gadaj 0.2.0

Git and AI Data Aggregator for Journaling — aggregate git history and AI coding-agent sessions into a time-windowed work log.

gadaj

Git and AI Data Aggregator for Journaling

gadaj aggregates git commit history and AI coding-agent session data into a time-windowed work log. Give it a window — gadaj -w 4h, gadaj -s yesterday — and it tells you what changed in the repo, which AI models were used, by whom, for how long, and at what cost. The output is ready to paste into a journal entry or hand to an AI for narrative drafting.

Built for small human+AI teams where Claude Code, Aider, or similar agents work alongside humans on the same codebase. gadaj is the missing link between raw session transcripts and a coherent record of the day's work.

══ GIT  2026-04-28 10:00 – 13:25 EEST ════════════════════════════════

  Commits   8 · 51d38b6 – e60d59b
  Authors   Samuel (8)
  Files     7 changed · +672 / -7

  Hash      Time   Author   Files      Message
  `51d38b6` 10:02  Samuel   3 · +95/-8 Implement feature X
  `8a2c1d9` 10:15  Samuel   1 · +12/-0 Fix typo
  `e60d59b` 13:15  Samuel   3 · +565/–5 Refactor component Y
  (5 more commits...)

══ CLAUDE CODE  2026-04-28 10:00 – 13:25 EEST ════════════════════════

  Sessions  2 in window

  Session 1  10:02 – 11:47  ~1.7h
  Session 2  12:10 – 13:24  ~1.2h   ← most recent

  Model                      In    Out   Cache↑   Cache↓      Cost
  claude-sonnet-4-6         180k  174k    293k     9.6M    ~$6.59
  ──────────────────────────────────────────────────────────────────
  2 sessions total                                          ~$6.59

══ SUMMARY ═══════════════════════════════════════════════════════════

  Source  Summary
  Git     8 commits over ~3.1h · Samuel
  CC      ~$6.59 over ~2.9h from 2 sessions
  Total   ~$6.59 over ~2.9h

---

Why gadaj?

A new category of small human+AI teams has emerged — 2–5 people working alongside Claude Code, Aider, or other coding agents on the same codebase simultaneously or in sequence. These teams quickly notice a gap: the agent produces work, but there is no easy way to see what happened across a session, who contributed what, or what it cost — especially when a human and one or more agents are working in parallel or in sequence on the same repo.

Existing tools cover only half the picture:

CC session analytics tools (ccusage, claude-monitor, ccstats, claude-token-analyzer) are purely consumption-focused — tokens in, dollars out, plan limits. They have no awareness of git history and no concept of a "work period" that spans multiple sessions or contributors.

Git analytics tools (git-quick-stats, gitinspector, git-dev-time, git-journal) know commits intimately but have zero awareness of AI agents, session costs, or model usage.

AI commit/report tools (commitloom, ai-git-cli, wsr) generate text from git history but don't aggregate multi-source activity data.

The closest adjacent project is git-ai (git-ai-project), which tracks AI attribution at the line level — but it solves code provenance and review, not work logging.

The gap: no tool combines git history + AI session data + multi-contributor time-windowed aggregation into a work log oriented toward human-readable output. gadaj fills this gap. It is not a task tracker, narrative generator, or project manager. It collects evidence and presents it clearly. You decide what story to tell.

Tool	Question answered
ccusage, claude-monitor	Am I about to hit my token limit?
git-quick-stats, gitinspector	How active is this repo over time?
gadaj	What did our team (human + AI) actually do in this window?

On the name: like grep or awk, gadaj is a proper noun that earns independence from its expansion. The acronym is a mnemonic, not a contract — as the tool grows to cover more sources, the name stays. Zero collisions on PyPI and GitHub at time of writing.

---

Install

pip install gadaj

Requires Python 3.8+. Depends only on tomli (automatically installed for Python < 3.11; stdlib tomllib is used on Python 3.11+).

---

Quickstart

# Summarise the last 4 hours (default, with commits table)
gadaj

# Everything since yesterday morning (commits shown by default)
gadaj -s yesterday

# Hide the commits table, show summary only
gadaj -c

# JSON output for an AI agent to draft a journal entry from
gadaj -w 8h --json

---

Usage

gadaj [options]

Time window

All sources are queried against a single time window. These flags control it.

Flag	Short	Default	Description
--window DURATION	-w	4h	Look back from now: 2h, 1.5d, 90m
--since DATETIME	-s	—	Window start. ISO date/datetime, "yesterday", "today", "N hours ago", weekday name
--until DATETIME	-u	now	Window end

--window and --since are mutually exclusive.

Git options

Flag	Short	Description
--git-range A..B	-g	Explicit hash range, overrides time window for git
--git-last N		Last N commits, ignoring time window
--git-author NAME	-a	Filter commits to this author (substring match)
--git-filter PATTERN	-f	Filter by commit message (case-insensitive)
--no-git		Exclude git section

Claude Code options

Flag	Short	Description
--cc-file PATH		Use a specific .jsonl instead of scanning by time
--no-cc		Exclude Claude Code section

Output options

Flag	Short	Description
--no-commits	-c	Hide the full commits table (shown by default)
--json	-j	Machine-readable JSON
--out FILE	-o	Write to file instead of stdout
--raw	-R	Raw token counts (debug)
--tz HOURS	-z	UTC offset override, e.g. 3, +2, -5, or auto

Examples

# Only git, last 12 commits
gadaj --git-last 12 --no-cc

# Only CC sessions from the last 2 days
gadaj -s "2 days ago" --no-git

# Filter to commits touching a specific area
gadaj -w 1d -f "software-factory"

# One author's commits this week
gadaj -s monday -a Samuel --no-cc

# Pipe JSON to another tool
gadaj -w 8h --json | jq '.summary'

---

Configuration

gadaj reads config from ~/.config/gadaj/config.toml, created with defaults on first run. To override for a specific repo, add .gadaj.toml at the repo root.

[authors]
"Your Full Name"  = "YourNick"
"Collaborator"    = "Nick"

[pricing]
# (input, output, cache_write, cache_read) — $/MTok
"claude-opus-4-7"           = [15.00, 75.00, 18.75,  1.500]
"claude-opus-4-6"           = [15.00, 75.00, 18.75,  1.500]
"claude-sonnet-4-6"         = [ 3.00, 15.00,  3.75,  0.300]
"claude-haiku-4-5"          = [ 0.80,  4.00,  1.00,  0.080]
"claude-haiku-4-5-20251001" = [ 0.80,  4.00,  1.00,  0.080]

[defaults]
window = "4h"
tz     = "auto"   # or e.g. "2", "3"

[thresholds]
cost_warn  = 1.0   # yellow → orange coloring boundary (USD)
cost_alert = 5.0   # orange → red coloring boundary (USD)

[colors]
# Author color palette for the commits table (ANSI color codes).
# Values are integers (standard or 256-color codes). The default palette
# has 8 colors in complementary pairs: green/red, cyan/magenta,
# yellow/blue, orange/violet.
# author_palette = [32, 31, 36, 35, 33, 34, 130, 93]

[author_colors]
# Explicit author name → ANSI color code mapping for commits table.
# If defined, these colors take precedence over the palette.
# Authors not listed here use palette colors in order of first appearance.
# "Samuel Marisa"     = 32   # green
# "Mikko Lastname"    = 31   # red

Update [pricing] when Anthropic changes rates. Update [authors] when your team changes. Use [author_colors] to pin specific authors to specific colors. Neither requires touching the package itself.

---

JSON output

gadaj --json emits a stable schema suitable for downstream scripts or AI agents. The summary key is first so an agent can stop reading early if it only needs the overview.

{
  "summary": {
    "total_cost_usd": 6.59,
    "contributors": [...]
  },
  "window": {
    "since": "2026-04-28T10:00:00+00:00",
    "until": "2026-04-28T13:25:00+00:00",
    "duration_hours": 3.4
  },
  "git": { "commits": [...], "files_changed": 7, ... },
  "cc":  { "sessions": [...], "models": {...}, "total_cost_usd": 6.59 }
}

The schema is stable across patch versions. Breaking changes require a minor version bump. See design.md §6 for the full schema definition.

---

How gadaj finds Claude Code sessions

Claude Code writes .jsonl session files to ~/.claude/projects/<hashed-path>/. gadaj hashes the current working directory the same way CC does, scans all .jsonl files in that directory, and includes any whose timestamp range overlaps the requested window — regardless of how many sessions that covers.

Sessions with a gap of more than 30 minutes between them are flagged in output (⚠ Xm gap). Parallel sessions (overlapping timestamps, e.g. two agents running at the same time) are labelled [parallel].

If you want to analyse a session from a different project, use --cc-file PATH to point directly at a .jsonl file.

---

Adding a new source

gadaj uses a simple collector protocol. To add support for Aider, OpenRouter, Toggl, or any other source:

1. Create gadaj/collectors/<name>.py
2. Subclass Collector and implement two things:

from gadaj.collectors.base import Collector

class AiderCollector(Collector):

    @property
    def source_name(self) -> str:
        return "Aider"

    def collect(self, since, until) -> list:
        # Parse Aider's .aider.chat.history.md or similar log file
        # Filter to the given window
        # Return a list of model instances
        ...

3. Register it in cli.py alongside GitCollector and CCCollector

No other files need to change. See design.md §4 for the full protocol including the available property for graceful degradation when a source is not present.

---

Development

git clone https://github.com/sjjsy/gadaj
cd gadaj
pip install -e ".[dev]"
pytest
pytest --cov=gadaj --cov-report=term-missing

gadaj has one runtime dependency (tomli) only on Python < 3.11. Dev dependencies are pytest and pytest-cov.

For architecture, data models, testing strategy, and design decisions, see design.md.

---

Changelog

See CHANGELOG.md.

---

License

MIT