genaicode 2.32.0 → 2.33.0
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 +1 -0
- package/dist/agents/session.d.ts +11 -0
- package/dist/agents/session.js.map +1 -1
- package/dist/browser/fake-model.d.ts +43 -0
- package/dist/browser/fake-model.js +133 -0
- package/dist/browser/fake-model.js.map +1 -0
- package/dist/browser/language-model.d.ts +74 -0
- package/dist/browser/language-model.js +83 -0
- package/dist/browser/language-model.js.map +1 -0
- package/dist/browser/nano.d.ts +83 -0
- package/dist/browser/nano.js +610 -0
- package/dist/browser/nano.js.map +1 -0
- package/dist/browser/page-agent.d.ts +74 -0
- package/dist/browser/page-agent.js +339 -0
- package/dist/browser/page-agent.js.map +1 -0
- package/dist/browser/tools.d.ts +36 -0
- package/dist/browser/tools.js +190 -0
- package/dist/browser/tools.js.map +1 -0
- package/dist/browser/webmcp.d.ts +32 -0
- package/dist/browser/webmcp.js +150 -0
- package/dist/browser/webmcp.js.map +1 -0
- package/dist/browser.d.ts +39 -0
- package/dist/browser.js +32 -0
- package/dist/browser.js.map +1 -0
- package/dist/react/agent-chat.css +141 -0
- package/dist/react/agent-chat.d.ts +12 -2
- package/dist/react/agent-chat.js +32 -4
- package/dist/react/agent-chat.js.map +1 -1
- package/dist/react/attachments.d.ts +41 -0
- package/dist/react/attachments.js +162 -0
- package/dist/react/attachments.js.map +1 -0
- package/dist/react/page-agent-chat.d.ts +38 -0
- package/dist/react/page-agent-chat.js +40 -0
- package/dist/react/page-agent-chat.js.map +1 -0
- package/dist/react.d.ts +6 -1
- package/dist/react.js +2 -0
- package/dist/react.js.map +1 -1
- package/dist/ui/web-client.js +154 -13
- package/docs/browser.md +200 -0
- package/docs/semver.md +2 -0
- package/package.json +5 -1
package/docs/browser.md
ADDED
|
@@ -0,0 +1,200 @@
|
|
|
1
|
+
# An agent in the page (`genaicode/browser`)
|
|
2
|
+
|
|
3
|
+
[← README](../README.md)
|
|
4
|
+
|
|
5
|
+
`genaicode/browser` runs an agent **in the web page itself**, on the browser's built-in model
|
|
6
|
+
(Gemini Nano through Chrome's [Prompt API](https://developer.chrome.com/docs/ai/prompt-api)),
|
|
7
|
+
and offers the page's tools to outside agents over [WebMCP](https://webmachinelearning.github.io/webmcp/).
|
|
8
|
+
No server, no API key, no cost per token, and nothing the user types leaves their device.
|
|
9
|
+
Nothing in it imports Node.
|
|
10
|
+
|
|
11
|
+
It is a small model. It is good at "show me X", "what does this page say about Y", "switch
|
|
12
|
+
that on": one or two tool calls and a short answer. It does not write code; for that, put a
|
|
13
|
+
coding agent behind your app with [`genaicode/app-agent`](app-agent.md).
|
|
14
|
+
|
|
15
|
+
```ts
|
|
16
|
+
import { createPageAgent, exposeWebMcp, type PageTool } from 'genaicode/browser';
|
|
17
|
+
|
|
18
|
+
const tools: PageTool[] = [
|
|
19
|
+
{
|
|
20
|
+
name: 'list_products',
|
|
21
|
+
description: 'The products in the shop, with id, name and price.',
|
|
22
|
+
run: () => products,
|
|
23
|
+
},
|
|
24
|
+
{
|
|
25
|
+
name: 'add_to_cart',
|
|
26
|
+
description: 'Put a product in the cart.',
|
|
27
|
+
inputSchema: { type: 'object', properties: { id: { type: 'string' } }, required: ['id'] },
|
|
28
|
+
write: true, // the user approves it first
|
|
29
|
+
summary: ({ id }) => `Add ${nameOf(id)} to your cart`,
|
|
30
|
+
run: ({ id }) => cart.add(String(id)),
|
|
31
|
+
},
|
|
32
|
+
];
|
|
33
|
+
|
|
34
|
+
export const assistant = createPageAgent({
|
|
35
|
+
name: 'helper',
|
|
36
|
+
instructions: 'You help shoppers on Example Shop. Prices are in euro.',
|
|
37
|
+
tools,
|
|
38
|
+
});
|
|
39
|
+
|
|
40
|
+
exposeWebMcp(tools); // the same tools for the browser's own agent and extensions
|
|
41
|
+
```
|
|
42
|
+
|
|
43
|
+
```tsx
|
|
44
|
+
import { PageAgentChat } from 'genaicode/react';
|
|
45
|
+
import 'genaicode/react/styles.css';
|
|
46
|
+
|
|
47
|
+
<PageAgentChat agent={assistant} emptyState={<p>Ask me about the shop.</p>} />;
|
|
48
|
+
```
|
|
49
|
+
|
|
50
|
+
`<PageAgentChat>` is [`<AgentChat>`](react.md) wired to the page agent, with what an
|
|
51
|
+
on-device model needs on top: a note where the browser has no model (`unavailable`), a
|
|
52
|
+
"Download the model" button with progress, Approve/Deny for `write` tools, a chip that shows
|
|
53
|
+
how full the context window is, and "New". Without React, read `assistant.get()` and
|
|
54
|
+
`assistant.subscribe()` and call `send`, `stop`, `reset`, `approve` and `prepare` yourself.
|
|
55
|
+
|
|
56
|
+
## Tools
|
|
57
|
+
|
|
58
|
+
A `PageTool` is a function with a `name`, a `description` and a JSON Schema (`inputSchema`,
|
|
59
|
+
an object schema). `run(args, signal)` returns text, or any value, which is sent as JSON.
|
|
60
|
+
|
|
61
|
+
- `write: true` marks a tool that changes something. It runs only after the user approved the
|
|
62
|
+
call; `summary(args)` is the line they read. An in-page agent asks in the chat, an outside
|
|
63
|
+
agent goes through `exposeWebMcp`'s `confirm`.
|
|
64
|
+
- `validate(args)` may throw to refuse a call before anyone is asked. Required arguments,
|
|
65
|
+
basic types and `enum` values are checked for you, and the problem goes back to the model.
|
|
66
|
+
- Keep the list short (a handful) and the outputs small: the model sees at most
|
|
67
|
+
`maxToolOutput` characters (default 2,000) of each result. `enum` arguments help a small
|
|
68
|
+
model the most.
|
|
69
|
+
|
|
70
|
+
A `PageTool` has the shape of `AppTool` from `genaicode/app-agent`, so one definition can serve
|
|
71
|
+
a coding agent on your server and the model in the page.
|
|
72
|
+
|
|
73
|
+
## How the model calls tools
|
|
74
|
+
|
|
75
|
+
The Prompt API has no tool calling, so `nano()` breaks every step into choices the browser
|
|
76
|
+
can constrain with a JSON Schema (`responseConstraint`):
|
|
77
|
+
|
|
78
|
+
1. Which action: one of the tool names, or `reply`.
|
|
79
|
+
2. The chosen tool's arguments, constrained by its `inputSchema`.
|
|
80
|
+
3. The tool runs; its result goes back and the model chooses again, up to `maxSteps` (default 6).
|
|
81
|
+
4. On `reply`, the answer streams as text.
|
|
82
|
+
|
|
83
|
+
The same exact call is never run twice in a task, bad arguments come back as an error the
|
|
84
|
+
model can fix, and a model that never picks `reply` is made to answer when the steps run out.
|
|
85
|
+
|
|
86
|
+
## The driver: `nano()`
|
|
87
|
+
|
|
88
|
+
`createPageAgent` is a conversation on top of `nano()`, which is a `CodingAgent` like every
|
|
89
|
+
driver in [`genaicode/agents`](agents.md): the same `run(task)`, events and result.
|
|
90
|
+
|
|
91
|
+
```ts
|
|
92
|
+
import { nano } from 'genaicode/browser';
|
|
93
|
+
|
|
94
|
+
const agent = nano({ instructions, tools });
|
|
95
|
+
if ((await agent.availability()) !== 'unavailable') {
|
|
96
|
+
const run = agent.run({ prompt: 'Which plans include SSO?', cwd: '' });
|
|
97
|
+
for await (const event of run) if (event.type === 'text-delta') render(event.text);
|
|
98
|
+
const { text, sessionId } = await run.result; // pass sessionId as task.resume to go on
|
|
99
|
+
}
|
|
100
|
+
```
|
|
101
|
+
|
|
102
|
+
- `availability()`: `available`, `downloadable`, `downloading` or `unavailable`.
|
|
103
|
+
- `prepare()` loads the model, downloading it first when needed. A download needs a user
|
|
104
|
+
gesture, so call it from a click; `onDownload(progress)` reports 0 to 1.
|
|
105
|
+
- Chrome keeps the first `responseConstraint` a session was given: a later prompt with another
|
|
106
|
+
schema fails with `UnknownError` or answers in the first one's shape. So `nano` asks every
|
|
107
|
+
constrained step on a copy of the conversation (`session.clone()`), tells it the task so far,
|
|
108
|
+
and lets the copy go; the conversation itself only hears the request for the reply.
|
|
109
|
+
- A schema-constrained call can still fail with an unexplained `UnknownError`. `nano` then asks
|
|
110
|
+
the model the same thing in words and checks the answer itself, and a reply that fails before
|
|
111
|
+
its first word is tried once more. An error that remains names the step: choosing the action,
|
|
112
|
+
filling in the arguments or writing the reply.
|
|
113
|
+
- `onModelCall({ step, ms, constrained, error })` reports every call to the model, for logs and
|
|
114
|
+
for finding the slow step.
|
|
115
|
+
- `task.onApproval` answers `write` tools (without it they are declined);
|
|
116
|
+
`task.permissions.approval` `'auto-approve'` and `'deny'` work as in every driver.
|
|
117
|
+
`permissions.sandbox: 'read-only'` refuses every `write` tool; a page has no other sandbox,
|
|
118
|
+
so `'workspace-write'`, `'unrestricted'` and `'yolo'` fail the task.
|
|
119
|
+
- `task.resume` continues a conversation in the same model session; `forget(sessionId)` frees it.
|
|
120
|
+
When the context window fills up, the browser drops the oldest messages and keeps the
|
|
121
|
+
instructions. `context(sessionId)` says how full it is.
|
|
122
|
+
- Usage is what the task added to the conversation (`usage.totalTokens`), and
|
|
123
|
+
`task.budget.maxTokens` stops a task that goes over it. There is no cost to cap.
|
|
124
|
+
- `task.cwd`, `env` and `mcpServers` mean nothing in a page. Pass `cwd: ''`.
|
|
125
|
+
|
|
126
|
+
`createPageAgent({ agent })` takes any other `CodingAgent` that runs in a browser instead.
|
|
127
|
+
|
|
128
|
+
## Files, images and voice
|
|
129
|
+
|
|
130
|
+
```ts
|
|
131
|
+
const assistant = createPageAgent({ tools, nano: { inputs: ['image', 'audio'] } });
|
|
132
|
+
assistant.send('What is on this screenshot?', [file]);
|
|
133
|
+
```
|
|
134
|
+
|
|
135
|
+
```tsx
|
|
136
|
+
<PageAgentChat agent={assistant} attachments />
|
|
137
|
+
```
|
|
138
|
+
|
|
139
|
+
`send(text, files)` takes `File`s or `Blob`s with a prompt, and `nano().run({ prompt, attachments })`
|
|
140
|
+
does the same one level down.
|
|
141
|
+
|
|
142
|
+
- A text file (by MIME type or extension: Markdown, JSON, CSV, source code) is read into the
|
|
143
|
+
prompt, up to `maxAttachmentText` characters (default 4,000). The context window is small, so a
|
|
144
|
+
long file is cut and the model is told so.
|
|
145
|
+
- An image or audio file goes to the model as it is, if `nano: { inputs }` asked for that kind
|
|
146
|
+
and the browser's model takes it. `assistant.get().model.inputs` (and `nano().inputs()`) lists
|
|
147
|
+
what it takes; elsewhere such a file fails the turn with a message that says why. Audio needs
|
|
148
|
+
a GPU in Chrome.
|
|
149
|
+
- Anything else (a PDF, an archive) is refused.
|
|
150
|
+
|
|
151
|
+
With `attachments`, `<PageAgentChat>` adds an attach button, takes pasted and dropped files, and
|
|
152
|
+
shows a microphone where the model takes audio: it records a voice message that is sent as an
|
|
153
|
+
audio file, so speech stays on the device too. The chat offers only the kinds the model takes.
|
|
154
|
+
`<AgentChat attachments={{ accept, voice, max }}>` has the same composer for any host:
|
|
155
|
+
`onSend(text, files)` gets the files, and a turn's `attachments` are drawn under its prompt.
|
|
156
|
+
|
|
157
|
+
## WebMCP
|
|
158
|
+
|
|
159
|
+
`exposeWebMcp(tools, { confirm, onCall })` registers the tools on `document.modelContext`
|
|
160
|
+
(or `navigator.modelContext` in the first implementations), so an agent outside the page,
|
|
161
|
+
such as the browser's own assistant, calls your functions instead of guessing at the DOM.
|
|
162
|
+
Read tools are marked `readOnlyHint`, `write` tools `consequentialHint`, and a `write` tool
|
|
163
|
+
runs only after `confirm(tool, args)` resolves true (default: the browser's `confirm()`
|
|
164
|
+
dialog). It returns a function that takes the tools back, and does nothing where WebMCP is
|
|
165
|
+
missing, so it is safe to call everywhere.
|
|
166
|
+
|
|
167
|
+
`webMcpTools()` goes the other way: it lists what the page already registered over WebMCP
|
|
168
|
+
(by any script, or by annotated forms) as `PageTool`s for the in-page agent. It returns an
|
|
169
|
+
empty list where the browser cannot list tools.
|
|
170
|
+
|
|
171
|
+
## Where it runs
|
|
172
|
+
|
|
173
|
+
| | |
|
|
174
|
+
| --------------- | ------------------------------------------------------------------------------------------ |
|
|
175
|
+
| Prompt API | Chrome 148+ on desktop (Windows, macOS, Linux, Chromebook Plus). Not on Android or iOS. |
|
|
176
|
+
| Hardware | About 22 GB of free disk, and 16 GB of RAM or a GPU with more than 4 GB of VRAM. |
|
|
177
|
+
| Languages | English, Japanese, Spanish, German, French (`nano({ language })`, default `en`). |
|
|
178
|
+
| First use | The browser downloads the model once (a few GB); after that it works offline. |
|
|
179
|
+
| WebMCP | A draft: Chrome origin trial (149 to 156), Edge behind a flag. Register a trial token. |
|
|
180
|
+
| Everything else | `availability()` is `unavailable`: show the page without the assistant, or use a fallback. |
|
|
181
|
+
|
|
182
|
+
Check `availability()` (or `<PageAgentChat>`'s `unavailable` slot) and treat the assistant as
|
|
183
|
+
an extra. A common setup loads it only where `'LanguageModel' in self`.
|
|
184
|
+
|
|
185
|
+
## Tests
|
|
186
|
+
|
|
187
|
+
`fakeLanguageModel(script)` stands in for the browser's model: replies come from `script` in
|
|
188
|
+
order and every prompt is recorded.
|
|
189
|
+
|
|
190
|
+
```ts
|
|
191
|
+
import { createPageAgent, fakeLanguageModel } from 'genaicode/browser';
|
|
192
|
+
|
|
193
|
+
const model = fakeLanguageModel(['{"action":"list_products"}', '{"action":"reply"}', 'Two products.']);
|
|
194
|
+
const assistant = createPageAgent({ tools, nano: { languageModel: model.factory } });
|
|
195
|
+
```
|
|
196
|
+
|
|
197
|
+
## Stability
|
|
198
|
+
|
|
199
|
+
Experimental, like the Prompt API and WebMCP themselves: names here may change in a minor
|
|
200
|
+
release.
|
package/docs/semver.md
CHANGED
|
@@ -23,6 +23,8 @@ These are covered by minor/patch compatibility within 2.x:
|
|
|
23
23
|
minor release until this section lists it as stable. See [agents.md](./agents.md).
|
|
24
24
|
- `genaicode/react` (`<AgentChat>` props, its `.gc-*` class names and `--gc-*` custom
|
|
25
25
|
properties) and `genaicode/ui` (`startEmbeddedWeb`), on the same terms.
|
|
26
|
+
- `genaicode/browser` (`nano`, `createPageAgent`, `exposeWebMcp`) and `<PageAgentChat>`, on the
|
|
27
|
+
same terms: the Prompt API and WebMCP under them are still moving.
|
|
26
28
|
|
|
27
29
|
## Additive changes (minor)
|
|
28
30
|
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "genaicode",
|
|
3
|
-
"version": "2.
|
|
3
|
+
"version": "2.33.0",
|
|
4
4
|
"description": "Drive the coding-agent CLIs you already have (Claude Code, Codex, Copilot, Cursor, Gemini, opencode...) behind one API and one CLI, plus a thin provider-neutral LLM layer.",
|
|
5
5
|
"author": "Grzegorz Tańczyk",
|
|
6
6
|
"repository": {
|
|
@@ -56,6 +56,10 @@
|
|
|
56
56
|
"types": "./dist/app-agent.d.ts",
|
|
57
57
|
"import": "./dist/app-agent.js"
|
|
58
58
|
},
|
|
59
|
+
"./browser": {
|
|
60
|
+
"types": "./dist/browser.d.ts",
|
|
61
|
+
"import": "./dist/browser.js"
|
|
62
|
+
},
|
|
59
63
|
"./vite": {
|
|
60
64
|
"types": "./dist/vite.d.ts",
|
|
61
65
|
"import": "./dist/vite.js"
|