adlc

Measure how fast one repository ships, on your own machine, and compare the numbers with industry benchmarks.

AI coding tools report acceptance rates, active users and lines generated. None of that says whether the team ships more. adlc looks at what happens after the code is written: how many pull requests merge, how long they take, how long people take to review them, how big they are, and how much new code gets rewritten.

It reads GitHub through your own gh login and prints the result in the terminal. Nothing is uploaded, and nothing is written to disk unless you ask.

What it measures

Metric	What it is
Throughput	Merged PRs per week, and per author per week
Open to merge	Time from PR open to merge
Time to first human review	Time from PR open to the first review by a person. Bots and AI reviewers are ignored. Also shown split by whether an AI reviewer had commented first
First human review to merge	Time from that first review to merge
PR size	Lines changed
Rework	Share of all line changes that rewrote code less than 21 days old, from git blame. Rework is a rough proxy for churn caused by AI: code that was generated, merged, and then had to be rewritten because it was wrong, sloppy or did not fit the codebase

Times are reported as median and 75th percentile. Each 75th percentile is placed in a band from the LinearB 2026 engineering benchmarks: elite, good, fair or needs focus.

Install

You need uv, git and the GitHub CLI, signed in with gh auth login.

Run it without installing:

uvx --from git+https://github.com/unblocked/ai-development-lifecycle.git adlc baseline --since 90d

Or put adlc on your path:

uv tool install git+https://github.com/unblocked/ai-development-lifecycle.git

Run it

From inside a full clone of the repository:

adlc baseline --since 90d

Or the most recent N merged pull requests, whatever their age:

adlc baseline --limit 3000

The clone must have full history. On a shallow clone git blame cannot date old lines, so adlc stops and tells you. Run git fetch --unshallow, or pass --skip-rework.

A run of 300 pull requests takes under a minute. The rework pass takes a few minutes on a large monorepo. Add --skip-rework for a quick first look.

No repository to hand? Run the bundled sample:

adlc baseline --sample

What you get

Delivery baseline for cline/cline ─────────────────────────────────────────────────────────────────────
Measured on 2026-09-13. 299 merged pull requests covering the last 57.9 days of a 90 day window.
Metric                                            Value   Band, LinearB 2026 p75
───────────────────────────────────────────────────────────────────────────────────────────────────────
Pull requests                                       299
Merged per week                                    36.1
Merged per author per week                          1.9   good, 1.2 to 2
Open to merge, median                             4.0 h
Open to merge, 75th percentile                   34.6 h   fair, 18.0 h to 40.0 h for pickup plus review
Time to first human review, median                2.3 h
Time to first human review, 75th percentile      18.3 h   needs focus, over 16.0 h
Time to first human review, after an AI review   91 min
Time to first human review, no AI review first    5.7 h
First human review to merge, median              48 min
First human review to merge, 75th percentile     16.0 h   fair, 14.0 h to 24.0 h
PR size, median changed lines                       210
PR size, 75th percentile                            657   needs focus, over 228 lines
Rework rate, code under 21 days old rewritten        7%   fair, 5% to 8%
New code deleted again within 30 days                6%   own trend

What stands out
  • Merged PRs per author per week: 1.9 (good, 1.2 to 2).
  • Open to merge: 4.0 h median, 34.6 h p75 (fair, 18.0 h to 40.0 h for pickup plus review).
  • First human review: 2.3 h median, 18.3 h p75 (needs focus, over 16.0 h).
  • People review 73% sooner when an AI reviewer got there first (91 min against 5.7 h).
  • ...

Caveats
  • ...

What this report assumes
  • AI reviewers with code comments: greptile-apps (141 PRs), ...
  • ...

Add --output-dir DIR to also write the report as Markdown and JSON, or --format markdown or --format json to print one of those instead of the table.

Options

Flag	Default	What it does
--since 90d	90 days	Window for pull requests, capped at 1,000
--limit N	off	The N most recent merged PRs instead of a window
--skip-rework	off	Skip the git blame pass
--rework-window 30d	30 days	Commits to scan for rework
--rework-age 30d	30 days	Age under which a deleted line counts as rework in the second rework row
--ai-reviewer LOGIN	built-in list	Treat this GitHub login as an AI reviewer. Repeatable
--ignore-author LOGIN	none	Drop PRs opened by this login, for mirror or sync bots. Repeatable
--format table	table	Print table, markdown or json
--output-dir DIR	off	Also write adlc-baseline-YYYY-MM-DD.md and .json into this directory
--repo-path .	current directory	Path to the clone

Reading the numbers

• Bands come from the LinearB 2026 engineering benchmarks.
• Compare a team with its own earlier baseline first, and with the benchmarks second.
• Report per team, never per person.
• Measure again in a quarter with the same window. The second measurement is the useful one.

Related

• engineering-social-graph: who reviews whom and where the experts are, from the same PR history
• The Context Maturity Field Guide

License

MIT