Model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
tags:
|
| 3 |
+
- codesync
|
| 4 |
+
- devshub
|
| 5 |
+
- decision-model
|
| 6 |
+
- developer-tools
|
| 7 |
+
pipeline_tag: text-classification
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# CodeSync Knack
|
| 11 |
+
|
| 12 |
+
**The free, private judgment model behind CodeSync.**
|
| 13 |
+
|
| 14 |
+
Every time you send CodeSync a message or it runs a shell command, it makes dozens of quick
|
| 15 |
+
calls: Is this command risky? Will it run forever? Are you asking to continue earlier work?
|
| 16 |
+
Is this a question or a task? Which mode fits? Knack makes those calls on your own machine,
|
| 17 |
+
in about 10 milliseconds.
|
| 18 |
+
|
| 19 |
+
- **Free.** Ships with every CodeSync install. No API key, no account, no usage limits.
|
| 20 |
+
- **Private.** Runs locally. Your prompts and commands never leave your machine to be judged.
|
| 21 |
+
- **Fast.** About 10 ms per decision on a GPU, about 250 ms on a plain CPU.
|
| 22 |
+
- **Small.** About 1 GB on disk. Installs itself in the background after CodeSync starts.
|
| 23 |
+
- **Accurate where it matters.** Trained on real CodeSync sessions, with extra focus on the
|
| 24 |
+
rare cases that cost the most when missed: destructive deletes, commands that never exit,
|
| 25 |
+
resuming interrupted work, refactor / test / documentation requests.
|
| 26 |
+
|
| 27 |
+
## Results
|
| 28 |
+
|
| 29 |
+
Measured on CodeSync's own judgments (version v3, November 2026):
|
| 30 |
+
|
| 31 |
+
| Benchmark | Knack v3 | What it measures |
|
| 32 |
+
|---|---:|---|
|
| 33 |
+
| Hand-labelled set | **97.9%** | 233 decisions labelled by hand |
|
| 34 |
+
| Held-out set | **95.5%** | 3,267 decisions from real sessions, never trained on |
|
| 35 |
+
| Rare-case set | **93.7%** | 4,120 decisions on the hardest, rarest cases |
|
| 36 |
+
| Latency | **13 ms** | median per request on a laptop GPU |
|
| 37 |
+
|
| 38 |
+
## How to get it
|
| 39 |
+
|
| 40 |
+
Install CodeSync. That's it.
|
| 41 |
+
|
| 42 |
+
```bash
|
| 43 |
+
curl -fsSL https://app.devshub.ai/install.sh | bash
|
| 44 |
+
```
|
| 45 |
+
|
| 46 |
+
The installer stays as fast as before. On first start, CodeSync downloads Knack in the
|
| 47 |
+
background and keeps working while it does. Once Knack is ready it becomes CodeSync's first
|
| 48 |
+
judgment engine, and new versions are picked up automatically.
|
| 49 |
+
|
| 50 |
+
To turn it off: `CODESYNC_KNACK_DISABLED=1`.
|
| 51 |
+
|
| 52 |
+
## What it decides
|
| 53 |
+
|
| 54 |
+
Knack answers CodeSync's fixed set of 29 decisions:
|
| 55 |
+
|
| 56 |
+
- **Every turn (21):** task intent, routing and reasoning depth, plan / ask / build mode,
|
| 57 |
+
new task vs continuing work, resume requests, screencast start and stop, browser and
|
| 58 |
+
messaging needs, whether the request is chat or a task, and more.
|
| 59 |
+
- **Every shell command (8):** risk level, activity type, long-running or not, whether it
|
| 60 |
+
verifies, mutates the repo, deploys, needs a running app, or starts a server.
|
| 61 |
+
|
| 62 |
+
It only answers the questions it was trained on. Anything else is passed to CodeSync's
|
| 63 |
+
cloud judgment service, so a new kind of question never gets a guess.
|
| 64 |
+
|
| 65 |
+
## Files
|
| 66 |
+
|
| 67 |
+
| File | Purpose |
|
| 68 |
+
|---|---|
|
| 69 |
+
| `manifest.json` | current version, archive URL, sha256 (read by CodeSync) |
|
| 70 |
+
| `knack-<version>.tar.gz` | the model folder |
|
| 71 |
+
|
| 72 |
+
Built by [DevsHub](https://app.devshub.ai) for [CodeSync](https://app.devshub.ai).
|