feat: Verdict-to-AutoFix pipeline - promote common Gemini corrections to Ponder

[martymcenroe]

Summary

Create a feedback loop where common Gemini verdict corrections become Ponder Stibbons auto-fixes, reducing LLM costs and improving iteration speed over time.

The Learning Loop

┌─────────────────────────────────────────────────────────────────────────┐
│                     THE LEARNING CYCLE                                  │
│                                                                         │
│   Gemini ──► Verdict ──► Analyzer ──► Pattern ──► Ponder Auto-Fix      │
│   Review     "BLOCK:     mines        detected    fixes it before      │
│              missing X"  patterns     (N>3)       Gemini sees it       │
│                              │                          │               │
│                              └────── Learning ──────────┘               │
│                                                                         │
└─────────────────────────────────────────────────────────────────────────┘

How It Works

Phase 1: Pattern Mining (Verdict Analyzer)

The Verdict Analyzer already processes Gemini verdicts. Extend it to track:

class VerdictPattern:
    pattern_type: str      # "missing_section", "wrong_format", "path_typo"
    regex_or_rule: str     # How to detect this in drafts
    fix_strategy: str      # How Ponder should fix it
    occurrence_count: int  # How many times seen
    example_verdicts: list # Sample verdict excerpts

Phase 2: Promotion Threshold

When a pattern occurs N times (configurable, default 3):

1. Verdict Analyzer flags it as "promotion candidate"
2. Human reviews and approves promotion
3. Pattern becomes a Ponder auto-fix rule

Phase 3: Auto-Fix Integration

Ponder loads promoted rules and applies them pre-review:

# In Ponder's auto_fix.py
def load_promoted_fixes() -> list[AutoFixRule]:
    """Load fixes promoted from verdict analysis."""
    rules_path = Path("docs/auto-fix-rules/promoted/")
    return [AutoFixRule.from_yaml(f) for f in rules_path.glob("*.yaml")]

Phase 4: Impact Measurement

Track metrics:

• Gemini calls saved (fixes that would have triggered BLOCK)
• Time saved (auto-fix is instant vs 30s+ LLM call)
• Accuracy (did the auto-fix actually match what Gemini would want?)

Example Pattern Promotion

Observed: Gemini blocks 5 LLDs for "Section 11 uses ### instead of ##"

Pattern extracted:

name: section_11_header_format
trigger: "Section 11" in verdict AND "header" in verdict
detection: r'^### 11\.'
fix: Replace with '## 11.'
promoted_from: verdicts/2026-02/LLD-099-v1.md, LLD-101-v2.md, ...

Result: Next time a draft has ### 11., Ponder fixes it before Gemini review.

Connection to "How the AgentOS Learns"

This is a key learning mechanism:

1. Gemini teaches - Blocks drafts with issues
2. Analyzer observes - Tracks patterns in blocks
3. System learns - Promotes patterns to auto-fixes
4. Quality improves - Fewer blocks, faster iteration

Update wiki page: How the AgentOS Learns

Files to Modify

File	Change
agentos/workflows/requirements/verdict_analyzer.py	Add pattern tracking
agentos/workflows/requirements/nodes/auto_fix.py	Load promoted rules
docs/auto-fix-rules/promoted/	Store promoted patterns
wiki/How-the-AgentOS-Learns.md	Document this loop

Dependencies

• feat: Ponder Stibbons - The Compositor (auto-fix layer for mechanical corrections) #307 - Ponder Stibbons (The Compositor) must exist first

Success Criteria

• Verdict Analyzer tracks correction patterns
• Patterns with N+ occurrences flagged for promotion
• Promoted patterns loaded by Ponder
• Metrics show reduced Gemini round-trips
• Wiki documents the learning loop

Labels

enhancement, workflow, learning