neosloc

sloccount counted lines of code and fed them to COCOMO to estimate the effort to build software. With LLMs writing code, lines are cheap and that estimate has lost its meaning. What is still expensive is everything around the code: how easily other systems and agents can call it, run it, rely on it and change it safely, and the knowledge that exists only in the code and its history.

neosloc measures that:

• ten integrability dimensions, each scored 0–4 with the evidence behind the level and the gaps phrased as actions;
• the retrofit effort to bring every dimension up to "solid", and whether the system is cheap to wrap;
• neoCOCOMO: what the codebase is worth when an agent can re-type it, next to classic COCOMO;
• optional LLM evaluators on Anthropic or OpenRouter: a probe attempts real integration tasks using only the docs, a judge checks the answers against the source, and a review audits every static level.

pip install neosloc
neosloc path/to/repo

Interface surface   ■■■□  3 solid      The framework generates a contract from code, so it tracks the implementation.
                      + routes:fastapi/flask: 15 route declarations in 3 files (app/routes/admin.py)
                      + generated-spec: FastAPI (auto OpenAPI) (app/main.py)
                      - Check the generated spec into the repo so changes are reviewable and diffable.
...
Integrability index (0-4, assessed dimensions):  1.80
Retrofit effort to level 3 (agent-assisted):     7.5 person-days

Value (neoCOCOMO, cost approach; see neosloc/value.py)
  Classic COCOMO (sloccount):    4.6 KSLOC -> 11.8 person-months, 6.4 months, $133k
  Behaviour captured:            38% (tests 0.14, contract 0.25, docs 1.00)
  Reproduce with agents:         x0.38 of classic -> 4.5 PM
  Rediscover uncaptured history: 102 fix / 474 commits, 19 authors -> 4.4 PM knowledge, 2.7 PM uncaptured
  Value:                         5.2 PM ~ $59k  (at $11k per PM)
  Knowledge at risk:             38% of replacement cost lives only in code and history

Documentation: https://ingmmo.com/neosloc/

Install

OpenRouter models need no extra package, only OPENROUTER_API_KEY.

Dimensions

Each dimension asks one question, answered per surface (service, CLI, library, frontend) with a ladder of checkable requirements: a surface reaches level N when it meets every requirement up to N, and the dimension takes the best surface's level. Dimensions that can't apply (events for a batch CLI, portability for a tool that stores no data) are reported as n/a and left out of the index.

Details, with every requirement: dimensions, retrofit effort, neoCOCOMO.

LLM evaluators

# Probe: a model attempts 8 integration tasks from the docs; answers are graded against the code
neosloc --agentic --scope both --transcripts runs/ .

# Probe with Claude, judge with another vendor through OpenRouter
neosloc --judge --judge-model openrouter:openai/gpt-5.6-luna .

# Review: a panel audits every static level against the code
neosloc --review --review-model openrouter:google/gemini-3.8-flash,openrouter:deepseek/deepseek-v4-pro-0813 .

Models are written provider:model. The default is anthropic:claude-opus-5-5 at effort medium. A comma-separated list runs a panel and reports agreement. Every model call shares one --max-cost budget (default $5), and the report states the total. See LLM evaluators and Using OpenRouter.

Scripting

--json output (reports and errors) follows a published JSON Schema (neosloc --schema). Exit statuses distinguish error classes (2 usage, 3 path, 4 evaluator setup). Diagnostics go to stderr: -q, --log-level, --log-format json. See JSON output and Command line.

Status

All thresholds and coefficients are uncalibrated, so treat the numbers as rankings. Calibration and limits describes how to fix that; --review and --agentic exist partly to do it.

Development

python -m unittest discover -s tests -t .     # no network, no API spend
python -m neosloc .

See Writing criteria and AGENTS.md. Releases: bump neosloc/__init__.py, add a CHANGELOG.md section, push a vX.Y.Z tag.

MIT licensed.