy multiple agent types per section (analyzer, reviewer, security, optimizer)
- Apply PHILOSOPHY.md standards ruthlessly
- Check module size, complexity, single responsibility
Phase 3: Issue Assembly
- Create GitHub issue for each finding
- Include severity, location, recommendation
- Tag with appropriate labels
- Add unique IDs, keywords, and file metadata
Phase 3.5: Post-Audit Validation [NEW]
- Scan merged PRs from last 30 days (configurable)
- Calculate confidence scores for PR-issue matches
- Auto-close high-confidence matches (≥90%)
- Tag medium-confidence matches (70-89%) for verification
- Add bidirectional cross-references between issues and PRs
- Target: <5% false positive rate
Phase 4: Parallel PR Generation
- Create worktree per remaining open issue (
worktrees/fix-issue-XXX)
- Run DEFAULT_WORKFLOW.md in each worktree
- Generate fix PR for each confirmed open issue
Phase 5: PM Review
- Invoke pm-architect skill
- Group PRs by category and priority
- Identify dependencies between fixes
Phase 6: Master Report
- Create master GitHub issue
- Link all related issues and PRs
- Prioritized action plan with recommendations
Philosophy Enforcement
This workflow ruthlessly applies:
- Ruthless Simplicity: Flag over-engineered modules
- Module Size Limits: Target <300 LOC per module
- Single Responsibility: One purpose per brick
- Zero-BS: No stubs, no TODOs, no dead code
- Anti-Fallback (#2805, #2810): Detect silent degradation and error swallowing patterns
- Structural Analysis (#2809): Flag oversized files, deeply nested code, and tangled dependencies
Detection Categories
Standard Categories
| Category | What It Detects |
|---|
| Security | Hardcoded secrets, missing input validation, string interpolation in queries |
| Reliability | Missing timeouts, bare except clauses, unhandled async |
| Dead Code | Unused imports, unreachable branches, stale TODOs |
| Test Gaps | Files without tests, tests without assertions |
| Doc Gaps | Public functions without docstrings, outdated docs |
Extended Categories
| Category | What It Detects |
| ---------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --- | -------------------------- |
| Silent Fallbacks | except: pass, broad catches that return defaults silently, fallback chains that mask failures, ?? defaultValue hiding missing config, dict.get(key, default) on required values, | | fallback in shell scripts |
| Error Swallowing | Catch blocks with no re-raise/re-throw, error-to-None/null transforms, catch-all discarding exceptions, log-only catch blocks, empty catch blocks, catch (Exception) returning false/default/empty collection |
| Result Dropping | Fire-and-forget async (_ = Task(), asyncio.create_task() without error handling), unchecked HTTP response status, discarded return values, Task.WhenAll/Promise.all/asyncio.gather without individual failure checks, unchecked subprocess.run() |
| Shell Anti-Patterns | \|\| true, >/dev/null 2>&1, 2>/dev/null, set +e, \|\| fallback_command, missing set -euo pipefail |
| Silent Truncation | Take(N)/[:N]/.slice(0,N) without logging, .Where()/list comprehensions that silently drop items that should be processed, string substring without bounds logging |
| Async Anti-Patterns | async void (C#), .Result/.Wait() sync-over-async, unawaited coroutines/promises, shared mutable state without synchronization, CancellationToken not propagated, Timer/CancellationTokenSource not disposed |
| Config Divergence | Env vars defined in deploy configs but read with silent fallbacks in code, IsDevelopment() guards that could leak to staging/prod, services expecting config that infrastructure doesn't provide |
| Validation Gaps | API endpoints without input validation, string interpolation in SQL/GraphQL/Cypher, missing pagination limits, missing request size limits, enum parsing from user input without validation, trusting deserialized external data without null checks |
| Health & Observability | Degraded reported when Unhealthy is appropriate, background worker failures not surfaced to /health, log-only error handling without metrics, permanent errors treated as transient (retried instead of dead-lettered), partial success marked as full success |
| Retry Anti-Patterns | Retry loops that fall through silently after exhaustion, circuit breakers that open without alerting, retry logic that eventually gives up without raising the last error |
| Structural Issues | Files >500 LOC, functions >50 lines, nesting >4 levels, >5 parameters, circular imports |
| Documentation | Point-in-time content, unprofessional tone (pirate speak, chatbot artifacts), quality/correctness gaps |
| Hardcoded Limits | Non-configurable numeric caps ([:N], max_X = N), silent truncation without logging, data loss from processing limits |
Multi-Agent Validation (v3.0)
Every finding is validated by 3 independent agents (analyzer, reviewer, architect). A finding is confirmed only if ≥2 agents agree. This eliminates false positives before any fixes are attempted.
Iterative Loop with Escalating Depth (v3.0)
Cycle 1: SEEK → VALIDATE (3 agents) → FIX → decision
Cycle 2: SEEK (deeper) → VALIDATE → FIX → decision
Cycle 3: SEEK (deepest) → VALIDATE → FIX → decision
...continues if thresholds not met
Loop rules:
- Minimum 3 cycles always run
- Continue past 3 if: any high/critical NEW findings emerged, or >3 medium NEW findings
- Maximum 6 cycles (safety valve)
- Each cycle: fresh eyes, dig deeper, challenge prior findings
- Fixes use the fix-agent which delegates to the full 22-step DEFAULT_WORKFLOW via Skill(skill="default-workflow")
- Fix-all-per-cycle rule (#2842): Every confirmed finding in a cycle MUST be
fixed before the cycle is complete. No partial cycles. No deferring findings to
"follow-up issues" or "next cycle". If SEEK finds issues, FIX must address ALL
of them.
- Loop decision based on NEW findings (#2842): The decision to continue is
based on whether the current cycle discovered NEW issues, not whether old issues
remain unfixed (they shouldn't — the fix-all rule prevents that).
- Fix verification step: After fixes, a verification step compares confirmed
findings against fix results to ensure nothing was skipped.
Run via recipe:
amplihack recipe run amplifier-bundle/recipes/quality-audit-cycle.yaml \
-c target_path=src/amplihack \
-c repo_path="." \
-c min_cycles=3 \
-c max_cycles=6 \
--verbose
Configuration
Override defaults via recipe context or environment:
Structured Inputs (recipe context, per #2843):
| Input | Default | Description |
|---|
target_path | src/amplihack | Directory to audit |
repo_path | . | Repository root; sets working_dir for agents |
min_cycles | 3 | Minimum audit cycles |
max_cycles | 6 | Maximum cycles (safety valve) |
validation_threshold | 2 | Min validators that must agree (out of 3) |
severity_threshold | medium | Minimum severity to report |
module_loc_limit | 300 | Flag modules exceeding this LOC |
fix_all_per_cycle | true | Must fix ALL findings before next cycle (#2842) |
categories | (all) | Comma-separated list of categories to check |
Available Categories: security, reliability, dead_code, silent_fallbacks,
error_swallowing, result_dropping, shell_anti_patterns, silent_truncation,
async_anti_patterns, config_divergence, validation_gaps, health_observability,
retry_anti_patterns, structural, hardcoded_limits, test_gaps, doc_gaps, documentation
Example invocation:
amplihack recipe run amplifier-bundle/recipes/quality-audit-cycle.yaml \
-c target_path=crates/amplihack-fleet/src \
-c repo_path="." \
-c min_cycles=3 \
-c max_cycles=6 \
-c severity_threshold=medium \
-c module_loc_limit=300 \
-c fix_all_per_cycle=true \
-c categories="security,reliability,dead_code,silent_fallbacks,error_swallowing" \
--verbose
Core Settings (environment):
AUDIT_PARALLEL_LIMIT: Max concurrent worktrees (default: 8)
Phase 3.5 Validation Settings:
AUDIT_PR_SCAN_DAYS: Days to scan for recent PRs (default: 30)AUDIT_AUTO_CLOSE_THRESHOLD: Confidence % for auto-close (default: 90)AUDIT_TAG_THRESHOLD: Confidence % for tagging (default: 70)AUDIT_ENABLE_VALIDATION: Enable Phase 3.5 (default: true)
