trusty-git-analytics

Three stages, one command

tga analyze runs the whole pipeline. Each stage is also a subcommand, because on a large history you will want to re-run one without paying for the others.

• tga collect — walk each configured repository, extract commit metadata and diff stats, resolve author identities, and write it all to SQLite. Optionally pull pull-request and issue metadata from GitHub, JIRA, Linear, or Azure DevOps alongside it.
• tga classify — run every unclassified commit through the cascade and write the verdict back. Rule tiers run in parallel.
• tga report — aggregate per author, per week, and per DORA metric, then write CSV, JSON, and Markdown into the output directory.

The database is a local SQLite file, so every number in a report is one you can go and check with a query.

The classification cascade

Naming what a commit actually did is the hard part, and a single heuristic gets it wrong often enough to be useless. tga tries tiers in order and takes the first confident answer: a manual override you pinned, the issue type from a linked ticket, a project-key mapping, an Aho-Corasick scan for conventional-commit prefixes, regex patterns, a weighted sum over several independent signals, and fuzzy heuristics for merges and reverts.

An LLM tier sits at the end for the commits the rules could not place, disabled by default and enabled with --use-llm. Its answers are accepted only above a confidence threshold you set. --no-external skips every network-bound source, which is what you want while iterating on a rule file.

The rule set is introspectable rather than a black box: tga rules list enumerates it, tga rules test "<message>" shows you which tier would fire, and tga override pins a verdict that outranks all of them.

What comes out

• tga author <email> — a per-engineer drill-down: commits, effort, pull requests, category mix.
• tga pr-metrics — pull-request metrics per engineer, once PR fetching is turned on.
• tga dora — all four DORA metrics, fed by tga deployments and tga incidents.
• tga aliases — merge the four email addresses one person has committed under, so the per-author numbers mean anything at all.

Getting started

tga install is an interactive wizard that writes the config for you. A hand-written one can be as small as a list of repository paths — every other section has a default. Note the package and binary are both tga, not the crate directory's longer name.

Acquisition due diligence

tga audit

One non-interactive command runs the whole pipeline across every repository your config names and hands the result to trusty-review, which renders an eight-section technical due-diligence report — the deliverable someone reads while deciding whether to buy the codebase. A stage that fails does not stop the sweep; it becomes a named line in the report's Gaps & Caveats instead of a zero in a cell.

It has its own page: the eight stages, the six flags, what lands in the output directory, and the six things it deliberately refuses to estimate.

Install and run tga audit →