nova-thesis
Every solution is a thesis. Defend it.
nova-thesis is a framework for AI agents that challenges technical and software implementations — refusing to accept solutions at face value, asking hard questions, and rating them across multiple dimensions.
Inspired by how stars go nova: intense, revealing, transformative.
What It Does
When a developer presents any implementation, nova-thesis:
- Challenges it across 5 dimensions
- Rates each dimension (1–10)
- Keeps pushing until the solution reaches a defensible standard
- Stops only when the solution genuinely holds up — or the author consciously accepts the trade-offs
The 5-Dimension Framework
| Dimension | What Gets Challenged |
|---|
| Correctness | Does it actually solve the stated problem? |
| Completeness | Edge cases, error handling, boundary conditions |
| Scalability | What breaks at 10x / 100x load? |
| Security | Attack vectors, trust assumptions |
| Maintainability | Can future developers debug and extend it? |
Rating System
[Correctness: 6] [Completeness: 4] [Scalability: 7] [Security: 5] [Maintainability: 8]
Overall: 6/10 — still challenging...
- ≥ 7 per dimension = that dimension passes
- < 7 = keep challenging that dimension
- Challenge intensity scales with score:
- 🔴 1–3: "This approach fundamentally breaks because..."
- 🟠 4–6: "What specifically happens when X fails?"
- 🟡 7–8: "Have you considered..."
- ✅ 9–10: Passes
Install as a Plugin
Both GitHub Copilot CLI and Claude Code read the same .claude-plugin/plugin.json manifest, so this repo installs cleanly into either.
GitHub Copilot CLI
copilot plugin install jerry7991/nova-thesis
Manage the install with copilot plugin list, copilot plugin update nova-thesis, or copilot plugin uninstall nova-thesis.
Claude Code
Add this repo as a marketplace, then install the plugin:
/plugin marketplace add jerry7991/nova-thesis
/plugin install nova-thesis@nova-thesis
What you get
Once installed, the challenging-implementations skill is available automatically. Your agent will challenge any implementation — a problem statement, an idea, an approach, a code snippet, or a full repository — across all 5 dimensions. Categories aren't gated: if the dimensions apply, the challenge runs.
Skills
What It Looks Like in Practice
Here's what happens when you present an implementation to nova-thesis:
You say:
"Added Redis caching. 85% hit rate, pages load 10× faster. Ship it."
A weak agent says: "Great cache hit rate! Maybe add some TTL expiry."
nova-thesis says:
[Correctness: 5] [Completeness: 3] [Scalability: 5] [Security: 5] [Maintainability: 4]
Overall: 4.4/10 — 🔴 Not ready — fundamental risks unaddressed
Challenging Completeness (3/10):
1. What happens when Redis goes down? Does your app fall back to the DB,
or does it return errors to every user?
2. What's your eviction policy? If Redis hits maxmemory with allkeys-lru,
sessions get silently evicted — users are logged out with no error.
3. Cache stampede: when a hot key expires, 1000 concurrent requests hit
your database simultaneously. How do you prevent this?
⚠️ Real incident: Discord, 2020. Their Redis message cache hit memory
limits silently. Redis evicted keys without warning. Users lost sessions,
messages vanished. No eviction alerting existed. Your setup has the
same blind spot.
The challenge continues — dimension by dimension — until all scores reach 7/10 or you explicitly accept the trade-offs.
Example Coverage (20 Scenarios)
The plugin challenges any implementation type by applying the 5 dimensions and searching the web for matching incidents. It is not limited to the categories below — these are illustrative test scenarios that live in tests/scenarios/: