Pipeline Plan 60

The plan file is ready. Let me provide the plan summary here since the artifacts directory has restricted permissions.

---

Implementation Plan: Ruthless Quality Validation (Issue #60)

Summary

Replace the "optimistically compliant" quality system with ruthless, multi-layer validation across 5 layers.

Files to Modify

File	Changes
scripts/sw-pipeline.sh	DoD zero-auto-pass, quality score gating, multi-perspective review, progressive deepening
scripts/sw-loop.sh	Intelligent completion, adaptive iteration budget, diminishing returns detection
scripts/sw-adversarial.sh	Multi-persona support (security, correctness, architecture auditors)
templates/pipelines/*.json (all 8)	Add quality_threshold, critical_blocks, progressive_deepening, multi_perspective
.claude/daemon-config.json	Enable adversarial, simulation, architecture by default
scripts/sw-pipeline-test.sh	Tests for all quality gates and blocking behavior

Implementation Steps

Step 1: Zero Auto-Pass DoD (sw-pipeline.sh:4638)

Replace *) item_passed=true with item_passed=false + optional AI verification. Unknown items fail by default. Also improve existing test/coverage/debug-statement checks to be more thorough.

Step 2: Quality Score as Hard Gate (sw-pipeline.sh:6195)

• Read quality_threshold from pipeline config (default 80, replacing hardcoded 60)
• Read critical_blocks config (default true)
• Critical findings ALWAYS block pipeline regardless of score
• Post detailed score breakdown table to GitHub issue

Step 3: Multi-Perspective Review (sw-adversarial.sh)

New multi_perspective_review() function running 3 distinct auditor personas:

1. Security auditor: injection, auth bypass, secrets, CSRF, SSRF
2. Correctness auditor: off-by-one, null access, race conditions, edge cases
3. Architecture auditor: naming violations, coupling, circular deps, pattern breaks

Each writes to adversarial-{persona}.md, combined summary to JSON.

Step 4: Progressive Deepening (sw-pipeline.sh:5825)

Each compound quality cycle goes deeper, not wider:

• Cycle 1 (fast): E2E, security scan, quality checks
• Cycle 2 (medium): + adversarial, negative prompting, DoD
• Cycle 3 (deep): + simulation, architecture, multi-perspective review
• Cycle 4+ (targeted): Only re-run audits that previously found issues

Controlled by progressive_deepening config flag. Legacy behavior preserved when false.

Step 5: Intelligent Build Completion (sw-loop.sh)

• Adaptive budget: Complexity <= 30 → 10 iter, <= 60 → 20, > 60 → 30 (only when not explicitly set)
• Diminishing returns: Warn when velocity < 5 lines over last 3 iterations
• Enhanced guard_completion: Reject if uncommitted changes or TODO/FIXME markers in diff

Step 6: Template Updates (all 8 templates)

Template	threshold	cycles	critical_blocks	progressive	multi_perspective
fast	50	2	true	false	false
hotfix	60	2	true	false	false
cost-aware	60	3	true	true	false
standard	70	4	true	true	true
autonomous	80	5	true	true	true
full	80	5	true	true	true
deployed	80	5	true	true	true
enterprise	90	7	true	true	true

Step 7: Enable Intelligence Defaults

Change adversarial_enabled, simulation_enabled, architecture_enabled to true in daemon-config.

Step 8: Tests (8 new test cases)

1. DoD unknown item fails (not auto-passes)
2. Quality score blocks below threshold
3. Quality score passes above threshold
4. Critical findings block regardless of score
5. Progressive deepening cycle 1 = fast (adversarial skipped)
6. Progressive deepening cycle 3 = deep (all audits run)
7. Multi-perspective review produces 3 persona files
8. Convergence plateau detection

Task Checklist

• Task 1: Replace DoD auto-pass with zero-auto-pass default in run_dod_audit()
• Task 2: Improve DoD verification for test, coverage, and debug checks
• Task 3: Make quality score threshold configurable (replace hardcoded 60)
• Task 4: Add critical-findings-always-block logic
• Task 5: Implement multi_perspective_review() in sw-adversarial.sh
• Task 6: Wire multi-perspective review into stage_compound_quality()
• Task 7: Implement progressive deepening (fast/medium/deep/targeted)
• Task 8: Add adaptive iteration budget to sw-loop.sh
• Task 9: Add diminishing returns detection to build loop
• Task 10: Enhance guard_completion() with uncommitted/TODO checks
• Task 11: Update all 8 pipeline templates
• Task 12: Enable intelligence defaults in daemon-config
• Task 13: Write tests for all quality gates
• Task 14: Run full test suite and fix regressions

Definition of Done

• *) item_passed=true replaced with item_passed=false + AI verification
• Quality score threshold configurable per template (not hardcoded 60)
• Critical findings block pipeline regardless of score
• Multi-perspective review generates 3 auditor reports
• Progressive deepening runs appropriate checks per cycle depth
• Adaptive iteration budget based on complexity scoring
• Diminishing returns detection active
• All 8 templates updated with new quality config
• Intelligence features enabled by default
• All new code tested with PASS/FAIL assertions
• Full test suite passes (npm test exits 0)
• Bash 3.2 compatible