brennhill/cost-per-accepted-change/cpac-cli

Cost per accepted change

This skill helps you measure cost per accepted change ("CPAC") — the canonical definition from aifinops.dev. The metric is the fully-loaded cost of producing software that reached production and stayed there, divided by the number of changes that did, size-normalized so one unit represents a comparable amount of substantive work.

Use this skill when the user wants to:

• Answer "is our AI investment paying back?" with a defensible single number
• Compute CPAC for a real GitHub repository over a date window
• Decompose CPAC into its five cost components for diagnosis
• Apply the 500-LOC size-normalization rule to a list of merged changes

Do not use this skill for:

• Comparing individual engineers (the metric is undefined at the individual level)
• Single-change "what did this PR cost" lookups (the metric is only defined over a population)
• Ranking unrelated teams against each other (work mix invalidates the comparison)

What's in this skill

A CLI (cpac) and an MCP server (cpac-mcp) sharing the same audit + calculator core.

CLI

# Pure math, given the 5 cost inputs and the denominator.
cpac calc --modelCost 1200 --infraCost 400 --engineeringTime 18000 \
          --reviewCost 6000 --reworkCost 2400 --acceptedChanges 42
 
# Audit a real GitHub repo window for the denominator. Requires gh CLI auth.
cpac audit --repo owner/name --since 2026-04-01 --until 2026-04-30
 
# Audit + full CPAC if cost inputs are passed too.
cpac audit --repo owner/name --since 2026-04-01 --until 2026-04-30 \
           --modelCost 5000 --infraCost 800 --engineeringTime 45000 \
           --reviewCost 12000 --reworkCost 3000 --json

Run cpac --help for the full flag list. Add --json to any command for machine-readable output.

MCP server

cpac-mcp is a stdio MCP server exposing three tools:

• calculate_cpac — pure math (5 cost components + denominator → CPAC value and per-component shares)
• normalize_changes — apply 500-LOC normalization to a list of { linesChanged } records
• audit_repo — walk a GitHub repository over a date window, fetch merged PRs, detect reverts, and emit the accepted-change-unit count

Add to Claude Code (or any MCP-aware client) by adding this to mcpServers:

{
  "mcpServers": {
    "cpac": {
      "command": "npx",
      "args": ["-y", "cost-per-accepted-change-cli", "cpac-mcp"]
    }
  }
}

How to use it well

1. Pick a window long enough that the denominator stabilizes. CPAC stabilizes around 25–100 accepted change units per window. Smaller teams need longer windows. A 5-engineer team should report quarterly, not monthly.

1. Run `audit_repo` (or `cpac audit`) first. This gives you the denominator and surfaces revert/repair candidates. Review the repairCandidates list manually — the FAQ rule is "if a reasonable reviewer would describe the follow-up as 'fixing what the original got wrong,' it invalidates." Adjust by passing --since/--until or by overriding acceptedChanges downstream.

1. Gather the five cost components for the same window. Numerator inputs:

• modelCost — LLM/API spend (your provider's billing dashboard, segmented by team if possible)
• infraCost — compute, storage, observability, tooling allocated to the production loop
• engineeringTime — fully-loaded hourly rate × hours allocated to delivery work
• reviewCost — fully-loaded hourly rate × hours spent reviewing/gating AI work
• reworkCost — fully-loaded hourly rate × hours spent reverting/fixing non-survivors

1. Call `calculate_cpac` (or `cpac calc`) to get the headline number plus per-component shares. The shares are diagnostic: a team where rework crept from 8% to 18% has a very different problem than one where it stayed at 8%, even if the headline moved by the same amount.

1. Pair the result with at least one leading indicator. Change failure rate (DORA), per-suggestion acceptance rate, DevEx pulse score. The headline tells you the system's cost; the components tell you the source; leading indicators tell you why.

1. Report the trend, not the snapshot. A single window in isolation tells you very little. Two consecutive windows in the same direction, with stable accounting, is the minimum signal.

Important guardrails

• The 30-day survival window means reporting lag. Current-month CPAC cannot be finalized until 30 days after the month closes. audit_repo flags PRs as tooRecentToEvaluate when they're inside the survival window.
• Hold accounting constant. Hourly rates, attribution rules, exclusions (vendored/generated/lockfiles), and window length must hold steady across the time series. Changing accounting mid-stream invalidates the trend.
• Don't target an absolute number. Targets invite gaming. Track the trend; investigate when it moves; do not legislate the number.
• Heuristic repair candidates need human judgment. The CLI flags PRs whose title looks like a fix and whose files overlap with an in-window PR. It does not auto-invalidate. Apply the FAQ's judgment rule.

Worked example (matches the homepage)

Inputs:
  modelCost         $1,200    (4.3%)
  infraCost           $400    (1.4%)
  engineeringTime  $18,000   (64.3%)
  reviewCost        $6,000   (21.4%)
  reworkCost        $2,400    (8.6%)
  acceptedChanges       42
 
  Total cost:    $28,000
  CPAC:          $666.67

Reference

• Canonical definition: <https://aifinops.dev>
• FAQ (window selection, normalization rule, anti-gaming): <https://aifinops.dev/faq>
• Instrumentation guide: <https://aifinops.dev/instrument>
• Originally defined in The Delivery Gap (Brenn Hill, 2026)
• Source: <https://github.com/brennhill/cost-per-accepted-change>