Every term the reports use, in plain language — the 60-second version. One line each, no mathematics required.
| Churn (CRN) | The total amount of change between two versions of a codebase: everything added, deleted, or edited. The headline number. |
| ADD / DEL / CHG | Every changed statement is counted in exactly one of three categories: added (new), deleted (removed), or changed in place (still there, but altered — an inline edit). Colours everywhere in the reports: green = added, blue = deleted, red = changed. |
| Moved (MOV) | A statement that moved to another place in the file without being changed. Counted as moved, not as a deletion plus an addition, and reported beside churn, never inside it — so reorganising code does not inflate the change figures. |
| LLOC vs SLOC | SLOC counts pure physical source lines — total lines minus blanks and comments (comment lines are reported separately as COM_LOC). LLOC counts logical statements — what the program actually does. Reformatting a file moves lots of lines and zero statements, which is why serious change measurement uses LLOC. |
| Inline edit (CHG_LLOC) | A change inside an existing statement — not an addition, not a deletion. The report column is CHG_LLOC. Humans do this constantly; AI tools barely do it at all. |
| Coverage | The percentage of files actually measured, stated on every run — so nothing is silently skipped. |
| REWORK | The share of all churn spent editing existing code in place (CHG ÷ CRN). Established hand-maintained projects sit near 17% — one statement in six. A measured agent-built codebase: 0.19% — one in five hundred. |
| REP_CHURN | The mirror of REWORK: the share of churn that was replacement — deletions plus additions — rather than editing. High replacement is the signature of generate-and-regenerate development. REWORK + REP_CHURN = 1, always. |
| TRUE_CHURN | Churn from files your developers actually wrote. Machine-generated files — lockfiles, minified bundles, code-generator output — are subtotalled out by named rules, nothing hidden. In one real npm release, 77% of the "churn" came from a single generated file. |
| Data metrics on the GUI: the DATA SHARE tile | Some "code" is data wearing code's syntax — firmware arrays, generated lookup tables — where one statement can hold thousands of values. CodeDelta splits working code from data, counts change inside data per element, and reports a working-code REWORK with the data noise removed. |
| Baseline & merge gate | A baseline is an accepted snapshot of findings. The gate fails a pull request only on findings newer than the baseline, or on a policy breach — measurement becomes enforcement only when you say so. |
| Post-review churn | How much a pull request changed after review started, shown in the PR comment. |
| Desktop app | Point-and-click on macOS, Windows or Linux — pick two versions, get the report. |
| On every pull request | A GitHub Action (GitLab equivalent included) measures each PR automatically and posts the summary as a comment on the PR itself — where reviewers already look. Can be set to block a merge on your policy. |
| Command line & CI | One command for engineers and pipelines (Jenkins, Docker, scheduled runs) — same numbers, headless. |
| What comes out | Human-readable HTML reports and the Code Browser — both versions side by side, with the classes and their relationships, so every figure traces to the lines that produced it — plus CSV/JSON for spreadsheets and dashboards, and SARIF for the GitHub Security tab. Everything runs inside your own infrastructure; no code leaves your machines. |
| Data Centre | The history database has its own card: what it holds, a health check, automatic backups before the first scan of the day or week to any folder (external, network or cloud-synced), and restore. Everything also at the command line. |
| MCP server | AI agents (Claude Code, Cursor and others) can call the real CodeDelta engine directly instead of improvising their own churn scripts. Included in every desktop download. |
| Cost | Free to test, with full functionality, until 2027. |
| AI agent | Software that calls an AI model and acts on the answer — running commands, writing files, sending requests. In your code or hidden in a dependency. |
| Agent Scan | Static detection of the AI layer: named agent SDKs and model calls, model output reaching execution (the dangerous pattern), and agent infrastructure left in the tree — workspaces, configs, committed credentials. Evidence with file and line, never guesswork. |
| AI-BOM | The AI Bill of Materials: an inventory of every place your software touches AI — which models, whose servers, what the output can reach, where data goes. Native or CycloneDX format; the artifact auditors and the EU AI Act ask about. |
| AI audit | The optional ML-based estimate of how AI-authored code looks. Off by default, and always labelled what it is: pointers for review, never verdicts — per-file AI authorship detection is not reliably possible, and we publish the evidence for that. |
| Agent trailers | When some AI tools commit code they append a machine-readable co-author line to the commit message. CodeDelta finds them by exact-pattern scan of the git history — deterministic, no guesswork. The count is a floor on AI involvement, never a total: the lines are voluntary and removable. |
| Threat Index | Compares two versions of a project and reports what in the change looks like an attack: the same new code in many files, obfuscated code, a large new file in a quiet release. Pointers for review, not verdicts. |
| Agent Tracer | Reads the session logs AI coding agents (Claude Code, Codex, Cursor, Copilot CLI, Gemini CLI, Antigravity) leave on a machine and reports what each agent ran, deleted, pushed and sent, to which hosts, under which model, and what the human refused. Offline. |
| AI Log Keeper | Saves those session logs and monitors token consumption. A free separate download for macOS, Windows and Linux. |
Every figure quoted here traces to a published, reproducible measurement — sources for all of it on the evidence page.

We can’t justify buying a third party tool! “We will just use Ai to make our own tool!”
Sure! Go ahead!
Except you will need a dedicated, experienced, prompt/software engineer with knowledge of code quality, CMVC, git hub, git lab, macOS, Linux and windows, programming language parsers, code syntax scanning, sql database architecture, software product design and testing plus documentation expertise including user guides.
Probably $100k as an absolute minimum (plus the same again to support it!) and you still won’t have 25 years of code churn tool experience - i.e “why true churn is not changed LOC/SLOC!”
CodeDelta is therefore a financial and technical no-brainer.
CodeDelta runs on macOS, Linux and Windows, free to try — or on every pull request. Download CodeDelta → · Questions or licensing: contact codedelta.app