@pikku/core 0.12.34 → 0.12.36

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.
@@ -5,6 +5,8 @@ import type {
5
5
  AIStreamChannel,
6
6
  } from '../wirings/ai-agent/ai-agent.types.js'
7
7
 
8
+ export type AIProviderOptions = Record<string, Record<string, unknown>>
9
+
8
10
  export type AIAgentRunnerParams = {
9
11
  model: string
10
12
  temperature?: number
@@ -36,12 +38,169 @@ export type AIAgentStepResult = {
36
38
  reasoningContent?: string
37
39
  }
38
40
 
41
+ export type AITranscriptionParams = {
42
+ model: string
43
+ audio: Uint8Array
44
+ providerOptions?: AIProviderOptions
45
+ maxRetries?: number
46
+ abortSignal?: AbortSignal
47
+ headers?: Record<string, string>
48
+ }
49
+
50
+ export type AITranscriptionResult = {
51
+ text: string
52
+ segments?: Array<{
53
+ text: string
54
+ startSecond: number
55
+ endSecond: number
56
+ }>
57
+ language?: string
58
+ durationInSeconds?: number
59
+ warnings?: unknown[]
60
+ providerMetadata?: Record<string, unknown>
61
+ responses?: unknown[]
62
+ }
63
+
64
+ export type AIGenerateSpeechParams = {
65
+ model: string
66
+ text: string
67
+ voice?: string
68
+ outputFormat?: string
69
+ instructions?: string
70
+ speed?: number
71
+ language?: string
72
+ providerOptions?: AIProviderOptions
73
+ maxRetries?: number
74
+ abortSignal?: AbortSignal
75
+ headers?: Record<string, string>
76
+ }
77
+
78
+ export type AIGenerateSpeechResult = {
79
+ audio: {
80
+ uint8Array: Uint8Array
81
+ base64: string
82
+ mediaType: string
83
+ format: string
84
+ }
85
+ warnings?: unknown[]
86
+ providerMetadata?: Record<string, unknown>
87
+ responses?: unknown[]
88
+ }
89
+
90
+ export type AIGenerateImagePrompt =
91
+ | string
92
+ | {
93
+ images: Array<Uint8Array | ArrayBuffer | string>
94
+ text?: string
95
+ mask?: Uint8Array | ArrayBuffer | string
96
+ }
97
+
98
+ export type AIGenerateImageParams = {
99
+ model: string
100
+ prompt: AIGenerateImagePrompt
101
+ n?: number
102
+ maxImagesPerCall?: number
103
+ size?: `${number}x${number}`
104
+ aspectRatio?: `${number}:${number}`
105
+ seed?: number
106
+ providerOptions?: AIProviderOptions
107
+ maxRetries?: number
108
+ abortSignal?: AbortSignal
109
+ headers?: Record<string, string>
110
+ }
111
+
112
+ export type AIGenerateImageResult = {
113
+ images: Array<{
114
+ uint8Array: Uint8Array
115
+ base64: string
116
+ mediaType: string
117
+ }>
118
+ warnings?: unknown[]
119
+ providerMetadata?: Record<string, unknown>
120
+ responses?: unknown[]
121
+ usage?: {
122
+ inputTokens?: number
123
+ outputTokens?: number
124
+ totalTokens?: number
125
+ }
126
+ }
127
+
128
+ export type AIEmbedParams = {
129
+ model: string
130
+ value: string
131
+ providerOptions?: AIProviderOptions
132
+ maxRetries?: number
133
+ abortSignal?: AbortSignal
134
+ headers?: Record<string, string>
135
+ }
136
+
137
+ export type AIEmbedResult = {
138
+ value: string
139
+ embedding: number[]
140
+ usage?: { tokens?: number }
141
+ warnings?: unknown[]
142
+ providerMetadata?: Record<string, unknown>
143
+ response?: unknown
144
+ }
145
+
146
+ export type AIEmbedManyParams = {
147
+ model: string
148
+ values: string[]
149
+ providerOptions?: AIProviderOptions
150
+ maxRetries?: number
151
+ abortSignal?: AbortSignal
152
+ headers?: Record<string, string>
153
+ maxParallelCalls?: number
154
+ }
155
+
156
+ export type AIEmbedManyResult = {
157
+ values: string[]
158
+ embeddings: number[][]
159
+ usage?: { tokens?: number }
160
+ warnings?: unknown[]
161
+ providerMetadata?: Record<string, unknown>
162
+ responses?: unknown[]
163
+ }
164
+
165
+ export type AIRerankParams<VALUE extends string | Record<string, unknown>> = {
166
+ model: string
167
+ query: string
168
+ documents: VALUE[]
169
+ topK?: number
170
+ providerOptions?: AIProviderOptions
171
+ maxRetries?: number
172
+ abortSignal?: AbortSignal
173
+ headers?: Record<string, string>
174
+ }
175
+
176
+ export type AIRerankResult<VALUE extends string | Record<string, unknown>> = {
177
+ ranking: Array<{
178
+ index: number
179
+ document: VALUE
180
+ score: number
181
+ }>
182
+ rerankedDocuments: VALUE[]
183
+ originalDocuments: VALUE[]
184
+ providerMetadata?: Record<string, unknown>
185
+ response?: unknown
186
+ }
187
+
39
188
  export interface AIAgentRunnerService {
40
189
  stream(
41
190
  params: AIAgentRunnerParams,
42
191
  channel: AIStreamChannel
43
192
  ): Promise<AIAgentStepResult>
44
193
  run(params: AIAgentRunnerParams): Promise<AIAgentStepResult>
194
+ transcribe?(params: AITranscriptionParams): Promise<AITranscriptionResult>
195
+ generateSpeech?(
196
+ params: AIGenerateSpeechParams
197
+ ): Promise<AIGenerateSpeechResult>
198
+ generateImage?(params: AIGenerateImageParams): Promise<AIGenerateImageResult>
199
+ embed?(params: AIEmbedParams): Promise<AIEmbedResult>
200
+ embedMany?(params: AIEmbedManyParams): Promise<AIEmbedManyResult>
201
+ rerank?<VALUE extends string | Record<string, unknown>>(
202
+ params: AIRerankParams<VALUE>
203
+ ): Promise<AIRerankResult<VALUE>>
45
204
  /** Return a new runner that uses the given API key for every LLM call.
46
205
  * Optional — runners that don't support per-key scoping leave this undefined. */
47
206
  withApiKey?(apiKey: string): AIAgentRunnerService
@@ -69,6 +69,20 @@ export type {
69
69
  AIAgentRunnerParams,
70
70
  AIAgentRunnerResult,
71
71
  AIAgentStepResult,
72
+ AIEmbedManyParams,
73
+ AIEmbedManyResult,
74
+ AIEmbedParams,
75
+ AIEmbedResult,
76
+ AIGenerateImageParams,
77
+ AIGenerateImagePrompt,
78
+ AIGenerateImageResult,
79
+ AIGenerateSpeechParams,
80
+ AIGenerateSpeechResult,
81
+ AIProviderOptions,
82
+ AIRerankParams,
83
+ AIRerankResult,
84
+ AITranscriptionParams,
85
+ AITranscriptionResult,
72
86
  AIAgentRunnerService,
73
87
  } from './ai-agent-runner-service.js'
74
88
  export type {
@@ -212,7 +212,6 @@ export class LocalMetaService implements MetaService {
212
212
  private secretsMetaCache: SecretDefinitionsMeta | null = null
213
213
  private credentialsMetaCache: CredentialDefinitionsMeta | null = null
214
214
  private variablesMetaCache: VariableDefinitionsMeta | null = null
215
- private emailMetaCache: EmailsMeta | null = null
216
215
  private middlewareGroupsMetaCache: MiddlewareGroupsMeta | null = null
217
216
  private permissionsGroupsMetaCache: PermissionsGroupsMeta | null = null
218
217
  private agentsMetaCache: AgentsMeta | null = null
@@ -262,7 +261,6 @@ export class LocalMetaService implements MetaService {
262
261
  this.secretsMetaCache = null
263
262
  this.credentialsMetaCache = null
264
263
  this.variablesMetaCache = null
265
- this.emailMetaCache = null
266
264
  this.middlewareGroupsMetaCache = null
267
265
  this.permissionsGroupsMetaCache = null
268
266
  this.agentsMetaCache = null
@@ -509,17 +507,20 @@ export class LocalMetaService implements MetaService {
509
507
  }
510
508
 
511
509
  async getEmailMeta(): Promise<EmailsMeta> {
512
- if (this.emailMetaCache) return this.emailMetaCache
513
-
510
+ // Read fresh every call — never cache. The email meta file is generated by
511
+ // `pikku all`/`pikku emails generate` and is regenerated DURING a long-lived
512
+ // session (sandbox boots the orchestrator before the user project's codegen
513
+ // has written it). Caching here once returned an empty {templates:{}} that
514
+ // then stuck for the whole session, leaving the console emails screen blank
515
+ // even after the file appeared. A local JSON read is essentially free.
514
516
  const content = await this.readFile('email/pikku-emails-meta.gen.json')
515
- this.emailMetaCache = content
517
+ return content
516
518
  ? JSON.parse(content)
517
519
  : {
518
520
  src: '',
519
521
  themeHash: '',
520
522
  templates: {},
521
523
  }
522
- return this.emailMetaCache!
523
524
  }
524
525
 
525
526
  async getEmailTemplateAssets(
@@ -6,6 +6,8 @@ export {
6
6
  } from './ai-agent-helpers.js'
7
7
  export { runAIAgent, resumeAIAgentSync } from './ai-agent-runner.js'
8
8
  export { streamAIAgent, resumeAIAgent } from './ai-agent-stream.js'
9
+ export { voiceInput } from './voice-input.js'
10
+ export { voiceOutput } from './voice-output.js'
9
11
  export {
10
12
  type RunAIAgentParams,
11
13
  type StreamAIAgentOptions,
@@ -0,0 +1,132 @@
1
+ import type { AIAgentRunnerService } from '../../services/ai-agent-runner-service.js'
2
+ import { pikkuAIMiddleware } from '../../types/core.types.js'
3
+ import type { AIContentPart } from './ai-agent.types.js'
4
+
5
+ function base64ToUint8Array(base64: string): Uint8Array {
6
+ const binary = atob(base64)
7
+ const bytes = new Uint8Array(binary.length)
8
+ for (let i = 0; i < binary.length; i++) {
9
+ bytes[i] = binary.charCodeAt(i)
10
+ }
11
+ return bytes
12
+ }
13
+
14
+ const MAX_AUDIO_SIZE = 50 * 1024 * 1024
15
+
16
+ // Portable SSRF guard: @pikku/core runs in edge runtimes (CF Workers) with no
17
+ // Node `dns`, so we cannot resolve hostnames to check for private targets.
18
+ // Reject the obvious internal literals; callers wanting stricter control pass
19
+ // an explicit `allowedAudioHosts` allowlist. (Does not defend against a public
20
+ // hostname that resolves to a private IP / DNS rebinding — out of reach here.)
21
+ function isPrivateHost(hostname: string): boolean {
22
+ const host = hostname.replace(/^\[|\]$/g, '').toLowerCase()
23
+ if (host === 'localhost' || host === '0.0.0.0' || host === '::1') return true
24
+ if (host.startsWith('fe80:') || host.startsWith('fc') || host.startsWith('fd'))
25
+ return true
26
+ const v4 = host.match(/^(\d{1,3})\.(\d{1,3})\.(\d{1,3})\.\d{1,3}$/)
27
+ if (v4) {
28
+ const [a, b] = [Number(v4[1]), Number(v4[2])]
29
+ if (a === 127 || a === 10 || a === 0) return true
30
+ if (a === 169 && b === 254) return true // link-local incl. cloud metadata
31
+ if (a === 172 && b >= 16 && b <= 31) return true
32
+ if (a === 192 && b === 168) return true
33
+ }
34
+ return false
35
+ }
36
+
37
+ async function fetchAsUint8Array(
38
+ url: string,
39
+ allowedAudioHosts?: string[]
40
+ ): Promise<Uint8Array> {
41
+ const parsed = new URL(url)
42
+ if (parsed.protocol !== 'https:' && parsed.protocol !== 'http:') {
43
+ throw new Error('Only HTTP(S) URLs are supported for audio')
44
+ }
45
+ if (allowedAudioHosts) {
46
+ if (!allowedAudioHosts.includes(parsed.hostname)) {
47
+ throw new Error(`Audio URL host is not allowed: ${parsed.hostname}`)
48
+ }
49
+ } else if (isPrivateHost(parsed.hostname)) {
50
+ throw new Error(
51
+ `Refusing to fetch audio from a private/internal host: ${parsed.hostname}`
52
+ )
53
+ }
54
+ const response = await fetch(url)
55
+ const contentLength = response.headers.get('content-length')
56
+ if (contentLength && parseInt(contentLength, 10) > MAX_AUDIO_SIZE) {
57
+ throw new Error('Audio file exceeds maximum size')
58
+ }
59
+ const buffer = await response.arrayBuffer()
60
+ if (buffer.byteLength > MAX_AUDIO_SIZE) {
61
+ throw new Error('Audio file exceeds maximum size')
62
+ }
63
+ return new Uint8Array(buffer)
64
+ }
65
+
66
+ export const voiceInput = (config?: {
67
+ language?: string
68
+ model?: string
69
+ allowedAudioHosts?: string[]
70
+ }) =>
71
+ pikkuAIMiddleware({
72
+ modifyInput: async (services, { messages, instructions }) => {
73
+ const transcribeAudio = (services as {
74
+ aiAgentRunner?: AIAgentRunnerService
75
+ }).aiAgentRunner?.transcribe
76
+ if (!transcribeAudio) return { messages, instructions }
77
+
78
+ const last = messages[messages.length - 1]
79
+ if (!last || last.role !== 'user' || typeof last.content === 'string') {
80
+ return { messages, instructions }
81
+ }
82
+
83
+ const parts = last.content as AIContentPart[]
84
+ if (!parts) return { messages, instructions }
85
+
86
+ const hasAudio = parts.some(
87
+ (p) => p.type === 'file' && !!p.mediaType?.startsWith('audio/')
88
+ )
89
+ if (!hasAudio) return { messages, instructions }
90
+
91
+ // Process parts in order, replacing each audio part with its transcription
92
+ // in place. Sequential, so no unbounded parallel downloads/transcriptions
93
+ // and original content ordering is preserved.
94
+ const updatedContent: AIContentPart[] = []
95
+ for (const p of parts) {
96
+ if (!(p.type === 'file' && p.mediaType?.startsWith('audio/'))) {
97
+ updatedContent.push(p)
98
+ continue
99
+ }
100
+ if (!config?.model) {
101
+ throw new Error(
102
+ 'voiceInput requires a transcription model (e.g. openai/whisper-1)'
103
+ )
104
+ }
105
+ const audioData = p.data
106
+ ? base64ToUint8Array(p.data)
107
+ : await fetchAsUint8Array(p.url!, config.allowedAudioHosts)
108
+ const result = await transcribeAudio({
109
+ model: config.model,
110
+ audio: audioData,
111
+ ...(config.language
112
+ ? {
113
+ providerOptions: {
114
+ openai: {
115
+ language: config.language,
116
+ },
117
+ },
118
+ }
119
+ : {}),
120
+ })
121
+ updatedContent.push({ type: 'text' as const, text: result.text })
122
+ }
123
+
124
+ return {
125
+ messages: [
126
+ ...messages.slice(0, -1),
127
+ { ...last, content: updatedContent },
128
+ ],
129
+ instructions,
130
+ }
131
+ },
132
+ })
@@ -0,0 +1,134 @@
1
+ import type { AIAgentRunnerService } from '../../services/ai-agent-runner-service.js'
2
+ import { pikkuAIMiddleware } from '../../types/core.types.js'
3
+
4
+ type VoiceOutputState = {
5
+ textBuffer?: string
6
+ }
7
+
8
+ function bufferToBase64(data: Uint8Array): string {
9
+ let binary = ''
10
+ for (let i = 0; i < data.length; i++) {
11
+ binary += String.fromCharCode(data[i]!)
12
+ }
13
+ return btoa(binary)
14
+ }
15
+
16
+ const SENTENCE_BOUNDARY = /[.!?]\s*$/
17
+
18
+ function isSentenceBoundary(text: string): boolean {
19
+ return SENTENCE_BOUNDARY.test(text)
20
+ }
21
+
22
+ async function synthesizeAudio(
23
+ aiAgentRunner: AIAgentRunnerService,
24
+ input: {
25
+ model: string
26
+ text: string
27
+ voice?: string
28
+ format?: string
29
+ instructions?: string
30
+ speed?: number
31
+ language?: string
32
+ }
33
+ ): Promise<{ bytes: Uint8Array; format: string }> {
34
+ const result = await aiAgentRunner.generateSpeech?.({
35
+ model: input.model,
36
+ text: input.text,
37
+ voice: input.voice,
38
+ outputFormat: input.format,
39
+ instructions: input.instructions,
40
+ speed: input.speed,
41
+ language: input.language,
42
+ })
43
+ if (!result) {
44
+ throw new Error(
45
+ 'voiceOutput requires an aiAgentRunner with generateSpeech support'
46
+ )
47
+ }
48
+ // Label chunks with the format the provider actually returned (config.format
49
+ // is only a request), falling back to the requested format then pcm16.
50
+ return {
51
+ bytes: result.audio.uint8Array,
52
+ format: result.audio.format || input.format || 'pcm16',
53
+ }
54
+ }
55
+
56
+ export const voiceOutput = (config?: {
57
+ model?: string
58
+ format?: string
59
+ voice?: string
60
+ instructions?: string
61
+ speed?: number
62
+ language?: string
63
+ }) =>
64
+ pikkuAIMiddleware<VoiceOutputState>({
65
+ modifyOutputStream: async (services, { event, state }) => {
66
+ const aiAgentRunner = (services as {
67
+ aiAgentRunner?: AIAgentRunnerService
68
+ }).aiAgentRunner
69
+ if (!aiAgentRunner?.generateSpeech) return event
70
+
71
+ if (event.type === 'done') {
72
+ const remaining = state.textBuffer ?? ''
73
+ if (remaining) {
74
+ state.textBuffer = ''
75
+ if (!config?.model) {
76
+ throw new Error(
77
+ 'voiceOutput requires a speech model (e.g. openai/tts-1)'
78
+ )
79
+ }
80
+ const audio = await synthesizeAudio(aiAgentRunner, {
81
+ model: config.model,
82
+ text: remaining,
83
+ voice: config?.voice,
84
+ format: config?.format,
85
+ instructions: config?.instructions,
86
+ speed: config?.speed,
87
+ language: config?.language,
88
+ })
89
+ return [
90
+ {
91
+ type: 'audio-delta' as const,
92
+ data: bufferToBase64(audio.bytes),
93
+ format: audio.format,
94
+ },
95
+ { type: 'audio-done' as const },
96
+ event,
97
+ ]
98
+ }
99
+ return [{ type: 'audio-done' as const }, event]
100
+ }
101
+
102
+ if (event.type !== 'text-delta') return event
103
+
104
+ state.textBuffer = `${state.textBuffer ?? ''}${event.text}`
105
+ if (!isSentenceBoundary(state.textBuffer)) return event
106
+
107
+ const text = state.textBuffer
108
+ state.textBuffer = ''
109
+
110
+ if (!config?.model) {
111
+ throw new Error(
112
+ 'voiceOutput requires a speech model (e.g. openai/tts-1)'
113
+ )
114
+ }
115
+ const audio = await synthesizeAudio(aiAgentRunner, {
116
+ model: config.model,
117
+ text,
118
+ voice: config?.voice,
119
+ format: config?.format,
120
+ instructions: config?.instructions,
121
+ speed: config?.speed,
122
+ language: config?.language,
123
+ })
124
+
125
+ return [
126
+ event,
127
+ {
128
+ type: 'audio-delta' as const,
129
+ data: bufferToBase64(audio.bytes),
130
+ format: audio.format,
131
+ },
132
+ ]
133
+ },
134
+ })