byollm 0.1.5-beta.2 → 0.1.5
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +17 -9
- package/dist/bin.js +1 -1
- package/dist/{chunk-GWC3CCIW.js → chunk-LB4MGI2X.js} +169 -74
- package/dist/chunk-LB4MGI2X.js.map +1 -0
- package/dist/index.d.ts +3 -1
- package/dist/index.js +1 -1
- package/package.json +2 -2
- package/dist/chunk-GWC3CCIW.js.map +0 -1
package/README.md
CHANGED
|
@@ -1,11 +1,16 @@
|
|
|
1
|
-
<!-- release-note 0.1.
|
|
2
|
-
> **`0.1.5
|
|
3
|
-
> is still `2
|
|
4
|
-
>
|
|
1
|
+
<!-- release-note 0.1.5 -->
|
|
2
|
+
> **`0.1.5` carries site tools and images. Update your hub or server first.**
|
|
3
|
+
> `PROTOCOL_VERSION` is still `2`. `llm.tools` lets the model call a site's
|
|
4
|
+
> own tools, which run on the site and never on this device. It is off until
|
|
5
|
+
> you turn it on for a service, and every call is in `byollm log`.
|
|
6
|
+
> `llm.vision` and `llm.image` work on local and HTTP models. A teammate's
|
|
7
|
+
> images are budgeted by `community.maxImageBytes`. A teammate's large job is
|
|
8
|
+
> now admitted when it could fit (#5738). Full note:
|
|
9
|
+
> `docs/release-notes/0.1.5.md`.
|
|
5
10
|
>
|
|
6
11
|
> This README used to open with a stack of release notes instead of with the
|
|
7
12
|
> package. The history moved to `docs/release-notes/` and `CHANGELOG.md`,
|
|
8
|
-
> where none of it was lost.
|
|
13
|
+
> where none of it was lost.
|
|
9
14
|
|
|
10
15
|
# `byollm`
|
|
11
16
|
|
|
@@ -237,12 +242,15 @@ locked to your own work — a protocol rule, not a setting you can change.
|
|
|
237
242
|
Every payload is treated as hostile input. Breakout is structurally
|
|
238
243
|
impossible: process-class backends get a fixed argv with the prompt on stdin,
|
|
239
244
|
a stripped environment, an empty scratch directory and hard timeout/output
|
|
240
|
-
caps; HTTP-class backends spawn nothing at all. The model has no tools
|
|
241
|
-
retrieval and no MCP.
|
|
245
|
+
caps; HTTP-class backends spawn nothing at all. The model has no tools that
|
|
246
|
+
run on your device, no retrieval and no MCP. A site's own tools (`llm.tools`,
|
|
247
|
+
off until you turn it on for a service) run on the site: the daemon seals each
|
|
248
|
+
call to that site, logs it first, and never executes it.
|
|
242
249
|
|
|
243
250
|
Prompt injection — steering what the model _says_ — is not prevented by
|
|
244
|
-
anything and we do not claim otherwise. It is bounded here because
|
|
245
|
-
|
|
251
|
+
anything and we do not claim otherwise. It is bounded here because nothing
|
|
252
|
+
the model says runs on the device: its output is inert, and a call to a site's
|
|
253
|
+
tool is sealed to that site and run there.
|
|
246
254
|
|
|
247
255
|
Full threat model, including what the OS stops us dropping:
|
|
248
256
|
[`docs/security.md`](https://github.com/oftomorrowinc/byollm/blob/main/docs/security.md).
|