Git tracks your code.
Diversion tracks the reasoning.
Diversion connects to Claude Code, captures the full agentic reasoning behind every commit, and makes it queryable in plain English.
An agent checking the reasoning before changing the code.
100% agentic. No human in the loop — just Claude using /diversion:ask
to consult past decisions before acting.
One plugin. Zero workflow change.
Trajectory layers on top of GitHub — your team keeps using PRs, branches, reviews, and CI exactly as before. Diversion runs alongside, capturing what GitHub can't.
Claude Code plugin
- Single distribution unit — install once
- Lives where your engineers already work
- No new dashboards, no new tools to learn
GitHub repo connection
- Two-way sync with your existing GitHub repo
- Branches, commits, and history stay consistent
- Your PR / review workflow is untouched
Reasoning storage
- Plugin hook captures every reasoning turn
- Linked directly to the commit it produced
- Stored alongside your repo, indexed in real time
Trajectory Query
- Ask questions in natural language
- Answers grounded in the actual session turns
- Available from your agent, CLI, or web UIAvailable from your agent, CLI, or web UI
Tribal knowledge, now a query away.
Diversion answers from the actual sessions that produced your code — not from documentation that's stale, or guesses based on the diff.
Stop reverse-engineering.
Start asking.
When something breaks at 2 AM, ask
Diversion what the function was meant to
handle and what edge cases the agent
considered — straight from the session
that wrote it.
What was the previous fix?
grounded in real intent, not guesswork.
Every architectural choice, fully sourced.
Before refactoring or extending a service,
surface the original session — the prompts,
the rejected alternatives, and the
reasoning behind the final call.
DynamoDB for the events service?
constraints. No more redoing analysis
your team already did.
New engineer? New agent? Quickly understand the codebase.
New hires and new agents ask the
codebase directly — how does this module
work, why two queue implementations?
Answers come from the sessions that built
it, not stale wikis.
why are there two queue
implementations?
knowledge. Agents that operate with
full context from their first prompt.
Start mirroring before the next outage.
Join the early access list.
Frequently asked Questions
Do I need to leave GitHub?
No. The Diversion Plugin connects to your existing GitHub repo via two-way sync. Your team's PRs, reviews, branches, and CI keep working exactly as before. Diversion runs alongside as the context layer.
Which AI agents are supported?
The Diversion Plugin ships first as a Claude Code plugin and captures every turn of every Claude Code session in your repo. Support for additional agents (Cursor, Copilot, custom agents) is on the roadmap.
Where is the reasoning data stored?
In Diversion, alongside your repo metadata. You control access. SOC 2 Type II. Encrypted at rest and in transit. Your code and reasoning are yours - we don't train on them, we don't read them, we don't share them.
Does it slow down my agent sessions?
No. Capture happens asynchronously via a plugin hook. Your sessions feel identical - you just gain a permanent, queryable record of what happened.
