commit → merged → stabilized → survived
Measure what survives, not what ships.
Your team ships more than ever. devXOS reads the Git history behind it and separates the code that lasted from the code that just landed — and shows whether AI is helping it last.
stabilization_ratio
Shipping got cheap. Proof got expensive.
AI tools multiplied your commits and halved your cycle time. Every dashboard is green — and none of them can tell you whether the code holds. The metrics that mattered before AI measure motion. The ones that matter now measure what remains.
durability_by_origin · org Δ
+29.3 → −12.5pp
Our own headline, AI code outlasting human code, came from how repositories were averaged. Counting only the repositories that measured both sides, the sign flipped. The engine caught it, and the number changed.
3,497 commits · 1,211 PRs
58 repos
Where signal discovery started: 58 repositories, 3,497 commits and 1,211 merged PRs of a real organization.
harness_profile
The agent works inside a harness. devXOS watches the harness.
Agentic engineering wraps AI in tests, reviews, and boundaries so speed doesn't outrun control. devXOS doesn't run inside that loop — it reads what the harness actually let through, across the three dimensions that decide whether a codebase is holding.
Maintainability
Is the code staying changeable?
Churn chains, duplication, moved code, and how much brand-new code gets rewritten within weeks.
- churn_chains
- duplicate_ratio
- moved_code_pct
- new_code_churn
Architecture
Is the structure holding its shape?
A structural harnessability score: type coverage, module boundaries, and the sensors the repo carries. Not yet calibrated against outcomes, and labeled that way.
- harnessability
- type_coverage
- module_boundaries
Behaviour
Is correctness holding up?
Reverts, fix latency, correction cascades, and change-failure rate — conditioned on who wrote the code.
- revert_rate
- fix_latency
- correction_cascades
- dora_cfr
Every number on this page is a field in the report — no metric ships without a name, a unit, and a documented computation.
Intelligence that audits itself.
usage_rollup × durability
Where the tokens went
Anonymous AI effort per repository, next to whether the code lasted. Shown only where a repository has at least four contributors, and named for what it is: a correlation at the repository grain, not a cause.
cross_correlations
Checked against DORA
Engine series are correlated with change-failure rate and time-to-restore, with significance tested against t-table critical values. A correlation we can't defend doesn't ship.
Runs where the history lives.
$ curl -fsSL https://devxos.ai/install.sh | shOne script. Python 3.11+ and Git are the only requirements.
$ devxos hook installA prepare-commit-msg hook records which AI tool touched each commit — Copilot, Claude, Cursor, Windsurf, and others.
$ devxos /path/to/repo --trendThe engine reads Git metadata and diffs locally and writes a Markdown report plus JSON metrics. Your code never leaves the machine.
$ devxos login && devxos pushOptional: push metrics to the platform for org-wide dashboards, trends over time, and cross-repo comparison.
What devXOS refuses to be.
Not a surveillance tool
The core analyzes systems, never individuals: no per-author scores, no leaderboards. Person-level numbers exist only in an opt-in module, under a contract the developer can verify, and ranking by volume or AI cost is refused there too.
Not a live dashboard
Point-in-time reports. Durability is measured in weeks, not in minutes of screen time.
Not an IDE plugin
It reads Git history and the integrations you connect. Nothing to install in anyone's editor.
Not a productivity tracker
It measures whether the work lasted, not how fast it was typed.
Run it on a repo you care about.
One command. The engine runs locally — the platform is optional.
$ curl -fsSL https://devxos.ai/install.sh | shPython 3.11+ · Git · zero dependencies · code never uploaded