@things-factory/ai-client-base 10.1.20 → 10.1.26
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/dist-server/service/assistant/index.d.ts +1 -0
- package/dist-server/service/assistant/index.js +1 -0
- package/dist-server/service/assistant/index.js.map +1 -1
- package/dist-server/service/assistant/registry-assistant.d.ts +10 -10
- package/dist-server/service/assistant/registry-assistant.js +11 -9
- package/dist-server/service/assistant/registry-assistant.js.map +1 -1
- package/dist-server/service/assistant/tool-scope.d.ts +25 -0
- package/dist-server/service/assistant/tool-scope.js +71 -0
- package/dist-server/service/assistant/tool-scope.js.map +1 -0
- package/dist-server/service/index.d.ts +1 -0
- package/dist-server/service/index.js +1 -0
- package/dist-server/service/index.js.map +1 -1
- package/dist-server/service/testing/tool-call-context.d.ts +10 -0
- package/dist-server/service/testing/tool-call-context.js +13 -0
- package/dist-server/service/testing/tool-call-context.js.map +1 -0
- package/dist-server/service/tool-registry/index.d.ts +3 -3
- package/dist-server/service/tool-registry/index.js +2 -1
- package/dist-server/service/tool-registry/index.js.map +1 -1
- package/dist-server/service/tool-registry/registry.d.ts +16 -0
- package/dist-server/service/tool-registry/registry.js +36 -0
- package/dist-server/service/tool-registry/registry.js.map +1 -1
- package/dist-server/service/tool-registry/types.d.ts +52 -4
- package/dist-server/service/tool-registry/types.js +4 -4
- package/dist-server/service/tool-registry/types.js.map +1 -1
- package/dist-server/tsconfig.tsbuildinfo +1 -1
- package/docs/grounded-analysis-contract.md +1 -1
- package/package.json +2 -2
- package/server/service/analysis-contract.test.ts +0 -1
- package/server/service/assistant/index.ts +1 -0
- package/server/service/assistant/registry-assistant.test.ts +131 -4
- package/server/service/assistant/registry-assistant.ts +29 -24
- package/server/service/assistant/tool-scope.ts +86 -0
- package/server/service/index.ts +1 -0
- package/server/service/testing/tool-call-context.ts +28 -0
- package/server/service/tool-registry/index.ts +7 -3
- package/server/service/tool-registry/registry.test.ts +50 -1
- package/server/service/tool-registry/registry.ts +47 -0
- package/server/service/tool-registry/types.ts +60 -10
|
@@ -13,6 +13,7 @@
|
|
|
13
13
|
*/
|
|
14
14
|
import { registryAssistantChat } from './registry-assistant'
|
|
15
15
|
import { clearToolRegistry, registerToolCategory } from '../tool-registry/index.js'
|
|
16
|
+
import { toolCallContextFixture } from '../testing/tool-call-context.js'
|
|
16
17
|
import type { AIToolCall, AIToolChatResult } from '../types.js'
|
|
17
18
|
|
|
18
19
|
const TC = (name: string, args: any = {}, id = 't1'): AIToolCall =>
|
|
@@ -60,6 +61,7 @@ beforeEach(() => {
|
|
|
60
61
|
kind: 'write',
|
|
61
62
|
name: 'addComponent',
|
|
62
63
|
description: 'adds a component to the document',
|
|
64
|
+
privilege: { category: 'board', privilege: 'mutation' },
|
|
63
65
|
schema: { type: 'object', properties: {} },
|
|
64
66
|
builder: () => {
|
|
65
67
|
ran.push({ name: 'addComponent', args: {}, context: null })
|
|
@@ -74,7 +76,7 @@ beforeEach(() => {
|
|
|
74
76
|
guidance: 'Never state a quantity you did not read.',
|
|
75
77
|
specs: [
|
|
76
78
|
{
|
|
77
|
-
kind: '
|
|
79
|
+
kind: 'read',
|
|
78
80
|
name: 'readOrderState',
|
|
79
81
|
description: 'reads live order state',
|
|
80
82
|
schema: { type: 'object', properties: {} },
|
|
@@ -85,6 +87,40 @@ beforeEach(() => {
|
|
|
85
87
|
}
|
|
86
88
|
]
|
|
87
89
|
} as any)
|
|
90
|
+
|
|
91
|
+
/*
|
|
92
|
+
* A category whose tools change something outside the conversation, each behind the same
|
|
93
|
+
* privilege its GraphQL door uses. This is the shape of board-import's `materializeImportSession`,
|
|
94
|
+
* which creates a Board.
|
|
95
|
+
*/
|
|
96
|
+
registerToolCategory({
|
|
97
|
+
name: 'board-import',
|
|
98
|
+
guidance: 'Create a board only after the user confirms.',
|
|
99
|
+
specs: [
|
|
100
|
+
{
|
|
101
|
+
kind: 'read',
|
|
102
|
+
name: 'getImportSession',
|
|
103
|
+
description: 'reads an import session',
|
|
104
|
+
privilege: { category: 'board-import', privilege: 'query' },
|
|
105
|
+
schema: { type: 'object', properties: {} },
|
|
106
|
+
builder: (args: any, context: any) => {
|
|
107
|
+
ran.push({ name: 'getImportSession', args, context })
|
|
108
|
+
return { status: 'completed' }
|
|
109
|
+
}
|
|
110
|
+
},
|
|
111
|
+
{
|
|
112
|
+
kind: 'external',
|
|
113
|
+
name: 'materializeImportSession',
|
|
114
|
+
description: 'creates a board from an import session',
|
|
115
|
+
privilege: { category: 'board-import', privilege: 'mutation' },
|
|
116
|
+
schema: { type: 'object', properties: {} },
|
|
117
|
+
builder: (args: any, context: any) => {
|
|
118
|
+
ran.push({ name: 'materializeImportSession', args, context })
|
|
119
|
+
return { boardId: 'b-1' }
|
|
120
|
+
}
|
|
121
|
+
}
|
|
122
|
+
]
|
|
123
|
+
} as any)
|
|
88
124
|
})
|
|
89
125
|
|
|
90
126
|
const input = (over: any = {}) => ({
|
|
@@ -98,7 +134,7 @@ const input = (over: any = {}) => ({
|
|
|
98
134
|
describe('what the model is offered', () => {
|
|
99
135
|
it('★ only the kinds that answer — a write tool is not offered to a surface with no document', async () => {
|
|
100
136
|
const { client, seen } = recordingClient()
|
|
101
|
-
await registryAssistantChat(input({ client }))
|
|
137
|
+
await registryAssistantChat(input({ client, toolCategories: ['figure-authoring', 'plant-ops'] }))
|
|
102
138
|
|
|
103
139
|
const offered = seen[0].tools.map(tool => tool.name)
|
|
104
140
|
expect(offered).toEqual(expect.arrayContaining(['listFigures', 'readOrderState']))
|
|
@@ -147,7 +183,7 @@ describe('running a tool', () => {
|
|
|
147
183
|
const { client } = recordingClient([okResult(undefined, [TC('listFigures', { q: 'roller' })]), okResult('two of them')])
|
|
148
184
|
|
|
149
185
|
const result = await registryAssistantChat(
|
|
150
|
-
input({ client, hostContext: { surface: 'figure-modeller', figureId: 'fig-1' } })
|
|
186
|
+
input({ client, toolCategories: ['figure-authoring'], hostContext: { surface: 'figure-modeller', figureId: 'fig-1' } })
|
|
151
187
|
)
|
|
152
188
|
|
|
153
189
|
expect(ran).toHaveLength(1)
|
|
@@ -158,10 +194,12 @@ describe('running a tool', () => {
|
|
|
158
194
|
expect(result.reply).toContain('two of them')
|
|
159
195
|
})
|
|
160
196
|
|
|
197
|
+
|
|
161
198
|
it('★ a write tool call is not executed — the same set decides offering and dispatch', async () => {
|
|
162
199
|
const { client } = recordingClient([okResult(undefined, [TC('addComponent')]), okResult('cannot do that')])
|
|
163
200
|
|
|
164
|
-
|
|
201
|
+
/* The category is allowed; the kind is what keeps the write out. */
|
|
202
|
+
await registryAssistantChat(input({ client, toolCategories: ['figure-authoring'] }))
|
|
165
203
|
|
|
166
204
|
expect(ran.map(one => one.name)).not.toContain('addComponent')
|
|
167
205
|
})
|
|
@@ -175,6 +213,95 @@ describe('running a tool', () => {
|
|
|
175
213
|
})
|
|
176
214
|
})
|
|
177
215
|
|
|
216
|
+
describe('what a tool is handed (A3)', () => {
|
|
217
|
+
it('★ exactly the shared shape — state and host side by side', async () => {
|
|
218
|
+
const { client } = recordingClient([okResult(undefined, [TC('listFigures')]), okResult('ok')])
|
|
219
|
+
const state = { domain: { id: 'dom-1' }, user: { id: 'u-1' } }
|
|
220
|
+
|
|
221
|
+
await registryAssistantChat(input({ client, state, toolCategories: ['figure-authoring'], hostContext: { figureId: 'fig-1' } }))
|
|
222
|
+
|
|
223
|
+
expect(ran[0].context).toEqual(toolCallContextFixture({ domainId: 'dom-1', userId: 'u-1', host: { figureId: 'fig-1' } }))
|
|
224
|
+
expect(ran[0].context.state).not.toHaveProperty('host')
|
|
225
|
+
})
|
|
226
|
+
|
|
227
|
+
it('★ host is an empty object when the surface sent none — a tool never reads it off undefined', async () => {
|
|
228
|
+
const { client } = recordingClient([okResult(undefined, [TC('listFigures')]), okResult('ok')])
|
|
229
|
+
|
|
230
|
+
await registryAssistantChat(input({ client, toolCategories: ['figure-authoring'] }))
|
|
231
|
+
|
|
232
|
+
expect(ran[0].context.host).toEqual({})
|
|
233
|
+
})
|
|
234
|
+
})
|
|
235
|
+
|
|
236
|
+
describe('who may use a tool (A1)', () => {
|
|
237
|
+
const grantAll = async () => true
|
|
238
|
+
const denyAll = async () => false
|
|
239
|
+
|
|
240
|
+
it('★ omitting the category list offers nothing and runs nothing', async () => {
|
|
241
|
+
const { client, seen } = recordingClient([okResult(undefined, [TC('listFigures')]), okResult('done')])
|
|
242
|
+
|
|
243
|
+
await registryAssistantChat(input({ client, toolAccess: grantAll }))
|
|
244
|
+
|
|
245
|
+
expect(seen[0].tools).toEqual([])
|
|
246
|
+
expect(ran).toHaveLength(0)
|
|
247
|
+
})
|
|
248
|
+
|
|
249
|
+
it('★ a tool that changes something outside is not offered here, even with the privilege', async () => {
|
|
250
|
+
const { client, seen } = recordingClient([okResult(undefined, [TC('materializeImportSession')]), okResult('done')])
|
|
251
|
+
|
|
252
|
+
await registryAssistantChat(input({ client, toolCategories: ['board-import'], toolAccess: grantAll }))
|
|
253
|
+
|
|
254
|
+
expect(seen[0].tools.map((t: any) => t.name)).not.toContain('materializeImportSession')
|
|
255
|
+
expect(ran.map(one => one.name)).not.toContain('materializeImportSession')
|
|
256
|
+
})
|
|
257
|
+
|
|
258
|
+
it('★ a tool behind a privilege the caller lacks is neither offered nor run when called by name', async () => {
|
|
259
|
+
const { client, seen } = recordingClient([okResult(undefined, [TC('getImportSession')]), okResult('done')])
|
|
260
|
+
|
|
261
|
+
await registryAssistantChat(input({ client, toolCategories: ['board-import'], toolAccess: denyAll }))
|
|
262
|
+
|
|
263
|
+
expect(seen[0].tools.map((t: any) => t.name)).not.toContain('getImportSession')
|
|
264
|
+
expect(ran.map(one => one.name)).not.toContain('getImportSession')
|
|
265
|
+
})
|
|
266
|
+
|
|
267
|
+
it('★ with no access check at all, a tool behind a privilege stays closed', async () => {
|
|
268
|
+
const { client, seen } = recordingClient([okResult(undefined, [TC('getImportSession')]), okResult('done')])
|
|
269
|
+
|
|
270
|
+
await registryAssistantChat(input({ client, toolCategories: ['board-import'] }))
|
|
271
|
+
|
|
272
|
+
expect(seen[0].tools.map((t: any) => t.name)).not.toContain('getImportSession')
|
|
273
|
+
expect(ran.map(one => one.name)).not.toContain('getImportSession')
|
|
274
|
+
})
|
|
275
|
+
|
|
276
|
+
it('the access check is asked with the privilege the tool declares, and a grant opens the tool', async () => {
|
|
277
|
+
const asked: any[] = []
|
|
278
|
+
const { client, seen } = recordingClient([okResult(undefined, [TC('getImportSession')]), okResult('done')])
|
|
279
|
+
|
|
280
|
+
await registryAssistantChat(
|
|
281
|
+
input({
|
|
282
|
+
client,
|
|
283
|
+
toolCategories: ['board-import'],
|
|
284
|
+
toolAccess: async (privilege: any) => {
|
|
285
|
+
asked.push(privilege)
|
|
286
|
+
return privilege.privilege === 'query'
|
|
287
|
+
}
|
|
288
|
+
})
|
|
289
|
+
)
|
|
290
|
+
|
|
291
|
+
expect(asked).toContainEqual({ category: 'board-import', privilege: 'query' })
|
|
292
|
+
expect(seen[0].tools.map((t: any) => t.name)).toContain('getImportSession')
|
|
293
|
+
expect(ran.map(one => one.name)).toContain('getImportSession')
|
|
294
|
+
})
|
|
295
|
+
|
|
296
|
+
it('guidance of a category whose every tool is withheld is not read to the model', async () => {
|
|
297
|
+
const { client, seen } = recordingClient()
|
|
298
|
+
|
|
299
|
+
await registryAssistantChat(input({ client, toolCategories: ['board-import'], toolAccess: denyAll }))
|
|
300
|
+
|
|
301
|
+
expect(seen[0].options.systemPrompt).not.toContain('Create a board only after the user confirms.')
|
|
302
|
+
})
|
|
303
|
+
})
|
|
304
|
+
|
|
178
305
|
describe('a client that cannot call tools', () => {
|
|
179
306
|
it('★ refuses instead of answering from nothing', async () => {
|
|
180
307
|
await expect(registryAssistantChat(input({ client: {} as any }))).rejects.toThrow(/cannot call tools/)
|
|
@@ -9,28 +9,24 @@
|
|
|
9
9
|
* operato-figure hit exactly that: the dock was mounted, the tools were registered, and the only
|
|
10
10
|
* mutation that could run them lived in a package figure does not load.
|
|
11
11
|
*
|
|
12
|
-
* ── Read
|
|
12
|
+
* ── Read only ──────────────────────────────────────────────────────────────────
|
|
13
13
|
* A registry category may also carry `write` and `action` tools. Those produce edit ops for a
|
|
14
14
|
* document the host owns, and a host with no document cannot apply them. Offering them here would
|
|
15
15
|
* declare a door with no road behind it — the model would call one and the call would go nowhere.
|
|
16
|
-
*
|
|
17
|
-
*
|
|
16
|
+
* `external` tools change something outside the conversation (board-import's
|
|
17
|
+
* `materializeImportSession` creates a Board) and this door has no confirm-and-apply step, so they
|
|
18
|
+
* are not offered either (A1, 2026-09-17). This runner offers, and dispatches, `read` only.
|
|
18
19
|
*
|
|
19
20
|
* The gate answers from the same list the definitions are built from, so what is offered and what
|
|
20
21
|
* may be dispatched cannot drift apart.
|
|
21
22
|
*/
|
|
22
23
|
import type { AIClient, AIMessage } from '../types.js'
|
|
23
|
-
import {
|
|
24
|
+
import { narrowToolScope, resolveToolScope, toolScopeGuidance, type ToolAccessCheck } from './tool-scope.js'
|
|
24
25
|
import { assistantChat, type AssistantChatResult } from './assistant-chat.js'
|
|
25
26
|
import { summarizeToolResult } from './summarize.js'
|
|
26
27
|
import type { GroundingSources } from '../utils/index.js'
|
|
28
|
+
import type { ToolCallContext } from '../tool-registry/types.js'
|
|
27
29
|
|
|
28
|
-
/** What a registry tool is handed when it runs. `state` is the caller's koa state. */
|
|
29
|
-
export interface RegistryToolCallContext {
|
|
30
|
-
state: any
|
|
31
|
-
/** What this surface is looking at — the host passes it through; tools read it as `host`. */
|
|
32
|
-
host?: any
|
|
33
|
-
}
|
|
34
30
|
|
|
35
31
|
export interface RegistryAssistantInput {
|
|
36
32
|
/** Provider client. Must implement `chatWithTools`. */
|
|
@@ -40,12 +36,17 @@ export interface RegistryAssistantInput {
|
|
|
40
36
|
/** Who the assistant is and what this surface is. The host owns this prose. */
|
|
41
37
|
systemPrompt: string
|
|
42
38
|
/**
|
|
43
|
-
* Categories this conversation may use.
|
|
44
|
-
*
|
|
45
|
-
*
|
|
46
|
-
* no tools at all. That is a legitimate surface (plain chat), so it is not corrected here.
|
|
39
|
+
* Categories this conversation may use. **Omitting it allows nothing** — the same as an empty
|
|
40
|
+
* list, a plain conversation with no tools (A1, 2026-09-17). It used to mean every registered
|
|
41
|
+
* category, which is how a Board-creating tool reached any conversation.
|
|
47
42
|
*/
|
|
48
43
|
toolCategories?: string[]
|
|
44
|
+
/**
|
|
45
|
+
* Answers whether the caller holds a tool's declared privilege. Without it, every tool that
|
|
46
|
+
* declares a privilege stays closed. The host builds it from the request — this package does
|
|
47
|
+
* not know auth.
|
|
48
|
+
*/
|
|
49
|
+
toolAccess?: ToolAccessCheck
|
|
49
50
|
/** Passed to tools as `context.host`. Client-supplied — never used for authorization. */
|
|
50
51
|
hostContext?: any
|
|
51
52
|
/** Caller's koa state (domain, user, tx). Tools that read the database need it. */
|
|
@@ -64,11 +65,6 @@ export interface RegistryAssistantResult extends AssistantChatResult<unknown> {
|
|
|
64
65
|
offeredTools: string[]
|
|
65
66
|
}
|
|
66
67
|
|
|
67
|
-
/** `read` and `external` are the two kinds that answer a question without changing a document. */
|
|
68
|
-
function answeringSpecs(filter: ToolCategoryFilter): ToolSpec[] {
|
|
69
|
-
return getToolSpecsFor(filter).filter(spec => spec.kind === 'read' || spec.kind === 'external')
|
|
70
|
-
}
|
|
71
|
-
|
|
72
68
|
/**
|
|
73
69
|
* Run one turn of a conversation whose only tools come from the registry.
|
|
74
70
|
*
|
|
@@ -82,8 +78,17 @@ export async function registryAssistantChat(input: RegistryAssistantInput): Prom
|
|
|
82
78
|
throw new Error('[registry-assistant] the configured AI client cannot call tools, so this conversation cannot run')
|
|
83
79
|
}
|
|
84
80
|
|
|
85
|
-
|
|
86
|
-
|
|
81
|
+
/*
|
|
82
|
+
* Only `read` is offered here. `external` means the tool changes something outside the
|
|
83
|
+
* conversation (board-import's `materializeImportSession` creates a Board); this door answers
|
|
84
|
+
* questions and has no confirm-and-apply step, so a tool with an effect does not go through it
|
|
85
|
+
* even when the caller holds the privilege.
|
|
86
|
+
*/
|
|
87
|
+
const scope = narrowToolScope(
|
|
88
|
+
await resolveToolScope(toolCategories, input.toolAccess),
|
|
89
|
+
spec => spec.kind === 'read'
|
|
90
|
+
)
|
|
91
|
+
const specs = scope.specs
|
|
87
92
|
const byName = new Map(specs.map(spec => [spec.name, spec]))
|
|
88
93
|
|
|
89
94
|
const tools = specs.map(spec => ({
|
|
@@ -96,12 +101,12 @@ export async function registryAssistantChat(input: RegistryAssistantInput): Prom
|
|
|
96
101
|
* The rules for using a tool belong to whoever registered it. Collect only the exposed ones, so
|
|
97
102
|
* a surface never reads instructions for a category it cannot call.
|
|
98
103
|
*/
|
|
99
|
-
const guidance =
|
|
104
|
+
const guidance = toolScopeGuidance(scope)
|
|
100
105
|
const prompt = guidance.length > 0 ? `${systemPrompt}\n\n${guidance.join('\n\n')}` : systemPrompt
|
|
101
106
|
|
|
102
|
-
const toolContext:
|
|
107
|
+
const toolContext: ToolCallContext = { state, host: hostContext ?? {} }
|
|
103
108
|
|
|
104
|
-
const result = await assistantChat<
|
|
109
|
+
const result = await assistantChat<ToolCallContext, unknown>({
|
|
105
110
|
initialConversation: messages,
|
|
106
111
|
tools,
|
|
107
112
|
options: {
|
|
@@ -0,0 +1,86 @@
|
|
|
1
|
+
/*
|
|
2
|
+
* Which registered tools one conversation may use — decided once per request, before anything is
|
|
3
|
+
* offered, so offering and dispatch read the same list.
|
|
4
|
+
*
|
|
5
|
+
* ── Two questions, one list ───────────────────────────────────────────────────────
|
|
6
|
+
* Which categories did this surface ask for? `toolCategories`
|
|
7
|
+
* Which of their tools may this caller use? each spec's `privilege`, asked of `access`
|
|
8
|
+
*
|
|
9
|
+
* **Nothing asked means nothing offered.** Omitting the category list used to mean "every
|
|
10
|
+
* registered category"; that default is the one that opened board-import's Board-creating tool to
|
|
11
|
+
* any conversation. An absent answer leans to the closed side (ADR-0047 ② · ADR-0053).
|
|
12
|
+
*
|
|
13
|
+
* **No access check means privileged tools stay closed.** A caller that cannot say who is asking
|
|
14
|
+
* gets only the tools that need no privilege.
|
|
15
|
+
*
|
|
16
|
+
* The permission check itself belongs to auth, which this package does not depend on, so the host
|
|
17
|
+
* hands it in (`ai-assistant` builds it from the request state).
|
|
18
|
+
*/
|
|
19
|
+
import { getToolCategories, type ToolCategoryFilter } from '../tool-registry/index.js'
|
|
20
|
+
import type { ToolPrivilege, ToolSpec } from '../tool-registry/types.js'
|
|
21
|
+
|
|
22
|
+
/** Answers "may this caller use a door guarded by this privilege". */
|
|
23
|
+
export type ToolAccessCheck = (privilege: ToolPrivilege) => Promise<boolean>
|
|
24
|
+
|
|
25
|
+
export interface ScopedCategory {
|
|
26
|
+
readonly name: string
|
|
27
|
+
readonly guidance: string
|
|
28
|
+
/** The tools of this category this caller may use. Never empty — a category with none is left out. */
|
|
29
|
+
readonly specs: ReadonlyArray<ToolSpec>
|
|
30
|
+
}
|
|
31
|
+
|
|
32
|
+
export interface ToolScope {
|
|
33
|
+
/** Categories with at least one usable tool, in registration order. */
|
|
34
|
+
readonly categories: ReadonlyArray<ScopedCategory>
|
|
35
|
+
/** Every usable tool, flat. Offering and dispatch both read this. */
|
|
36
|
+
readonly specs: ReadonlyArray<ToolSpec>
|
|
37
|
+
}
|
|
38
|
+
|
|
39
|
+
export const EMPTY_TOOL_SCOPE: ToolScope = Object.freeze({ categories: [], specs: [] })
|
|
40
|
+
|
|
41
|
+
const privilegeKey = (p: ToolPrivilege) => [p.category, p.privilege, !!p.owner, !!p.super].join('|')
|
|
42
|
+
|
|
43
|
+
export async function resolveToolScope(requested: ToolCategoryFilter, access?: ToolAccessCheck): Promise<ToolScope> {
|
|
44
|
+
if (!requested || requested.length === 0) return EMPTY_TOOL_SCOPE
|
|
45
|
+
|
|
46
|
+
/* One check per distinct privilege — a category of ten read tools asks once. */
|
|
47
|
+
const answers = new Map<string, Promise<boolean>>()
|
|
48
|
+
const mayUse = (spec: ToolSpec): Promise<boolean> => {
|
|
49
|
+
const privilege = spec.privilege
|
|
50
|
+
if (!privilege) return Promise.resolve(true)
|
|
51
|
+
if (!access) return Promise.resolve(false)
|
|
52
|
+
const key = privilegeKey(privilege)
|
|
53
|
+
if (!answers.has(key)) answers.set(key, access(privilege).catch(() => false))
|
|
54
|
+
return answers.get(key)!
|
|
55
|
+
}
|
|
56
|
+
|
|
57
|
+
const categories: ScopedCategory[] = []
|
|
58
|
+
for (const category of getToolCategories()) {
|
|
59
|
+
if (!requested.includes(category.name)) continue
|
|
60
|
+
const allowed = await Promise.all(category.specs.map(mayUse))
|
|
61
|
+
const specs = category.specs.filter((_, i) => allowed[i])
|
|
62
|
+
if (specs.length > 0) categories.push({ name: category.name, guidance: String(category.guidance ?? '').trim(), specs })
|
|
63
|
+
}
|
|
64
|
+
return fromCategories(categories)
|
|
65
|
+
}
|
|
66
|
+
|
|
67
|
+
/**
|
|
68
|
+
* Keep only the tools a surface can run — e.g. a surface with no document keeps `read`. A category
|
|
69
|
+
* left with nothing drops out, so its guidance is not read to a model that cannot call it.
|
|
70
|
+
*/
|
|
71
|
+
export function narrowToolScope(scope: ToolScope, keep: (spec: ToolSpec) => boolean): ToolScope {
|
|
72
|
+
return fromCategories(
|
|
73
|
+
scope.categories
|
|
74
|
+
.map(category => ({ ...category, specs: category.specs.filter(keep) }))
|
|
75
|
+
.filter(category => category.specs.length > 0)
|
|
76
|
+
)
|
|
77
|
+
}
|
|
78
|
+
|
|
79
|
+
/** The rules of the categories in scope, for the system prompt. */
|
|
80
|
+
export function toolScopeGuidance(scope: ToolScope): string[] {
|
|
81
|
+
return scope.categories.map(category => category.guidance).filter(Boolean)
|
|
82
|
+
}
|
|
83
|
+
|
|
84
|
+
function fromCategories(categories: ScopedCategory[]): ToolScope {
|
|
85
|
+
return { categories, specs: categories.flatMap(category => category.specs) }
|
|
86
|
+
}
|
package/server/service/index.ts
CHANGED
|
@@ -0,0 +1,28 @@
|
|
|
1
|
+
/*
|
|
2
|
+
* The one way a test builds what a registered tool is handed (A3, 2026-09-17).
|
|
3
|
+
*
|
|
4
|
+
* The twin's space tools read `ctx.state.host` while the door passed `ctx.host`, and their tests
|
|
5
|
+
* stayed green because each test wrote its own context in the shape the tool expected. A fixture
|
|
6
|
+
* owned by each test agrees with its tool by construction. This one is owned by the package that
|
|
7
|
+
* defines the neutral door, and that door's tests check it hands tools exactly this shape — so a tool
|
|
8
|
+
* test built on it fails when the tool reads a field no door fills.
|
|
9
|
+
*/
|
|
10
|
+
import type { ToolCallContext, ToolCallMode } from '../tool-registry/types.js'
|
|
11
|
+
|
|
12
|
+
export interface ToolCallContextFixtureInput {
|
|
13
|
+
domainId?: string
|
|
14
|
+
userId?: string
|
|
15
|
+
/** Extra request-state fields a tool reads (e.g. `tx`). Merged after domain and user. */
|
|
16
|
+
state?: Record<string, any>
|
|
17
|
+
host?: Record<string, any>
|
|
18
|
+
mode?: ToolCallMode
|
|
19
|
+
}
|
|
20
|
+
|
|
21
|
+
export function toolCallContextFixture(input: ToolCallContextFixtureInput = {}): ToolCallContext {
|
|
22
|
+
const context: ToolCallContext = {
|
|
23
|
+
state: { domain: { id: input.domainId ?? 'domain-test' }, user: { id: input.userId ?? 'user-test' }, ...(input.state ?? {}) },
|
|
24
|
+
host: input.host ?? {}
|
|
25
|
+
}
|
|
26
|
+
if (input.mode) context.mode = input.mode
|
|
27
|
+
return context
|
|
28
|
+
}
|
|
@@ -13,15 +13,17 @@ export {
|
|
|
13
13
|
getToolKind,
|
|
14
14
|
hasTool,
|
|
15
15
|
getCategorySpecs,
|
|
16
|
-
clearToolRegistry
|
|
16
|
+
clearToolRegistry,
|
|
17
|
+
getToolRegistryWarnings
|
|
17
18
|
} from './registry'
|
|
18
19
|
|
|
19
20
|
/* 대화별 노출 제한 타입은 registry 소유(정의가 거기 있다). */
|
|
20
|
-
export type { ToolCategoryFilter } from './registry'
|
|
21
|
+
export type { ToolCategoryFilter, ToolRegistryWarning } from './registry'
|
|
21
22
|
|
|
22
23
|
export type {
|
|
23
24
|
ToolSpec,
|
|
24
25
|
ToolSpecBase,
|
|
26
|
+
ToolPrivilege,
|
|
25
27
|
ReadToolSpec,
|
|
26
28
|
WriteToolSpec,
|
|
27
29
|
ActionToolSpec,
|
|
@@ -33,5 +35,7 @@ export type {
|
|
|
33
35
|
ExternalToolBuilder,
|
|
34
36
|
ReadToolContext,
|
|
35
37
|
WriteToolContext,
|
|
36
|
-
ExternalToolContext
|
|
38
|
+
ExternalToolContext,
|
|
39
|
+
ToolCallContext,
|
|
40
|
+
ToolCallMode
|
|
37
41
|
} from './types'
|
|
@@ -14,7 +14,7 @@ import {
|
|
|
14
14
|
getToolKind,
|
|
15
15
|
hasTool,
|
|
16
16
|
getCategorySpecs,
|
|
17
|
-
clearToolRegistry, getCategoryGuidance } from './registry'
|
|
17
|
+
clearToolRegistry, getCategoryGuidance, getToolRegistryWarnings } from './registry'
|
|
18
18
|
import type { ToolCategory, ToolSpec } from './types'
|
|
19
19
|
|
|
20
20
|
// 테스트 간 격리
|
|
@@ -34,6 +34,8 @@ function makeCategory(name: string, specs: Partial<ToolSpec>[]): ToolCategory {
|
|
|
34
34
|
description: 'desc',
|
|
35
35
|
schema: {},
|
|
36
36
|
builder: () => ({}),
|
|
37
|
+
/* A tool that changes something must name its door (A1). The fixture does, like guidance. */
|
|
38
|
+
...(s.kind && s.kind !== 'read' ? { privilege: { category: 'fixture', privilege: 'mutation' } } : {}),
|
|
37
39
|
...s
|
|
38
40
|
})) as ToolSpec[]
|
|
39
41
|
}
|
|
@@ -174,6 +176,7 @@ describe('registry — spec 보존', () => {
|
|
|
174
176
|
kind: 'write',
|
|
175
177
|
name: 't1',
|
|
176
178
|
description: 'main desc',
|
|
179
|
+
privilege: { category: 'fixture', privilege: 'mutation' },
|
|
177
180
|
themeNote: '⚠ note',
|
|
178
181
|
schema: { type: 'object' },
|
|
179
182
|
builder: () => null
|
|
@@ -258,3 +261,49 @@ describe('registry — 지침 없는 카테고리는 등록되지 않는다', ()
|
|
|
258
261
|
expect(getToolCategories()[0].guidance).toContain('fixture guidance')
|
|
259
262
|
})
|
|
260
263
|
})
|
|
264
|
+
|
|
265
|
+
describe('registry — a tool that changes something names its door (A1, stage one: surface)', () => {
|
|
266
|
+
test.each(['write', 'action', 'external'] as const)(
|
|
267
|
+
'★ a %s tool without a privilege registers, and the omission is recorded as a code',
|
|
268
|
+
kind => {
|
|
269
|
+
registerToolCategory({
|
|
270
|
+
name: 'no-door',
|
|
271
|
+
guidance: '- fixture guidance for no-door',
|
|
272
|
+
specs: [{ kind, name: 'changesThings', description: 'd', schema: {}, builder: () => null } as any]
|
|
273
|
+
})
|
|
274
|
+
expect(getToolKind('changesThings')).toBe(kind)
|
|
275
|
+
expect(getToolRegistryWarnings()).toEqual([
|
|
276
|
+
expect.objectContaining({
|
|
277
|
+
code: 'tool-privilege-undeclared',
|
|
278
|
+
params: { category: 'no-door', tool: 'changesThings', kind }
|
|
279
|
+
})
|
|
280
|
+
])
|
|
281
|
+
}
|
|
282
|
+
)
|
|
283
|
+
|
|
284
|
+
test('a privilege with no privilege name is recorded the same way', () => {
|
|
285
|
+
registerToolCategory({
|
|
286
|
+
name: 'half-door',
|
|
287
|
+
guidance: '- fixture guidance for half-door',
|
|
288
|
+
specs: [
|
|
289
|
+
{ kind: 'external', name: 'halfDoor', description: 'd', schema: {}, builder: () => null, privilege: { category: 'board' } } as any
|
|
290
|
+
]
|
|
291
|
+
})
|
|
292
|
+
expect(getToolRegistryWarnings().map(w => w.params.tool)).toEqual(['halfDoor'])
|
|
293
|
+
})
|
|
294
|
+
|
|
295
|
+
test('a declared tool and a read tool leave no warning', () => {
|
|
296
|
+
registerToolCategory(makeCategory('declared', [{ name: 'w1', kind: 'write' }, { name: 'r1', kind: 'read' }]))
|
|
297
|
+
expect(getToolRegistryWarnings()).toEqual([])
|
|
298
|
+
})
|
|
299
|
+
|
|
300
|
+
test('re-registering a category replaces its warnings instead of piling them up', () => {
|
|
301
|
+
const undeclared = { kind: 'write', name: 'w9', description: 'd', schema: {}, builder: () => null } as any
|
|
302
|
+
registerToolCategory({ name: 'again', guidance: '- g', specs: [undeclared] })
|
|
303
|
+
registerToolCategory({ name: 'again', guidance: '- g', specs: [undeclared] })
|
|
304
|
+
expect(getToolRegistryWarnings()).toHaveLength(1)
|
|
305
|
+
|
|
306
|
+
registerToolCategory({ name: 'again', guidance: '- g', specs: [{ ...undeclared, privilege: { category: 'c', privilege: 'mutation' } }] })
|
|
307
|
+
expect(getToolRegistryWarnings()).toEqual([])
|
|
308
|
+
})
|
|
309
|
+
})
|
|
@@ -29,6 +29,24 @@ import type { ToolCategory, ToolSpec } from './types'
|
|
|
29
29
|
|
|
30
30
|
const categories: Map<string, ToolCategory> = new Map()
|
|
31
31
|
|
|
32
|
+
/** A registration that was accepted but is not yet what the contract requires. Code + params, not prose. */
|
|
33
|
+
export interface ToolRegistryWarning {
|
|
34
|
+
code: 'tool-privilege-undeclared'
|
|
35
|
+
params: { category: string; tool: string; kind: ToolSpec['kind'] }
|
|
36
|
+
/** For a person reading it; the code and params are what callers act on. */
|
|
37
|
+
message: string
|
|
38
|
+
}
|
|
39
|
+
|
|
40
|
+
const warnings: ToolRegistryWarning[] = []
|
|
41
|
+
|
|
42
|
+
/**
|
|
43
|
+
* What registration accepted but the contract does not yet allow. Stage one of A1 records these
|
|
44
|
+
* instead of refusing; a host shows them at boot, and when the list is empty the rule becomes a refusal.
|
|
45
|
+
*/
|
|
46
|
+
export function getToolRegistryWarnings(): ReadonlyArray<ToolRegistryWarning> {
|
|
47
|
+
return warnings.slice()
|
|
48
|
+
}
|
|
49
|
+
|
|
32
50
|
/**
|
|
33
51
|
* 카테고리 등록. tool name 충돌 시 throw, 이름 중복 등록 시 덮어쓰기 + warn.
|
|
34
52
|
*/
|
|
@@ -55,6 +73,34 @@ export function registerToolCategory(category: ToolCategory): void {
|
|
|
55
73
|
)
|
|
56
74
|
}
|
|
57
75
|
|
|
76
|
+
/*
|
|
77
|
+
* A tool that changes something must say which privilege it goes through (A1, 2026-09-17).
|
|
78
|
+
*
|
|
79
|
+
* The conversation door checks one privilege — may this person talk to the assistant. Without a
|
|
80
|
+
* privilege on the tool, that single check opened every tool behind it: a user holding only
|
|
81
|
+
* `ai-assistant:mutation` could reach board-import's `materializeImportSession` and create a
|
|
82
|
+
* Board.
|
|
83
|
+
*
|
|
84
|
+
* **This is stage one of two: surface, then refuse** (ADR-0037, "먼저 드러내고, 목록이 비면 막는다").
|
|
85
|
+
* Refusing now would stop every app whose tools predate the `privilege` field from booting, and
|
|
86
|
+
* those apps compile against an installed copy that does not have the field yet. So an undeclared
|
|
87
|
+
* tool still registers, and the omission is recorded as a code (`getToolRegistryWarnings`), not
|
|
88
|
+
* a log line — a warning only a log carries is not distinguishable from no warning. When the list
|
|
89
|
+
* is empty across the apps, this becomes a refusal.
|
|
90
|
+
*/
|
|
91
|
+
warnings.splice(0, warnings.length, ...warnings.filter(w => w.params.category !== category.name))
|
|
92
|
+
for (const spec of category.specs) {
|
|
93
|
+
if (spec.kind !== 'read' && !(spec.privilege?.category && spec.privilege?.privilege)) {
|
|
94
|
+
warnings.push({
|
|
95
|
+
code: 'tool-privilege-undeclared',
|
|
96
|
+
params: { category: category.name, tool: spec.name, kind: spec.kind },
|
|
97
|
+
message:
|
|
98
|
+
`tool '${spec.name}' in category '${category.name}' is '${spec.kind}' but declares no privilege — ` +
|
|
99
|
+
'declare the same { category, privilege } its GraphQL door uses'
|
|
100
|
+
})
|
|
101
|
+
}
|
|
102
|
+
}
|
|
103
|
+
|
|
58
104
|
if (categories.has(category.name)) {
|
|
59
105
|
console.warn(
|
|
60
106
|
`[board-ai/tool-registry] category '${category.name}' already registered — overwriting. ` +
|
|
@@ -180,4 +226,5 @@ export function getCategorySpecs(categoryName: string): ReadonlyArray<ToolSpec>
|
|
|
180
226
|
*/
|
|
181
227
|
export function clearToolRegistry(): void {
|
|
182
228
|
categories.clear()
|
|
229
|
+
warnings.length = 0
|
|
183
230
|
}
|