Pr Evidence

PR Evidence Extractor

Build a facts-only evidence pack from a PR diff.

Invocation

/pr-evidence owner/repo#123
/pr-evidence https://github.com/owner/repo/pull/123
/pr-evidence 123                  # Uses current repo

Or provide diff directly when prompted.

Purpose

Extract factual evidence from a PR without interpretation or value judgment. This produces structured data that can be used for review, documentation, or further analysis.

Process

Step 1: Fetch PR Diff

gh pr diff {pr_number} --repo {owner}/{repo}

Step 2: Analyze Changed Files

For each file in the diff:

• File path
• Change type (added/modified/deleted)
• Lines added/removed
• File category (source, test, docs, config, etc.)

Step 3: Extract Key Code Changes

For each significant code change:

• File path and line numbers
• What was added/modified/removed (factual description)
• No interpretation of intent or quality

Step 4: Observe Design Decisions

Look for observable patterns (not inferred intent):

• Constants and thresholds introduced
• Error handling patterns used
• Early returns or guard conditions
• Algorithm choices (name the technique, don't evaluate)
• Scope restrictions or exclusions

Step 5: Catalog Test Changes

For each test file:

• Test name/description
• What scenario it tests
• Fixtures added or modified

Step 6: Catalog Doc Changes

For each documentation file:

• Section modified
• Content added/removed

Output Format

<evidence_pack>
  <changed_files>
    <file>
      <path>src/feature.ts</path>
      <type>modified</type>
      <additions>50</additions>
      <deletions>20</deletions>
    </file>
  </changed_files>

  <key_code_changes>
    <change>
      <location>src/feature.ts:42-60</location>
      <description>Added function validateInput that checks for null and returns early</description>
    </change>
    <change>
      <location>src/feature.ts:100</location>
      <description>Changed THRESHOLD constant from 0.01 to 0.001</description>
    </change>
  </key_code_changes>

  <key_design_decisions_observed>
    <decision>
      <pattern>Early return on invalid input</pattern>
      <location>src/feature.ts:45</location>
    </decision>
    <decision>
      <pattern>Uses Bellman-Ford algorithm for cycle detection</pattern>
      <location>src/graph.ts:120-180</location>
    </decision>
  </key_design_decisions_observed>

  <tests_changed>
    <test>
      <file>tests/feature.test.ts</file>
      <description>Added test "handles null input gracefully"</description>
    </test>
    <test>
      <file>tests/fixtures.ts</file>
      <description>Added fixture invalidInputFixture with null values</description>
    </test>
  </tests_changed>

  <docs_changed>
    <doc>
      <file>README.md</file>
      <description>Updated "Configuration" section with new threshold parameter</description>
    </doc>
  </docs_changed>

  <notable_constants_and_thresholds>
    <constant>
      <name>THRESHOLD</name>
      <value>0.001</value>
      <location>src/feature.ts:10</location>
    </constant>
    <constant>
      <name>MAX_ITERATIONS</name>
      <value>100</value>
      <location>src/graph.ts:15</location>
    </constant>
  </notable_constants_and_thresholds>
</evidence_pack>

Rules

1. Facts only: Describe what changed, not whether it's good or bad
2. No interpretation: Don't infer intent beyond what's observable
3. Cite everything: Every claim needs a file:line reference
4. Complete coverage: Include all significant changes, not just interesting ones
5. Structured output: Use consistent XML format for downstream processing

What to Include

• All file changes with statistics
• Function/class additions and modifications
• Constant and threshold values
• Algorithm implementations (name the technique)
• Test scenarios added/modified
• Documentation updates
• Configuration changes

What NOT to Include

• Opinions on code quality
• Recommendations or suggestions
• Speculation about intent
• Evaluation of correctness
• Comparisons to alternatives