Skip to content

build-plan.md

Latest commit

 

History

History
117 lines (76 loc) · 2.1 KB

build-plan.md

File metadata and controls

117 lines (76 loc) · 2.1 KB
Raw

Build Plan

Layer 1: Shippable Spine

Goal: a minimal but impressive system that judges one patch against one task in one repo.

Task 1: Repository Scaffold

  • TypeScript CLI
  • test runner
  • lint/typecheck scripts
  • docs
  • mergecode.yaml schema draft

Done when: mergecode --help works and tests run.

Task 2: Patch Worktree Runner

  • create isolated worktree
  • apply patch
  • capture apply errors
  • cleanup option

Done when: fixture patch applies without touching the source repo.

Task 3: Verification Runner

  • run configured commands
  • capture status, duration, output summary
  • timeout handling

Done when: a failing test produces structured evidence.

Task 4: Diff Parser

  • parse unified diff
  • touched files
  • added/deleted lines
  • file categories
  • deleted test detection

Done when: fixture diffs produce stable summaries.

Task 5: First Rubric Engine

  • hard gates
  • weighted categories
  • findings with evidence
  • verdict

Done when: known good/bad fixture patches produce expected verdicts.

Task 6: Markdown Report

  • 30-second summary
  • verdict
  • why
  • verification
  • diff risk
  • findings

Done when: report is useful without opening source code.

Layer 2: Depth Roadmap

Stage A: Real Fixture Corpus

Create 5-10 realistic repo tasks with multiple candidate patches each:

  • good patch
  • test-passing but bad patch
  • failing patch
  • overbroad patch

Stage B: Deleted/Weakened Test Analysis

Detect:

  • removed assertions
  • .skip / .only
  • snapshot churn
  • changed expected values

Stage C: Architecture Drift

Use repo context:

  • existing module map
  • similar utility detection
  • layer boundary hints
  • duplicate logic candidates

Stage D: Human Calibration

Compare MergeCode verdicts with human labels.

Output:

  • confusion matrix
  • false positives
  • false negatives
  • rubric changes

Stage E: Agent Comparison

Run Claude/Codex/Cursor/Aider patches through same tasks.

Output:

  • mergeable rate
  • test-pass vs mergeable gap
  • recurring failure modes

Build Rule

Each task must leave the repo working. Do not start Stage B until the spine can produce a complete report.