Skip the archaeology when you come back
Decision Detection runs for free on every checkpoint. The add-on unlocks full decision details, rationale notes, and LLM-powered refinement.
The problem
You open a project after three weeks away. There are 40 new commits. Most are minor: formatting, small fixes, dependency bumps. But somewhere in there, someone changed the auth strategy, swapped the database driver, and restructured the API routes.
Those three commits are the ones standing between you and momentum. Everything else is noise. But git treats them all the same, so you scroll through git log trying to figure out what to pay attention to.
How it works
You commit
Any editor, any terminal, any AI coding agent. Just git commit as usual.
The hook classifies
A global post-commit hook runs silently in the background. It scores your checkpoint summary text for decision signals like architectural choices, migrations, and pivots, with git-diff patterns as a confidence boost.
The decision is saved
If the commit is high-signal, it is tagged with a category and saved to .keepgoing/keepgoing.db. Low-signal commits are quietly ignored.
Example: making a commit that triggers detection
~/my-app git commit -m "ci: add deploy workflow for staging"
[main abc1234] ci: add deploy workflow for staging
2 files changed, 45 insertions(+)
[KeepGoing] Decision detected: infrastructure (92% confidence)
What gets detected
The heuristics look for commits that change the direction or shape of a project. Here are the main categories.
Infrastructure
CI/CD pipelines, deployment configs, hosting changes
ci: add deploy workflow for staging
chore: switch from Heroku to Fly.io
Authentication
Auth middleware, session handling, security policies
feat: replace session cookies with JWT
fix: add CSRF protection to auth endpoints
Database
Migrations, schema changes, database engine swaps
feat: add user_preferences table migration
refactor: switch from Postgres to SQLite
Architecture
New patterns, service extraction, major refactors
refactor: extract payment service from monolith
feat: add event bus for async processing
Dependencies
Major version bumps, library replacements, new packages
chore: upgrade React 18 to 19
feat: replace moment.js with date-fns
Always-on, everywhere
Decision detection runs via a global git post-commit hook. It does not require VS Code, JetBrains, or any specific editor to be open. If you can run git commit, decisions are captured.
Install once and every repository you work in is covered. No per-project setup, no editor dependency, no polling.
keepgoing hook installFeatures
Decision cards in sidebar
Browse detected decisions in a dedicated VS Code panel with category badges, timestamps, and quick navigation to the commit.
Rationale annotation
Add "why" notes to any decision. When you come back in three weeks, the reasoning is right there.
Category badges
Each decision is tagged with its category: infra, auth, migration, dependency, architecture, and more.
LLM refinement
Run keepgoing refine to improve classification with AI. Confirms or rejects heuristic candidates and finds decisions the heuristic missed. Works with your own Anthropic API key or Claude CLI.
Auto-refinement
Enable with keepgoing refine --enable and every detected decision is automatically refined in the background. No manual step needed.
MCP tool access
AI assistants can query your decisions via the get_decisions MCP tool.
CLI access
Browse decisions from the terminal with keepgoing decisions. Filter by branch and limit results.
Local and private
Decisions are stored in .keepgoing/keepgoing.db. Nothing leaves your machine.
Common questions
How does it decide what is a "decision"?
The heuristic scores your checkpoint summary text using weighted signal categories: decision verbs, change verbs, contrast language, and design markers. Git-diff patterns serve as an optional confidence boost. Detection runs in two tiers: the always-on heuristic runs for everyone on every checkpoint save, and the optional LLM refinement tier (via keepgoing refine) uses your Anthropic API key or Claude CLI to confirm, reject, or discover decisions the heuristic missed. You can also enable auto-refinement so this happens in the background after every detection.
Do I need Claude CLI installed for LLM refinement?
No. If you have an Anthropic API key, you can store it with keepgoing config set api-key sk-ant-... or set the ANTHROPIC_API_KEY environment variable. KeepGoing checks for an API key first, then falls back to Claude CLI. Refinement uses the Haiku model, so typical cost is under $0.01 per decision.
Will it slow down my commits?
No. The post-commit hook runs in the background and never blocks your terminal.
Does it require VS Code?
No. Decision detection runs via a global git post-commit hook that works with any editor, terminal, or AI coding agent. VS Code provides a sidebar panel for browsing decisions, but the detection itself is editor-independent.
Can I use it on multiple devices?
Each license activates on one device at a time. Deactivate from one device and reactivate on another whenever you want.
What if I want a refund?
Contact us within 14 days of purchase for a full refund, no questions asked. See our refund policy.
Lifetime license. No subscription, no recurring fees. Activate on one device at a time.