Does it actually work? We ran the numbers.

# Boris Cherny's Methodology — Benchmark Policy
> Enterprise TypeScript monorepo (~1571 source files).
> No MCP tools. Use bash: grep, find, cat, head.

## Step 1: Plan before you search (mandatory)
Before running any grep or reading any file:
1. Break the task into sub-questions
2. List relevant packages and file patterns
3. Decide search terms upfront (primary + 2-3 alternatives)
4. Then execute the plan

## Step 2: Search systematically
- Use `grep -rn` to search with line numbers
- Check ALL 4 packages: medusa, admin, cli, plugins
- Run multiple searches — first search rarely catches everything
- For exhaustive tasks: minimum 3 grep passes

## Step 3: Verification loop
After initial findings, ask yourself:
- "What search terms did I miss?"
- "Did I check all packages?"
- "Are there 5+ instances or did I stop too early?"
Run 1-2 more targeted greps to verify completeness.

## Things you must NOT do
- Don't stop after 1-2 examples when asked to find ALL instances
- Don't skip the planning step
- Don't report without specific file:line citations

_Every missed instance is a gap. Iterate until coverage is complete._

# Codebase: Medusa e-commerce monorepo
# ~1,571 TypeScript source files, 4 packages
# Repo: github.com/medusajs/medusa (open source)

# 4 isolated worktrees — no shared state:
# medusa-gr-final/   → GR v3 (AST index, graph_continue → graph_read)
# medusa-jcm/        → JCM  (jcodemunch-mcp, SSE port 8201)
# medusa-normal/     → Normal (bash/grep only)
# medusa-boris/      → Boris (bash/grep + Boris CLAUDE.md)

# Per prompt — all 4 modes run simultaneously:
with ThreadPoolExecutor(max_workers=4) as pool:
    futures = {pool.submit(run_mode, mode, prompt): mode
               for mode in ["gr","jcm","normal","boris"]}

# Each mode call:
result = subprocess.run(
    ["claude", "-p", prompt,
     "--model", "claude-sonnet-4-6",
     "--dangerously-skip-permissions",
     "--no-session-persistence",
     *mcp_flags],          # only set for gr/jcm
    cwd=worktree_path,
    timeout=None           # no timeout — some prompts take 10+ min
)

# LLM judge — 5-dimension rubric (100 pts total):
# findings_accuracy  /25 — real file paths, real line numbers?
# coverage_breadth   /25 — all 4 packages checked?
# depth_quality      /20 — explains WHY each issue is a problem?
# fix_completeness   /20 — all instances found, not just first few?
# actionability      /10 — can a dev act on this immediately?

# Cost calculation (cache-filtered):
cost = (
    (input_tokens - cache_read - cache_write) * 3.00/1M
  + cache_write_tokens * 3.75/1M
  + cache_read_tokens  * 0.30/1M
  + output_tokens      * 15.00/1M
)