git-impact

Translate your git commits into plain-English business impact — for standups, manager updates, and performance review prep.

📅 Thursday, May 8, 2026

✅ Shipped secure multi-tenant data access layer       [PR #142]
   → Prevents cross-company data leaks. Required for SOC2 compliance sign-off.

✅ Fixed login failures for admin users                [a1b2c3d, ENG-431]
   → Unblocks Q2 portal launch, was blocking 3 engineers.

⏳ In progress: Refactoring authentication flow        [inferred]
   → Will reduce login latency by ~40%, targeting end of week.

📁 6 files changed across auth + database layers
   4 commits on main

No API key, ever. Translation runs inside your AI editor.

Install

Run this once inside any repo — no install step needed, npx fetches it on demand:

npx git-impact init

The wizard auto-detects which AI editors you already use and installs for them. Three short questions, then it self-installs:

  git-impact init
  ────────────────────────────────────

  Repo: /your/project
  Files will be committed with your team.

  What does your company/product do? (1–2 sentences)
  > B2B SaaS for workforce analytics

  What does your manager care most about?
  > Shipping on time, not breaking prod

  Technical terms to translate? (optional)
  Format: "RLS=data security, MFA=login security"
  >

  Which AI editors should I install for? (comma-separated, or "all")
  Options: claude, copilot, cursor, gemini, opencode
  (.agents/skills/ is always written — covers Goose, Amp, Codex, Letta, Roo)
  Detected in this repo: claude, cursor
  [press Enter to use detected]
  >

Then it creates everything and tells you what it did:

  ────────────────────────────────────
  git-impact installed

  ✅  .git-impact/context.json
  ✅  .gitignore
  ✅  .agents/skills/git-impact/SKILL.md          ← cross-vendor baseline
  ✅  .agents/skills/git-impact/references/...
  ✅  .claude/skills/git-impact/SKILL.md
  ✅  .claude/skills/git-impact/references/...
  ✅  .cursor/skills/git-impact/SKILL.md
  ✅  .cursor/skills/git-impact/references/...
  ✅  CLAUDE.md
  ✅  .git-impact/manifest.json

  Next steps:
  1. git add .git-impact/context.json && git commit -m "chore: add git-impact"
  2. In Claude Code: type "/git-impact" or say "do my standup"
  3. After a few standups: `git-impact view` to see the rolling dashboard

Using it

Once installed, just talk to your AI editor naturally:

| What you say | What happens | |---|---| | /git-impact or do my standup | Translates commits since your last standup (survives weekends) | | since yesterday, since 3d, since 2026-05-01 | Explicit lookback window | | what did I ship this week? | Looks back 7 days | | generate a performance review | Synthesises history into review-prep evidence | | make a presentation for today | Adds a bespoke shareable HTML page (opt-in) | | set up context for this repo | Re-runs the init questions inline |

Works in any editor that supports the Agent Skills format. Five editors get a vendor-specific mirror (Claude Code, Cursor, GitHub Copilot + VS Code, Gemini CLI, OpenCode). Every other Agent Skills adopter — Goose, Amp, OpenAI Codex, Letta, Roo Code, and future ones — picks up the same skill from the cross-vendor .agents/skills/ baseline that's always written. One canonical SKILL.md folder, mirrored where each vendor's docs say it needs to live.

How it works

git-impact ships two things: an Agent Skills–compatible skill that handles the prompting (Claude Code, Cursor, Copilot, Gemini CLI, OpenCode get a vendor-specific mirror; Goose, Amp, Codex, Letta, Roo pick up the .agents/skills/ baseline), and an MCP server that exposes typed data tools (git activity, history, context). The skill calls the MCP tools — no sqlite3 shelling, no hand-rolled SQL.

your commits  →  MCP get_git_activity  →  privacy redaction  →  AI translation
                          ↓
              MCP save_impact_entry (with provenance + refs)
                          ↓
              .git-impact/result.html (rolling dashboard)

Two safeguards built in:

• Privacy redaction (default-on) — filenames matching .env*, .aws/, .ssh/, *credentials*, *secret*, *.pem, *.key get replaced with [redacted-secret-file] before they reach the model. Obvious-looking secret values in commit bodies (Stripe keys, AWS access key ids, GitHub PATs, Slack tokens, JWTs) become [redacted-secret].
• Per-bullet provenance — every translated bullet is tagged pr / commit_body / commit_message / ticket / inferred. Inferred bullets surface with an amber chip in the HTML report so you (and your manager) can tell what's grounded vs. guessed.

History is saved locally to .git-impact/history.db (SQLite, gitignored) so review-prep can synthesise weeks or months of standups.

What gets installed

your-project/
├── .agents/
│   └── skills/git-impact/                      # Cross-vendor Agent Skills baseline
│       ├── SKILL.md                            # Workflow router (orchestrates MCP tools)
│       └── references/                         # Loaded on demand (progressive disclosure)
│           ├── mode-standup.md
│           ├── mode-review.md
│           ├── mode-init.md
│           ├── translation-rules.md
│           └── html-template.md
├── .claude/skills/git-impact/                  # Same skill folder, mirrored for Claude Code
├── .cursor/skills/git-impact/                  # …mirrored for Cursor
├── .github/skills/git-impact/                  # …mirrored for GitHub Copilot + VS Code
├── .gemini/skills/git-impact/                  # …mirrored for Gemini CLI
├── .opencode/skills/git-impact/                # …mirrored for OpenCode
│   (Goose, Amp, Codex, Letta, Roo — read from .agents/skills/ above, no mirror)
├── .git-impact/
│   ├── context.json                            # Team glossary  ← commit this
│   ├── manifest.json                           # Install record ← commit this
│   ├── history.db                              # Local history  ← gitignored
│   ├── result.html                             # Rolling dashboard ← gitignored
│   └── standups/YYYY-MM-DD.html                # Bespoke presentations (opt-in)
└── CLAUDE.md                                   # Usage reminder block (managed)

Commit the SKILL files and context.json. Everything else stays local.

Re-running npx git-impact init is safe — it updates existing files in place.

context.json

Lives at .git-impact/context.json. Commit it so everyone on the team gets the same translation:

{
  "companyDescription": "B2B SaaS for workforce analytics",
  "managerPriorities": "Shipping on time, not breaking prod",
  "glossary": {
    "RLS": "data security layer",
    "MFA": "login security",
    "TabPFN": "AI predictions"
  },
  "privacy": {
    "redact": true,
    "filePatterns": ["*internal-only*"],
    "valuePatterns": ["MY_COMPANY_NAME"]
  }
}

The privacy block is optional — defaults work for most teams. Set redact: false to disable the filter, or extend the patterns to catch internal names / customer identifiers.

Translation rules

The skill follows these rules to keep output useful rather than vague:

1. What + why, not what + how — "Fixed login failures for admin users → unblocks Q2 launch" not "updated auth middleware"
2. Apply the glossary — every term in context.json is replaced with its plain-English equivalent
3. Provenance is mandatory — every saved bullet is tagged pr / commit_body / commit_message / ticket / inferred. Inferred bullets render with an amber chip so the reader can spot what was grounded vs guessed.
4. Group related commits — four auth commits become one bullet, not four
5. Three statuses — done (✅), in_progress (⏳), blocked (🚫). WIP work surfaces the expected outcome, not the work-so-far.
6. Use numbers — "reduced latency by ~40%" beats "improved performance"

Full rules in skill/references/translation-rules.md.

Performance review prep

Use the standup daily for a few weeks. History builds up in history.db. Then:

generate performance review prep for the last 90 days    ← Claude Code
generate review for Q2-2026                              ← any supported editor

Output (think of this as evidence pack you paste into your own writing — not a finished review):

Performance Review Prep — Q2 2026

Led the security layer redesign that enabled SOC2 compliance sign-off.

🚀 Security
   • Shipped multi-tenant data isolation, unblocking a $2M enterprise deal      [Apr 8, PR #98]
   • Resolved 2 auth vulnerabilities before the audit window                    [May 2, PR #114]

✅ Features
   • Delivered admin portal ahead of schedule (was blocking 3 engineers)        [Apr 22, PR #103]

🔧 Reliability
   • Reduced login latency by ~40% through auth flow refactor                   [May 8, inferred]

📊 47 commits across 58 working days

CLI reference

The CLI is just for installing and viewing reports — translation always happens inside your AI editor.

git-impact init                    # set up / update this repo
git-impact view                    # open the HTML report of saved standups
git-impact view --date 2026-05-09  # open a specific day directly

All commands accept -p <path> to point at a different repo. By default the repo is auto-detected from the current directory.

MCP server

The MCP server exposes git data as typed tools so AI editors can call them directly. No API key — your editor does the translation in its own session.

The skill folder gets installed automatically by npx git-impact init, but the MCP server has to be registered in your editor's MCP config separately — otherwise the skill loads but save_impact_entry / render_dashboard aren't callable and nothing persists. Same JSON block everywhere, just different file paths:

| Editor | Config file | |---|---| | Claude Code | ./.claude/settings.json (per-repo) | | Antigravity | ~/.gemini/antigravity/mcp_config.json (user-global) | | Other editors | See your editor's MCP setup docs — the block below works for any stdio MCP host |

{
  "mcpServers": {
    "git-impact": {
      "command": "npx",
      "args": ["git-impact-mcp"]
    }
  }
}

Restart the editor after editing the config. Available tools:

| Tool | What it does | |---|---| | get_git_activity | Reads commits + file changes, applies privacy redaction | | get_github_activity | Pulls PR titles + descriptions (needs github_token in context) | | get_last_standup_date | Powers the "since last standup" default | | save_impact_entry | Persists a translation with provenance + refs | | get_history | Returns saved entries for a date range | | render_dashboard | Regenerates .git-impact/result.html and returns the file:// URL | | update_context | Updates context.json (company / glossary / privacy / token) |

Local development

git clone https://github.com/efraimnabil/git-impact
cd git-impact
npm install
npm run build

# Test the installer
node dist/cli/index.js init

# Open the HTML report
node dist/cli/index.js view

# Run the MCP server
node dist/mcp/server.js

# Tests
npm test                  # one-shot
npm run test:watch        # watch mode

Publishing a new version:

npm version patch          # bump version
npm publish --otp=123456   # 6-digit code from your authenticator
git push --follow-tags

Project structure

src/
├── init/
│   ├── installer.ts        # install() + runInitWizard() + detectEditors()
│   ├── installer.test.ts
│   └── templates.ts        # CLAUDE.md block + context.json template
├── readers/
│   ├── git.ts              # reads git log, diffs, file changes
│   ├── github.ts           # reads PRs via GitHub API (optional)
│   ├── redact.ts           # privacy filter (filenames + secret-shaped values)
│   └── redact.test.ts
├── storage/
│   ├── db.ts               # per-repo SQLite history + context.json
│   └── db.test.ts
├── report/
│   ├── render.ts           # builds the rolling HTML dashboard
│   └── html.ts             # HTML template + provenance chips
├── mcp/
│   ├── server.ts           # MCP server entry point
│   ├── tools.ts            # typed data tools (get_git_activity, save_impact_entry, …)
│   ├── resources.ts        # MCP resources (context, history overview)
│   ├── prompts.ts          # computed prompt templates (standup, review)
│   └── repo.ts             # auto-detects repo root from cwd
└── cli/
    └── index.ts            # init + view commands
skill/
├── SKILL.md                # ~60-line router (picks a mode, points at references/)
└── references/             # loaded on demand (Agent Skills progressive disclosure)
    ├── mode-standup.md     # default mode — translate commits → bullets
    ├── mode-review.md      # synthesise history → performance review prep
    ├── mode-init.md        # interactive context.json setup
    ├── translation-rules.md
    └── html-template.md

Roadmap

Shipped:

• [x] Skill + MCP architecture (skill orchestrates, tools provide)
• [x] Per-bullet provenance tags + refs
• [x] Privacy redaction (default-on)
• [x] "Since last standup" default
• [x] Editor auto-detection in init
• [x] Test baseline (Vitest, 40 tests)
• [x] Multi-editor: Claude Code, Copilot, Cursor, Gemini CLI
• [x] Canonical SKILL.md format (Agent Skills) — vendor mirrors for OpenCode; .agents/ baseline picked up by Goose, Amp, Codex, Letta, Roo
• [x] Auto-migration from pre-0.7 per-editor instruction files

Next:

• [ ] PR description writer — git-impact-pr skill produces ready-to-paste PR bodies
• [ ] Multi-repo standup — aggregate across ~/work/* in one update
• [ ] Changelog mode — since v1.2.0 produces user-facing release notes
• [ ] Team / cloud features — shared history, shareable evidence-pack URLs

MIT License