tga 2.2.0

Developer productivity analytics — git commit collection, classification, and reporting
Documentation

trusty-git-analytics

Analyze git repositories to measure developer productivity — classify commit work types, track weekly velocity, and export CSV/JSON/Markdown reports.

What It Does

tga walks one or more local git repositories, collects every commit into a SQLite database, classifies each commit into a work category (feature, bugfix, refactor, etc.) using a multi-tier classification cascade (including the new weighted-sum Tier 2.5 added in 1.3.0), then aggregates the results into per-author, per-week, DORA, velocity, and quality reports. It is a feature-complete Rust port of gitflow-analytics with the same YAML config schema and the same SQLite schema — existing config files work without modification.

1.4.0 adds four new external classification sources (Linear, Shortcut, Confluence, Datadog) and fixes a SQLite WAL data-loss bug (issue #298) that could cause up to 22 minutes of classification work to be silently discarded on exit.

Installation

From crates.io (recommended)

cargo install tga

This installs the tga binary to ~/.cargo/bin/. Ensure ~/.cargo/bin is in your PATH.

From source

git clone https://github.com/bobmatnyc/trusty-tools
cd trusty-tools/crates/trusty-git-analytics
cargo install --path .

Verify installation

tga --version
tga --help

Quick Start

Run Your First Analysis

Step 1 — Create a config.yaml:

repositories:
  - path: ~/code/my-project
    name: my-project

output:
  directory: ./reports
  formats: [csv, json, markdown]

Step 2 — Run the full pipeline:

tga analyze --config config.yaml

Step 3 — Find reports in ./reports/:

A full run writes 14 files: 9 CSV, 4 JSON, and 1 Markdown report. The most commonly used ones are:

reports/
├── authors.csv         # Per-author commit summary
├── weekly_activity.csv # Week-by-week breakdown
├── ... (7 more CSV files: DORA, velocity, quality, etc.)
├── report.json         # Full structured payload
├── ... (3 more JSON files)
└── report.md           # Narrative Markdown report

Configuration

Minimal config.yaml

repositories:
  - path: ~/code/my-repo
    name: my-repo

All other sections are optional. When output.formats is omitted, all three formats (CSV, JSON, Markdown) are written.

Full config reference

Key Type Default Description
repositories list required Repos to analyze
developer_aliases map {} Canonical name → list of emails/aliases
team object — Alternative to developer_aliases; roster with email
output.directory path ./reports Where reports are written
output.formats list [csv, json, markdown] csv, json, and/or markdown
output.include_unclassified bool false Include commits with no category
output.include_merges bool false Include merge commits
output.include_files bool false Include file-level change detail
classification.rules_file path — Path to custom rules YAML/JSON
classification.use_llm bool false Enable LLM fallback tier
classification.llm_model string gpt-4o-mini LLM model identifier
classification.confidence_threshold float 0.7 Minimum acceptance confidence
classification.llm_fallback_threshold float 0.65 Commits with confidence above this value skip the LLM tier. Raised from 0.0 in 1.3.0; see LLM fallback threshold.
classification.llm_fallback_concurrency uint 8 Max concurrent LLM requests during fallback
github.token string $GITHUB_TOKEN GitHub PAT for PR fetch. Required scopes: public_repo for public repos, repo for private repos. Without a token, GitHub rate-limits anonymous traffic to 60 requests/hour and most PRs will be missed.
github.org string — Org slug for org-wide PR queries
github.repo string — Single repo slug (owner/name)
github.fetch_prs bool false Fetch pull request metadata. Must be true for tga pr-metrics to return data — when left at the default, tga collect writes zero rows to pull_requests (issue #211).
github.ticket_regex string — Override regex for detecting GitHub ticket refs in commit messages
jira.url string — JIRA base URL
jira.username string — JIRA API username (email for Cloud)
jira.token string — JIRA API token
jira.project_key string — Project key filter (e.g. API)
jira.ticket_regex string — Override regex for detecting JIRA ticket refs in commit messages
linear.ticket_regex string — Override regex for detecting Linear ticket refs in commit messages
pm.azure_devops.organization_url string — ADO org URL (e.g. https://dev.azure.com/myorg)
pm.azure_devops.pat string — Azure DevOps Personal Access Token
pm.azure_devops.project string — Default ADO project name
pm.azure_devops.fetch_on_reference bool false Fetch work items when AB#N refs appear in commits
pm.azure_devops.fetch_prs bool false Fetch ADO pull requests and reviewer data
pm.azure_devops.ticket_regex string AB#(\d+) Override regex for detecting ADO work item refs in commit messages
pm.bitbucket.workspace string — Bitbucket Cloud workspace slug
pm.bitbucket.repo_slug string — Repository slug within the workspace
pm.bitbucket.fetch_prs bool false Fetch Bitbucket Cloud pull request metadata
pm.bitbucket.token string $BITBUCKET_TOKEN Bearer token (App password or OAuth)
pm.bitbucket.username string — Atlassian account username for Basic auth
pm.bitbucket.app_password string — Atlassian App password for Basic auth (alternative to token)
cache.directory path — Cache directory (supports ~)
version string — Schema version; stored for compatibility
profile string — Named profile; stored for compatibility
dora.deployment_source string git_tags Source for tga deployments collect. One of git_tags, github_releases, github_actions, manual.
dora.deployment_tag_pattern regex ^v?[0-9]+\.[0-9]+\.[0-9]+(-...)?$ Tags matching this regex are ingested as deployments.
dora.production_branch string main Default branch that production deployments come from.
dora.failure_signals list [] One signal per entry; work_type (classification category) and/or commit_message_pattern (regex) + within_hours window.
dora.datadog_dir path — Directory of Datadog incident exports (.json). Currently a stub — JIRA SRE path is the default MTTR source.
jira.jira_project_mappings map {} Project key → work type (issue #206). Fires as Tier 1.6 — outranks the generic ticket regex.
jira.jira_project_mapping_confidence float 0.88 Per-verdict confidence for the JIRA mapping tier.

Paths support ~ expansion. Config files from the Python gitflow-analytics tool load without changes — unknown keys are silently ignored.

developer_aliases vs team.members

developer_aliases (Python-compatible flat map):

developer_aliases:
  "Alice Smith":
    - "alice@company.com"
    - "asmith@company.com"
    - "alice@personal.dev"
  "Bob Jones":
    - "bob@company.com"
    - "129991831+bobgithub@users.noreply.github.com"

team.members (structured roster with canonical email):

team:
  members:
    - name: Alice Smith
      email: alice@company.com
      aliases:
        - asmith@company.com
        - alice@personal.dev

When developer_aliases is non-empty it takes precedence over team.members. Use developer_aliases when migrating an existing Python config file; use team.members for new setups where canonical email matters for downstream tooling.

Example: multi-repo config with GitHub

See configs/example-config.yaml for a working example that covers multiple repositories, developer aliases, and CSV+Markdown output.

Common Workflows

First-time setup

# 1. Generate config.yaml interactively
tga install

# 2. Collect all history across all repos (full branch coverage, tga 2.0.0+)
tga collect

# 3. Classify all collected commits
tga classify

# 4. Generate reports
tga report

Routine weekly run

# Collect last 4 weeks (incremental; already-collected weeks are skipped)
tga collect --weeks 4

# Classify any unclassified commits
tga classify

# Regenerate reports
tga report

Scoped re-run for one service

# Re-collect only one repo, forcing re-walk of the last 8 weeks
tga collect --repos my-service --weeks 8 --force

# Re-classify only that repo
tga classify --force --repos my-service --weeks 8

# Reports always cover the full DB — no scoping needed
tga report

Upgrade from tga ≤ 1.5.4 (branch coverage fix)

# Re-collect all history to pick up commits from non-default branches
tga collect --force

# Re-classify everything (force overwrites prior verdicts)
tga classify --force

# Regenerate reports
tga report

Restrict collection to specific branches

# Walk only 'main' and 'release/1.0' branches for all repos
tga collect --branch main,release/1.0

# Walk only 'main' for one specific repo
tga collect --repos my-service --branch main

Monitor fetch health across many repos

tga fetches via libgit2 with full TLS support — https:// remotes (GitHub, GitLab, Bitbucket, etc.) work out of the box without a system git pre-fetch. On macOS the native Security framework is used; on Linux openssl is vendored statically; on Windows Schannel is used.

By default tga collect always fetches each repo before walking and prints an end-of-run summary:

Fetch summary: 116 / 118 repos updated (2 failure(s), 0 skipped)
  - ml_pricing_engine: could not find remote 'origin'
  - datapipelines: authentication required

# Use --strict-fetch in CI to fail the run if any fetch fails
tga collect --strict-fetch

# See per-repo success lines too (useful for debugging network topology)
tga collect --verbose-fetch

# Skip fetching entirely (stale data warning is printed to stderr)
tga collect --no-fetch

Per-repo fetch timeouts can be set in the YAML config to prevent a slow remote from blocking the entire run:

repositories:
  - path: ~/code/slow-repo
    name: slow-repo
    fetch_timeout_secs: 30   # optional; enforcement is pending a future release

Update rules and re-classify

# Preview which rules will fire (dry-run)
tga rules test "feat: add login page" --rules ./my-rules.yaml

# Re-classify all commits with the new rules (force overwrites existing verdicts)
tga classify --rules ./my-rules.yaml --force

# Scope re-classification to the last 4 weeks only
tga classify --rules ./my-rules.yaml --force --weeks 4

Fix existing database (backfill operations)

# Fix on_default_branch = 0 for pre-1.4.1 databases
tga backfill reachability

# Re-extract ticket IDs after extending ticket patterns
tga backfill ticket-ids

# Score historical commit effort (persists in fact_commit_effort)
tga backfill effort --repos my-service

# Preview revert-flag changes without writing
tga backfill revert-flags --dry-run

DORA metrics pipeline

# Ingest deployment events (default: git tags matching dora.deployment_tag_pattern)
tga deployments collect

# Ingest incidents from JIRA SRE issues or Datadog
tga incidents collect

# Compute and display the four DORA metrics
tga dora

# Limit to events since the start of Q2
tga dora --since 2026-04-01

CLI Subcommand Reference

Subcommand Purpose Key flags
tga analyze Full pipeline (collect → classify → report) in one command --weeks, --from, --to, --force, --skip-collect, --skip-classify, --dry-run
tga collect Stage 1: extract commits into the database --repos, --branch, --weeks, --from, --to, --force, --head-only, --no-fetch, --strict-fetch, --verbose-fetch, --dry-run
tga classify Stage 2: classify collected commits --repos, --weeks, --since, --until, --force, --rules, --use-llm, --no-external
tga report Stage 3: generate CSV/JSON/Markdown reports --output, --formats, --author
tga author Per-engineer drill-down (commits, effort, PRs, categories) <email>, --format, --since, --until
tga pr-metrics Pull-request metrics per engineer --weeks, --csv, --output
tga install Interactive first-time config wizard --output, --force
tga aliases Manage developer identity aliases list, merge <src> <dst>, add-login <email> <provider> <login>
tga backfill Retroactive maintenance on existing rows --repos, --weeks, --since, --until, --dry-run
tga override Pin classification verdicts (Tier 0) add, list, remove
tga rules Introspect the active rule set list, show <sha>, test "<message>"
tga deployments DORA: ingest deployment events collect [--source]
tga incidents DORA: ingest production incidents collect [--source]
tga dora DORA: compute and display all four metrics --since

CLI Reference

All subcommands accept these global flags:

Flag Default Description
--config <PATH> config.yaml Path to config YAML
--database <PATH> tga.db Path to SQLite database
-v / -vv / -vvv warnings only Increase log verbosity

tga analyze

Run the full pipeline: collect → classify → report.

tga analyze [--config <PATH>] [--database <PATH>] [--output <DIR>]
            [--skip-collect] [--skip-classify] [--weeks <N>]
            [--from <DATE>] [--to <DATE>] [--dry-run]
            [--validate-only] [--no-validate]
Flag Description
--skip-collect Skip Stage 1; use commits already in the database
--skip-classify Skip Stage 2; use existing classifications
--output <DIR> Override output.directory from config
--weeks <N> Limit collection to the last N weeks (overrides config start_date)
--from <DATE> Start date for collection (ISO 8601 YYYY-MM-DD); mutually exclusive with --weeks
--to <DATE> End date for collection (ISO 8601 YYYY-MM-DD); defaults to today
--dry-run Perform all steps against an in-memory database; the on-disk database is left untouched
--validate-only Run configuration validation and exit (0 on success, 1 on errors)
--no-validate Skip pre-flight configuration validation 
# Full pipeline
tga analyze --config config.yaml

# Re-run reports only (commits already collected and classified)
tga analyze --skip-collect --skip-classify --output ./reports-v2

tga collect

Stage 1: extract commits from git repositories into the database.

tga collect [--config <PATH>] [--database <PATH>]
            [--repos <NAME,...>] [--branch <NAME[,NAME…]>]
            [--from <DATE>] [--to <DATE>] [--weeks <N>]
            [--since <DATE>] [--until <DATE>] [--dry-run]
            [--force-refresh-prs] [--validate-only] [--no-validate]
            [--head-only]

Branch coverage (tga 2.0.0+): By default tga collect walks ALL local branches (refs/heads/*) and remote tracking refs (refs/remotes/origin/*), so commits on PR branches, feature branches, and hotfixes are not silently excluded. This fixes a data-integrity bug (#331) that was losing ~56% of commits in multi-branch repos when using the HEAD-only default from tga ≤ 1.5.4.

To restore the legacy HEAD-only behaviour, pass --head-only globally or set head_only: true on individual repository entries in your YAML config.

Migration from tga ≤ 1.5.4: existing databases are missing commits from non-default branches. Re-collect with --force to recover that history:

tga collect --force   # re-walk all weeks with full branch coverage
Flag Description
--repos <NAME,...> Comma-separated list of repository names to collect; others are skipped
--branch <NAME[,NAME…]> Restrict the revwalk to these branch names (seeds from both refs/heads/<name> and refs/remotes/origin/<name>); mutually exclusive with --head-only
--from <DATE> Collect commits on or after this ISO 8601 date; mutually exclusive with --weeks
--to <DATE> Collect commits on or before this ISO 8601 date; defaults to today
--since <DATE> Legacy alias for --from (Python-predecessor compatibility); --from takes precedence
--until <DATE> Legacy alias for --to (Python-predecessor compatibility); --to takes precedence
--weeks <N> Limit collection to the last N weeks; --weeks takes precedence over --from/--to
--dry-run Run collection against an in-memory database; the on-disk database is left untouched
--force-refresh-prs Re-fetch ADO pull requests even when already cached (backfills pre-v1.0.9 rows)
--validate-only Run configuration validation and exit (0 on success, 1 on errors)
--no-validate Skip pre-flight configuration validation
--no-fetch Skip the pre-walk git fetch; use only local refs. A warning is printed to stderr so you know the data may be stale.
--strict-fetch Exit non-zero if any repository's fetch fails (default: failures are shown in the summary but exit code is 0). Useful for CI.
--verbose-fetch Print a success line per fetched repository in the fetch summary (default: only failures are shown).
--head-only Restore legacy HEAD-only revwalk (tga ≤ 1.5.4 behaviour); mutually exclusive with --branch; also available per-repo via head_only: true in YAML 
tga collect --repos my-project --from 2024-01-01 --to 2024-03-31
tga collect --weeks 4           # collect last 4 weeks across all repos
tga collect --force             # re-collect all history (e.g. after upgrading to 2.0.0)
tga collect --branch main       # walk only the main branch for all repos
tga collect --branch main,release/1.0 --repos my-service  # branch + repo filter
tga collect --head-only         # legacy HEAD-only walk for all repos

tga classify

Stage 2: run the classification cascade over collected commits.

tga classify [--config <PATH>] [--database <PATH>]
             [--repos <NAME,...>] [--weeks <N>] [--since <DATE>] [--until <DATE>]
             [--force] [--rules <PATH>] [--use-llm] [--backfill-complexity]
             [--no-external]
Flag Description
--repos <NAME,...> Limit re-classification to these repository names (comma-separated)
--weeks <N> Limit to commits in the last N ISO weeks; mutually exclusive with --since/--until
--since <DATE> Lower bound on author timestamp (ISO 8601 YYYY-MM-DD); mutually exclusive with --weeks
--until <DATE> Upper bound on author timestamp (ISO 8601 YYYY-MM-DD); mutually exclusive with --weeks
--force Re-classify commits that already have a verdict (useful after updating rules)
--rules <PATH> Override classification.rules_file from config. Custom rules default to priority 110 (above built-in 100) and are standalone by default (extend_defaults: false).
--use-llm Enable LLM fallback regardless of config setting
--backfill-complexity Fill missing complexity scores (1–5) for already-classified commits via the LLM, without re-running the full cascade; category, confidence, and method are left untouched
--no-external Disable all external classification sources (JIRA, GitHub Issues) for this run, even if configured in the rules file or config 
tga classify --rules ./custom-rules.yaml --use-llm
tga classify --no-external                         # offline / CI mode
tga classify --force --repos my-service --weeks 4  # re-classify one repo, last 4 weeks
tga classify --force --since 2026-01-01            # re-classify from a date forward

tga report

Stage 3: generate reports from classified commits.

tga report [--config <PATH>] [--database <PATH>]
           [--output <DIR>] [--formats <FMT,...>]
           [--author <EMAIL>]
Flag Description
--output <DIR> Override output.directory from config
--formats <FMT,...> Comma-separated: csv, json, markdown
--author <EMAIL> Scope report to one canonical identity (matches canonical_email case-insensitively) 
# Full team report
tga report --output ./q1-reports --formats csv,json

# Single-engineer report
tga report --author alice@example.com --output ./alice-q1

# List available canonical identities
tga aliases list

If --author is supplied but the email is not in the authors table, tga report exits non-zero with a suggestion to run tga aliases list.

tga pr-metrics

Aggregate pull-request metrics per engineer from the pull_requests cache.

tga pr-metrics [--config <PATH>] [--database <PATH>]
               [--weeks <N>] [--csv] [--output <PATH>]
Flag Description
--weeks <N> Limit metrics to PRs created within the last N weeks
--csv Emit CSV instead of an aligned text table
--output <PATH> Write output to a file (CSV with --csv, otherwise the text table) 
tga pr-metrics --weeks 12 --csv --output pr-metrics.csv

The pr_comments_given and avg_revisions columns are reserved for future use and currently always output 0.0; the underlying review-comment and revision-count data is not yet tracked.

tga backfill

Retroactive maintenance operations that update existing commit rows in place (outside the normal collect → classify → report pipeline). All global filter flags (--repos, --weeks, --since, --until) scope the operation uniformly.

Note: --branch is collect-only. Commits in the database do not carry branch attribution after collection, so there is no branch filter for backfill operations.

tga backfill <SUBCOMMAND> [--dry-run]
             [--repos <NAME,...>] [--weeks <N>] [--since <DATE>] [--until <DATE>]
Subcommand Description
ai-detection Re-run LLM classification on low-confidence prior LLM verdicts
revert-flags Scan commit messages for revert patterns and set is_revert
ticket-ids Scan commit messages for ticket refs and update ticket_id/ticketed
reachability Re-run the full reachability scan and upsert fact_commit_reachability without re-collecting (fixes on_default_branch=0 rows in existing databases, issue #290)
effort Compute empirical effort scores (XS/S/M/L/XL) for historical commits and persist in fact_commit_effort
Global flag Description
--dry-run Report how many rows would change without writing
--repos <NAME,...> Limit to these repository names (comma-separated)
--weeks <N> Limit to commits in the last N ISO weeks; mutually exclusive with --since/--until
--since <DATE> Lower bound on author timestamp (ISO 8601 YYYY-MM-DD); mutually exclusive with --weeks
--until <DATE> Upper bound on author timestamp (ISO 8601 YYYY-MM-DD); mutually exclusive with --weeks 
tga backfill revert-flags --dry-run
tga backfill ticket-ids
tga backfill reachability                      # fix on_default_branch for all repos
tga backfill reachability --repos my-service   # single repo only
tga backfill effort --repos api --weeks 4      # effort for one repo, last 4 weeks
tga backfill ticket-ids --since 2026-01-01     # ticket IDs from a date forward

tga override

Manage manual classification overrides (Tier 0). Rows here pin a commit's verdict regardless of what the rule-based or LLM tiers would produce.

tga override <SUBCOMMAND>
Subcommand Description
add <SHA> <WORK_TYPE> <CHANGE_TYPE> [--notes <TEXT>] [--repo <PATH>] Insert (or replace) an override row for a commit SHA
list [--repo <PATH>] List every override row, optionally filtered by repository
remove <SHA> [--yes] Delete the override row(s) for a SHA (--yes skips confirmation) 
tga override add abc1234 feature feature --notes "manual review"
tga override list
tga override remove abc1234 --yes

tga rules

Introspect the classification rule set (issue #209). Useful for tuning rules and answering "why was this commit classified as X?".

tga rules list                       # every loaded rule, sorted by priority
tga rules show <commit-sha>          # verdict + method recorded for a commit
tga rules test "feat: add login"     # dry-run the cascade against a message

tga deployments / tga incidents / tga dora

DORA metrics infrastructure (issues #207, #208, #212, #213).

# Step 1: ingest deployment events (default: git tags matching dora.deployment_tag_pattern)
tga deployments collect

# Step 2 (optional): ingest production incidents from JIRA SRE issues
tga incidents collect

# Step 3: compute Deployment Frequency, Lead Time, Change Failure Rate, MTTR.
# Also rebuilds the `deployment_failures` derived join from the current
# `dora.failure_signals` config — safe to re-run after a config edit.
tga dora

The four DORA metrics land in pre-computed SQL views (v_deployment_frequency, v_lead_time, v_change_failure_rate, v_mttr) so dashboards can read them directly without re-aggregating.

Tag & Release-Branch Reachability (issues #279, #290)

tga collect automatically populates fact_commit_reachability with five columns that distinguish "deployed via cherry-pick to a release branch and tagged" from "abandoned WIP":

Column Type Meaning
on_default_branch boolean true if the commit is reachable from the repo's default branch (main/master)
on_any_tag boolean true if any git tag reaches this commit
reachable_from_tags JSON array Tag names that contain this commit
on_release_branch boolean true if commit is on any configured release branch
release_branches JSON array Matching release-branch names

on_default_branch fix (1.4.1 — issue #290): Prior to 1.4.1, on_default_branch was declared in the schema but never populated — every row had on_default_branch = 0. tga collect now auto-detects each repo's default branch via refs/remotes/origin/HEAD (the symref set by git clone), falling back through refs/heads/main, refs/heads/master, refs/remotes/origin/main, refs/remotes/origin/master. If none of these refs exist for a repo, a warning is logged and on_default_branch stays 0 for that repo's commits. To fix existing databases without re-running the full collection pipeline, see tga backfill reachability below.

This resolves a systematic blind spot: bug fixes and security patches cherry-picked to release/* branches, tagged for production, and never merged to main previously showed on_default_branch = false — making them indistinguishable from abandoned WIP. With this feature, the 32% "merged rate" for bug fixes turns out to be much higher when deployed-via-tag commits are counted.

Config:

reachability:
  track_tags: true               # default: true
  track_release_branches: true   # default: true
  release_branch_patterns:       # default: ["release/*", "hotfix/*", "chore/release-*", "v*"]
    - "release/*"
    - "hotfix/*"
    - "chore/release-*"
    - "v*"

Set track_tags: false to skip the tag scan (useful for trunk-based repos with thousands of tags). The default is true because the scan is O(repo + refs), not O(repo × refs × commits).

Optionally disable the scan for a single run:

tga collect --skip-tag-reachability

Useful derived queries:

-- What % of bug fixes actually shipped (via any path)?
SELECT
  COUNT(*) FILTER (WHERE on_default_branch OR on_any_tag OR on_release_branch)
  * 100.0 / COUNT(*) AS shipped_pct
FROM classifications cl
JOIN commits c ON c.classification_id = cl.id
JOIN fact_commit_reachability fcr ON fcr.commit_sha = c.sha
WHERE cl.category = 'bug_fix';

-- Cherry-pick rate: commits reachable from tags but not on main
SELECT repo, COUNT(*)
FROM fact_commit_reachability
WHERE on_any_tag AND NOT on_default_branch
GROUP BY repo;

Pipeline Architecture

git repos ──┐
             │   collect      SQLite (tga.db)   classify       SQLite    report
GitHub API ──┼──────────────► [commits]        ──────────────► [classif]─────────► CSV (×9)
JIRA API ────┤   (libgit2,    [authors]         (7-tier                 ► JSON
Linear API ──┤   reqwest)     [pull_requests]   cascade,                ► Markdown
ADO API  ────┘                [work_items]      Rayon-parallel)

Stage 1 — collect (tga::collect): opens each repository with libgit2, walks the configured branch, extracts commit metadata and diff stats, resolves author identities, fetches GitHub PR / JIRA issue / Linear / Azure DevOps work item metadata via REST/GraphQL, and writes everything to SQLite.

Stage 2 — classify (tga::classify): reads unclassified commits from the database, runs each message through the classification cascade (see below), and writes a classification verdict back. Rule-based tiers execute in parallel via Rayon.

Stage 3 — report (tga::report): reads the classified database, aggregates per-author, per-week, DORA, velocity, and quality statistics, and writes the configured output formats to the output directory.

Classification

Classification Cascade

Each commit message is tested against tiers in order. The first tier to produce a confident result wins.

Tier 0 — Manual Override (confidence 1.0): looks up the (commit_hash, repo_path) pair in the classification_overrides table. Managed via tga override add|list|remove.

Tier 1.5 — Issue Type (confidence 0.90): when the commit has ticket references resolving to rows in issue_cache, maps the upstream issue type (bug, story, task, spike, etc.) directly to a change_type.

Tier 1.6 — JIRA Project Mapping (default confidence 0.88, override via jira.jira_project_mapping_confidence): when jira.jira_project_mappings is configured, maps the JIRA project key prefix of any [A-Z]+-\d+ reference to a change_type. Fires before the regex tier so project mappings outrank the generic jira-ticket regex rule (issue #206). Example config:

jira:
  jira_project_mappings:
    TQL: bug_fix
    APEX: integration
    INFRA: platform_infrastructure
    SEC: security
  jira_project_mapping_confidence: 0.88   # optional, default 0.88

Tier-0 manual overrides and exact-keyword conventional-commit prefixes (e.g. fix:) still beat this tier.

Tier 1 — Exact (Aho-Corasick): builds a single finite-state machine from every keyword list across every rule and scans the message in O(n) time. Matches feat:, fix:, chore:, etc. Confidence 0.85–0.95.

Tier 2 — Regex: applies pre-compiled regex patterns from the rule set. Handles anchored conventional-commit patterns (^feat(\([^)]*\))?!?:) and JIRA ticket IDs (\b[A-Z][A-Z0-9]+-\d+\b).

Tier 2.5 — Weighted Sum (new in 1.3.0): composes five independent signals into per-category scores and emits a verdict when the argmax score meets the minimum confidence threshold (default 0.55). Active for all commits including when extend_defaults: false. See Weighted-Sum Tier.

Tier 3 — Fuzzy heuristics: detects merge commits (via is_merge flag or Merge pull request prefix) and reverts (via Revert prefix). No external dependencies. Suppressed when extend_defaults: false.

Tier 7 — LLM fallback (optional, async): calls an OpenAI-compatible API (OpenRouter by default, AWS Bedrock behind the bedrock cargo feature) when tiers 0–3 leave a commit below the fallback threshold. Disabled by default; enable with analysis.llm_classification.enabled: true or --use-llm. Results are only accepted when confidence >= confidence_threshold (default 0.7). See LLM fallback threshold.

Default Rules

ID Category Keywords / Patterns
cc-feat feature feat:, feature:, ^feat(...)?!?:
cc-fix bugfix fix:, bugfix:, hotfix, ^fix(...)?!?:
cc-chore chore chore:, ^chore(...)?!?:
cc-docs documentation docs:, doc:, ^docs?(...)?!?:
cc-refactor refactor refactor:, ^refactor(...)?!?:
cc-test test test:, tests:, ^tests?(...)?!?:
cc-ci ci ci:, ^ci(...)?!?:
cc-perf performance perf:, ^perf(...)?!?:
cc-style style style:, ^style(...)?!?:
cc-build build build:, ^build(...)?!?:
cc-revert revert revert:, ^revert(...)?!?:
breaking-change breaking breaking change, breaking-change
jira-ticket feature (ticketed) \b[A-Z][A-Z0-9]+-\d+\b
kw-bug bugfix bug, defect
kw-security bugfix (security) security, cve-, vulnerability

Commits that match no rule are assigned category uncategorized with confidence 0.0.

Custom Rules File

Supply your own rules as a YAML (or JSON) file:

# my-rules.yaml
version: "1.0"
extend_defaults: false   # true = merge with built-ins; false = standalone (default)
rules:
  - id: my-deploy
    category: deployment
    keywords:
      - "deploy:"
      - "release:"
    patterns:
      - "(?i)^deploy(ment)?:"
    priority: 80
    confidence: 0.9

tga classify --rules ./my-rules.yaml
# or in config.yaml:
# classification:
#   rules_file: ./my-rules.yaml

Rules YAML schema

Each entry under rules: is a Rule with these fields:

Field Type Default Description
id string required Unique rule identifier (used in logs and tga rules test)
category string required The classification label written to the database
subcategory string null Optional leaf label (e.g. "sql-injection" under category: "security")
keywords list of strings [] Exact substrings, matched case-insensitively via Aho-Corasick
pattern / patterns string or list [] Regex patterns (see note below)
priority integer 110 Higher = checked first. Custom-rule default (110) beats built-ins (peak 115 for cc-revert, 100 for cc-feat/cc-fix)
confidence float 0–1 0.85 Score attached to the verdict; only accepted if ≥ confidence_threshold

pattern: vs patterns: — both key names are accepted:

# Singular string (most natural for a single regex)
pattern: "(?i)^feat[:(]"

# Plural list (multiple alternates for one rule)
patterns:
  - "(?i)^feat[:(]"
  - "(?i)^feature[:(]"

Using pattern: with a single string is exactly equivalent to patterns: with a single-element list. Both forms were silently mishandled in versions ≤ 1.2.0 (issue #259); if you had rules that never fired, this was the cause.

The top-level RuleSet also has two optional fields:

Field Type Default Description
version string null Schema version tag (informational)
extend_defaults bool false When true, custom rules are merged with the built-in set; same-id entries override the built-in

Weighted-Sum Tier 2.5 (new in 1.3.0)

The weighted-sum tier is a lightweight ensemble that sits between the regex tier and the fuzzy tier. It was introduced in 1.3.0 (issue #270) after evaluating fuzzy-logic crates — none of them provided sufficient recall improvement over the stepped keyword approach at acceptable binary size cost.

How it works

Five signals are evaluated independently and their scores are summed per category. The category with the highest total score wins, provided the score meets the min_confidence threshold (default 0.55). Confidence is clamped to [min_confidence, 0.95].

Signal How scored
Keyword density 0 matches → 0.0; 1 match → 0.40; 2 matches → 0.60; 3+ matches → 0.75
Ticket prefix Uniform +0.05 when a PROJ-123 style prefix is detected
Message length Short (<12 chars): boosts KTLO/Merge (+0.10), Maintenance (+0.05); suppresses Feature (−0.05). Long (>80 chars): boosts Feature/PlatformWork (+0.10), Maintenance/Bugfix (+0.05).
Merge indicator Dedicated weight vector fires when is_merge=true or message starts with "merge…"
File paths Test files → Maintenance/Bugfix signal; docs/markdown → Content signal; manifests → KTLO/Maintenance signal

If two or more categories share the exact same top score, the tier returns no verdict (tie-breaking falls through to fuzzy or LLM).

The tier is active even when extend_defaults: false. Unlike the fuzzy tier, which emits fixed built-in category strings (merge, feature, chore), the weighted-sum tier picks its verdict dynamically by signal composition and does not embed built-in category knowledge directly.

Configuration

classification:
  weighted_sum:
    enabled: true         # default: true. Set false to disable this tier entirely.
    min_confidence: 0.55  # default: 0.55. Minimum score to emit a verdict.

To disable just this tier without touching anything else:

classification:
  weighted_sum:
    enabled: false

LLM Fallback Threshold Migration

Breaking change in 1.3.0: classification.llm_fallback_threshold default raised from 0.0 to 0.65.

At 0.0 (the old default), the LLM fallback would only fire when every deterministic tier returned confidence exactly 0.0 — effectively never. This made use_llm: true a no-op for most real-world commit messages. The new default of 0.65 routes any deterministic verdict below 0.65 to the LLM when use_llm: true.

Affected setups: configs that set use_llm: true (or --use-llm) and relied on the LLM never re-classifying commits that the deterministic tiers had already classified with confidence below 0.65. In practice this affects configs where the fuzzy tier's outputs (0.40 for short chore messages, 0.60 for bare ticket refs) were intentionally accepted as final.

Migration options:

Option 1 — restore the old behavior explicitly:

classification:
  llm_fallback_threshold: 0.0   # pin to old default

Option 2 — accept the new default and tune upward if the LLM fires too often:

classification:
  llm_fallback_threshold: 0.70  # only route very low-confidence verdicts to LLM

Users with use_llm: false (the default) are not affected by this change.

Rule Priority and extend_defaults (issue #259)

Custom rules default to priority 110 — one step above the highest built-in rule priority (100). This means user rules win over the default ruleset without needing an explicit priority: entry in every rule.

Custom rule files also default to standalone mode (extend_defaults: false): only the rules in the file are applied. Opt in to merging with the built-in defaults by adding extend_defaults: true.

Fuzzy-tier gating (1.2.2): the fuzzy heuristic tier (which emits the built-in category strings merge, feature, chore) is suppressed when extend_defaults: false. This aligns with the principle that extend_defaults: false means "no built-in hardcoded classification." If you see method=fuzzy_match rows for a config with extend_defaults: false, upgrade to 1.2.2.

Weighted-sum tier and extend_defaults (1.3.0): unlike the fuzzy tier, the weighted-sum tier (Tier 2.5) remains active even with extend_defaults: false. It picks its verdict via signal composition rather than emitting fixed built-in strings, so disabling it requires classification.weighted_sum.enabled: false.

# my-rules.yaml — standalone by default (no built-in rules loaded)
extend_defaults: false   # optional: explicitly document the default
rules:
  - id: my-deploy
    category: deployment
    keywords: ["deploy:", "release:"]
    # priority defaults to 110 — beats the built-in cc-feat (100)

# my-addons.yaml — augments the defaults
extend_defaults: true
rules:
  - id: my-payments
    category: payments
    keywords: ["payment:", "billing:"]
    confidence: 0.92

Diagnosing rules that never fire: if your custom categories never appear in reports, run:

tga rules list --rules ./my-rules.yaml   # shows all loaded rules with their matchers
tga rules test "feat: add login" --rules ./my-rules.yaml   # dry-run a message
tga classify -vv --rules ./my-rules.yaml   # verbose mode shows per-rule decisions

A WARN rule has no keywords or patterns in the log output means a rule's YAML key was silently dropped — the most common cause is writing pattern: (singular) in a version before 1.2.1. Upgrade to 1.2.1+ and both pattern: and patterns: are accepted.

See also: Multi-source classification for external ticket system integration.

Multi-source classification (issue #260)

tga classify can consult external ticket systems — JIRA Cloud/Server and GitHub Issues — as high-confidence classification signals before the commit-message rule tiers run.

External sources fire as Tier 0.5 — after manual overrides, before commit-message rules — with a fixed confidence of 0.92. They are enabled by default when configured; use --no-external to opt out for a run.

Priority model (highest to lowest):

  1. Manual overrides (tga override add, confidence 1.0)
  2. External sources — JIRA issue type / GitHub Issues labels (confidence 0.92)
  3. Custom commit-message rules (default priority 110)
  4. Built-in TGA commit-message rules (priority 100)
  5. LLM fallback (when use_llm: true)

Configuration — add a sources: list to config.yaml under classification::

classification:
  no_external: false    # opt-out flag; sources fire by default once configured
  sources:

    # --- JIRA source ---
    - type: jira
      base_url: "https://yourco.atlassian.net"
      token_env: JIRA_API_TOKEN      # name of the env var holding your API token
      email_env: JIRA_USER_EMAIL     # optional — only needed for JIRA Cloud Basic auth
      project_keys: ["PROJ", "ENG"]  # empty list = match all projects
      field_mappings:
        # Maps JIRA issue type → tga category. Case-sensitive on the JIRA side.
        issue_type:
          Bug: bug_fix
          Story: new_feature
          Task: tech_debt_refactoring
          Sub-task: tech_debt_refactoring
          Epic: new_feature
        # Maps JIRA labels → tga category. Labels are case-sensitive.
        labels:
          ktlo: tech_debt_refactoring
          security: security
        # Maps JIRA components → tga category.
        components:
          "CI/CD": devops
          Platform: platform_infrastructure

    # --- GitHub Issues source ---
    - type: github_issues
      repo: "acme/widgets"           # "owner/repo" slug
      token_env: GITHUB_TOKEN        # name of the env var holding your PAT
      label_mappings:
        bug: bug_fix
        enhancement: new_feature
        dependencies: tech_debt_refactoring
        documentation: documentation
        security: security

Field-mapping precedence within a JIRA source (highest to lowest): issue_type → labels → components. The first mapping that matches wins.

Credential handling: tokens are read from the named environment variables at runtime — never store them directly in config files:

export JIRA_API_TOKEN="your-jira-api-token"
export JIRA_USER_EMAIL="you@yourco.com"   # JIRA Cloud only
export GITHUB_TOKEN="your-github-pat"
tga classify --config config.yaml

Failure mode: a missing token, HTTP 401/404, or network error causes the source to be skipped with a WARN log and the pipeline falls through to commit-message rules for affected commits. Classification never panics on source failures.

Caching: each unique ticket is fetched at most once per tga classify run (in-memory dedup, never persisted to disk). A 15 000-commit run against a project with 500 unique JIRA tickets makes 500 API calls, not 15 000.

Disabling external sources: use --no-external to skip all sources for a run (useful in CI or offline environments):

tga classify --no-external

See examples/multi-source-config.yaml for a fully annotated example.

# --- Linear source (new in 1.4.0) ---
- type: linear
  api_key_env: LINEAR_API_KEY         # Personal API Key (not OAuth)
  team_keys: ["ENG", "INFRA"]         # Empty = match all Linear teams.
                                      # Use this to disambiguate from JIRA keys
                                      # when both follow the PROJ-NNN pattern.
  field_mappings:
    issue_type:
      Bug: bug_fix
      Feature: new_feature
      Improvement: tech_debt_refactoring
    labels:
      security: security
      ktlo: tech_debt_refactoring
    cycle:
      "Sprint 42": new_feature        # active cycle/sprint → category

# --- Shortcut source (new in 1.4.0) ---
- type: shortcut
  api_token_env: SHORTCUT_API_TOKEN
  field_mappings:
    story_type:
      bug: bug_fix
      feature: new_feature
      chore: tech_debt_refactoring
    labels:
      security: security
    workflow_state:
      "In Progress": new_feature      # state name → category

# --- Confluence source (new in 1.4.0) ---
# Note: Confluence confidence is fixed at 0.80 (lower than the default 0.92).
# References are extracted from /wiki/spaces/.../pages/<id> URLs and
# Smart Commit tags ([CONF-<id>] in commit messages).
- type: confluence
  base_url: "https://yourco.atlassian.net"
  token_env: CONFLUENCE_API_TOKEN
  email_env: CONFLUENCE_EMAIL         # Confluence Cloud Basic auth email
  label_mappings:
    runbook: devops
    architecture: tech_debt_refactoring
    security: security

# --- Datadog source (new in 1.4.0) ---
# Matches commit SHAs (7–40 hex chars) in commit messages against the
# Datadog Events API deployment events. Confidence is fixed at 0.95.
- type: datadog
  api_key_env: DD_API_KEY
  app_key_env: DD_APP_KEY
  site: datadoghq.com                 # optional, default "datadoghq.com"
  default_category: devops            # category written when a deployment
                                      # event is found (default: "devops")
  lookback_days: 30                   # how far back to search events (default: 30)


**Linear key disambiguation**: Linear keys and JIRA keys share the same
`PROJ-NNN` shape. Set `team_keys` on the Linear source to the exact team
prefixes you use — keys that do not match any configured prefix are silently
skipped by the Linear source (and may still be picked up by JIRA if that
source is also configured).

**Shortcut reference formats**: two forms are recognized automatically:
- `[ch1234]` — bracket form (in commit message body or subject)
- `sc-1234` — sc-prefix form (common in branch names carried into messages)

**Confluence reference formats**: two forms are recognized:
- `/wiki/spaces/MYSPACE/pages/12345` — full URL (page ID extracted)
- `[CONF-12345]` — Smart Commit tag (page ID after the hyphen)

**Credential handling for new sources**:

```bash
export LINEAR_API_KEY="lin_api_..."  # pragma: allowlist secret
export SHORTCUT_API_TOKEN="..."      # pragma: allowlist secret
export CONFLUENCE_API_TOKEN="..."    # pragma: allowlist secret
export CONFLUENCE_EMAIL="you@yourco.com"
export DD_API_KEY="..."              # pragma: allowlist secret
export DD_APP_KEY="..."              # pragma: allowlist secret
tga classify --config config.yaml

See examples/multi-source-config.yaml for a fully annotated example covering all six sources.

See also: Rules YAML schema for commit-message rule authoring.

Operational Notes

Fixing on_default_branch = 0 for existing databases (1.4.1 — issue #290)

If your tga.db was built before 1.4.1, every row in fact_commit_reachability has on_default_branch = 0 because the column was never populated. You can fix this without re-running the full collection pipeline (which can take 20+ minutes on large corpora):

# Verify the problem (sum should be 0 on a pre-1.4.1 DB)
sqlite3 tga.db "SELECT SUM(on_default_branch), COUNT(*) FROM fact_commit_reachability;"

# Fix in-place — upserts all five reachability columns, no row deletion
tga backfill reachability

# Verify the fix (sum should now be > 0 for repos with main/master)
sqlite3 tga.db "SELECT SUM(on_default_branch), SUM(on_any_tag), COUNT(*) FROM fact_commit_reachability;"
sqlite3 tga.db "SELECT repo, SUM(on_default_branch) FROM fact_commit_reachability GROUP BY repo LIMIT 10;"

The backfill is idempotent — running it multiple times produces identical results and never deletes rows. Only the five reachability columns are touched.

Auto-detection logic: for each repository, tga tries (in order):

  1. refs/remotes/origin/HEAD (symref — set by git clone, points at whatever the remote considers its default branch)
  2. refs/heads/main
  3. refs/heads/master
  4. refs/remotes/origin/main
  5. refs/remotes/origin/master

If none of the above resolve to a commit, a WARN is logged and on_default_branch remains 0 for that repo.

SQLite WAL durability (fixed in 1.4.0 — issue #298)

tga opens its database in WAL (Write-Ahead Log) mode for concurrent-read performance. Prior to 1.4.0, the WAL was never explicitly checkpointed on exit, so the main .db file could lag behind the -wal sidecar. On a large corpus this meant up to 22 minutes of classification work was held only in the WAL and could be lost if the file was copied or the process was killed.

1.4.0 fix: tga classify now calls PRAGMA wal_checkpoint(TRUNCATE) on clean exit, flushing all WAL frames into the main database file and zeroing the WAL. In addition, a periodic PRAGMA wal_checkpoint(PASSIVE) fires every classification.checkpoint_every commits (default 0 = disabled) to cap the data-loss window during long runs.

classification:
  checkpoint_every: 1000  # PASSIVE checkpoint every 1000 classified commits
                          # (0 = off, which is the default)

The checkpoint is non-fatal: if it fails (e.g., SQLITE_CORRUPT), tga logs a WARN and continues. You can manually flush at any time:

sqlite3 tga.db 'PRAGMA wal_checkpoint(TRUNCATE)'

Output Formats

CSV

Nine CSV files are written when csv is in the format list (per-author, weekly activity, DORA metrics, velocity, quality, and related breakdowns). The two most commonly used are documented below.

authors.csv — one row per author:

Column Description
name Canonical author name
email Canonical author email
commit_count Total commits
insertions Total lines added
deletions Total lines deleted
files_changed Total files changed
first_commit ISO 8601 timestamp of earliest commit
last_commit ISO 8601 timestamp of most recent commit

weekly_activity.csv — one row per week/author/repository bucket:

Column Description
week ISO week label, e.g. 2024-W03
author Author name
repository Repository name
commit_count Commits in this bucket
insertions Lines added in this bucket
deletions Lines deleted in this bucket

JSON

Four JSON files are written when json is in the format list: report.json, velocity_summary.json, quality_summary.json, and dora_summary.json. The primary one is documented below.

report.json — full structured payload:

{
  "generated_at": "2024-03-15T10:00:00Z",
  "period_start": "2024-01-01T00:00:00Z",
  "period_end":   "2024-03-14T23:59:59Z",
  "total_commits": 347,
  "total_authors": 8,
  "category_breakdown": { "feature": 120, "bugfix": 45, ... },
  "authors": [
    {
      "name": "Alice Smith",
      "email": "alice@company.com",
      "commit_count": 87,
      "insertions": 4200,
      "deletions": 1100,
      "files_changed": 310,
      "categories": { "feature": 50, "bugfix": 20, ... },
      "first_commit": "...",
      "last_commit": "..."
    }
  ],
  "repositories": [
    {
      "name": "my-project",
      "commit_count": 347,
      "author_count": 8,
      "insertions": 18000,
      "deletions": 6000,
      "top_categories": [["feature", 120], ["bugfix", 45]]
    }
  ],
  "weekly_activity": [
    {
      "week": "2024-W03",
      "author": "Alice Smith",
      "repository": "my-project",
      "commit_count": 12,
      "insertions": 500,
      "deletions": 120,
      "categories": { "feature": 8, "bugfix": 4 }
    }
  ]
}

Markdown

report.md — a narrative report containing a summary header, per-author commit table, category breakdown, and weekly activity section. Suitable for pasting into Confluence or a PR description.

Development

Build and Test

# Build everything
cargo build

# Build release binary
cargo build --release

# Run all tests
cargo test

# Lint (zero warnings required)
cargo clippy -- -D warnings

# Format check (CI gate)
cargo fmt --check

# Auto-format
cargo fmt

# Generate rustdoc
cargo doc --open

Running Against Real Repos

configs/example-config.yaml is a working example that analyzes repositories using developer_aliases. Copy it, adjust paths and names to match your setup, then run:

tga analyze --config configs/example-config.yaml --database tga.db

CI Gates

The GitHub Actions workflow (ci.yml) requires:

  • cargo fmt -- --check
  • cargo clippy --all-targets -- -D warnings
  • cargo test
  • cargo doc --no-deps with RUSTDOCFLAGS="-D warnings"

Crate Structure

Single tga crate (consolidated from the original 5-crate workspace):

Module Path Purpose
tga::core src/core/ Shared types, config, DB schema, migrations, error types
tga::collect src/collect/ Stage 1: git extraction (libgit2), GitHub/JIRA/Linear/ADO clients, PmAdapter trait
tga::classify src/classify/ Stage 2: seven-tier classification cascade
tga::report src/report/ Stage 3: CSV/JSON/Markdown output
commands (binary-private) src/commands/ Subcommand handlers wired into src/main.rs

Performance

Benchmarked against the Duetto production corpus (72,608 commits, 20 JIRA projects) on Apple M4 Max (128 GB RAM) with tga 1.3.0:

Metric Value
CPU classify throughput (rule cascade only) ~113,000 commits/sec
Coverage — rule tiers only (--no-external) 64.3% of 72,608 commits
Coverage — with JIRA external source 67.7% of 72,608 commits
Peak RSS (full classify pass) 235 MB
weighted_sum tier contribution (new in 1.3.0) 9.6–11.0% of classified commits

JIRA external-source classification adds 3.4 percentage points of coverage at the cost of ~23 minutes of network-bound JIRA API time (the rule-cascade itself completes in 0.64s on this corpus). Use --no-external for fast iteration on rule files.

Full benchmark: docs/trusty-git-analytics/regression-testing/v1.3.0-2026-05-27.md

License

Non-commercial use only. See LICENSE for terms.

Elastic License 2.0