File size: 14,275 Bytes
064bfd6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 | import type { ToolResultBlockParam } from '@anthropic-ai/sdk/resources/index.mjs'
import memoize from 'lodash-es/memoize.js'
import { z } from 'zod/v4'
import {
type AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
logEvent,
} from '../../services/analytics/index.js'
import {
buildTool,
findToolByName,
type Tool,
type ToolDef,
type Tools,
} from '../../Tool.js'
import { logForDebugging } from '../../utils/debug.js'
import { lazySchema } from '../../utils/lazySchema.js'
import { escapeRegExp } from '../../utils/stringUtils.js'
import { isToolSearchEnabledOptimistic } from '../../utils/toolSearch.js'
import { getPrompt, isDeferredTool, TOOL_SEARCH_TOOL_NAME } from './prompt.js'
export const inputSchema = lazySchema(() =>
z.object({
query: z
.string()
.describe(
'Query to find deferred tools. Use "select:<tool_name>" for direct selection, or keywords to search.',
),
max_results: z
.number()
.optional()
.default(5)
.describe('Maximum number of results to return (default: 5)'),
}),
)
type InputSchema = ReturnType<typeof inputSchema>
export const outputSchema = lazySchema(() =>
z.object({
matches: z.array(z.string()),
query: z.string(),
total_deferred_tools: z.number(),
pending_mcp_servers: z.array(z.string()).optional(),
}),
)
type OutputSchema = ReturnType<typeof outputSchema>
export type Output = z.infer<OutputSchema>
// Track deferred tool names to detect when cache should be cleared
let cachedDeferredToolNames: string | null = null
/**
* Get a cache key representing the current set of deferred tools.
*/
function getDeferredToolsCacheKey(deferredTools: Tools): string {
return deferredTools
.map(t => t.name)
.sort()
.join(',')
}
/**
* Get tool description, memoized by tool name.
* Used for keyword search scoring.
*/
const getToolDescriptionMemoized = memoize(
async (toolName: string, tools: Tools): Promise<string> => {
const tool = findToolByName(tools, toolName)
if (!tool) {
return ''
}
return tool.prompt({
getToolPermissionContext: async () => ({
mode: 'default' as const,
additionalWorkingDirectories: new Map(),
alwaysAllowRules: {},
alwaysDenyRules: {},
alwaysAskRules: {},
isBypassPermissionsModeAvailable: false,
}),
tools,
agents: [],
})
},
(toolName: string) => toolName,
)
/**
* Invalidate the description cache if deferred tools have changed.
*/
function maybeInvalidateCache(deferredTools: Tools): void {
const currentKey = getDeferredToolsCacheKey(deferredTools)
if (cachedDeferredToolNames !== currentKey) {
logForDebugging(
`ToolSearchTool: cache invalidated - deferred tools changed`,
)
getToolDescriptionMemoized.cache.clear?.()
cachedDeferredToolNames = currentKey
}
}
export function clearToolSearchDescriptionCache(): void {
getToolDescriptionMemoized.cache.clear?.()
cachedDeferredToolNames = null
}
/**
* Build the search result output structure.
*/
function buildSearchResult(
matches: string[],
query: string,
totalDeferredTools: number,
pendingMcpServers?: string[],
): { data: Output } {
return {
data: {
matches,
query,
total_deferred_tools: totalDeferredTools,
...(pendingMcpServers && pendingMcpServers.length > 0
? { pending_mcp_servers: pendingMcpServers }
: {}),
},
}
}
/**
* Parse tool name into searchable parts.
* Handles both MCP tools (mcp__server__action) and regular tools (CamelCase).
*/
function parseToolName(name: string): {
parts: string[]
full: string
isMcp: boolean
} {
// Check if it's an MCP tool
if (name.startsWith('mcp__')) {
const withoutPrefix = name.replace(/^mcp__/, '').toLowerCase()
const parts = withoutPrefix.split('__').flatMap(p => p.split('_'))
return {
parts: parts.filter(Boolean),
full: withoutPrefix.replace(/__/g, ' ').replace(/_/g, ' '),
isMcp: true,
}
}
// Regular tool - split by CamelCase and underscores
const parts = name
.replace(/([a-z])([A-Z])/g, '$1 $2') // CamelCase to spaces
.replace(/_/g, ' ')
.toLowerCase()
.split(/\s+/)
.filter(Boolean)
return {
parts,
full: parts.join(' '),
isMcp: false,
}
}
/**
* Pre-compile word-boundary regexes for all search terms.
* Called once per search instead of tools×terms×2 times.
*/
function compileTermPatterns(terms: string[]): Map<string, RegExp> {
const patterns = new Map<string, RegExp>()
for (const term of terms) {
if (!patterns.has(term)) {
patterns.set(term, new RegExp(`\\b${escapeRegExp(term)}\\b`))
}
}
return patterns
}
/**
* Keyword-based search over tool names and descriptions.
* Handles both MCP tools (mcp__server__action) and regular tools (CamelCase).
*
* The model typically queries with:
* - Server names when it knows the integration (e.g., "slack", "github")
* - Action words when looking for functionality (e.g., "read", "list", "create")
* - Tool-specific terms (e.g., "notebook", "shell", "kill")
*/
async function searchToolsWithKeywords(
query: string,
deferredTools: Tools,
tools: Tools,
maxResults: number,
): Promise<string[]> {
const queryLower = query.toLowerCase().trim()
// Fast path: if query matches a tool name exactly, return it directly.
// Handles models using a bare tool name instead of select: prefix (seen
// from subagents/post-compaction). Checks deferred first, then falls back
// to the full tool set — selecting an already-loaded tool is a harmless
// no-op that lets the model proceed without retry churn.
const exactMatch =
deferredTools.find(t => t.name.toLowerCase() === queryLower) ??
tools.find(t => t.name.toLowerCase() === queryLower)
if (exactMatch) {
return [exactMatch.name]
}
// If query looks like an MCP tool prefix (mcp__server), find matching tools.
// Handles models searching by server name with mcp__ prefix.
if (queryLower.startsWith('mcp__') && queryLower.length > 5) {
const prefixMatches = deferredTools
.filter(t => t.name.toLowerCase().startsWith(queryLower))
.slice(0, maxResults)
.map(t => t.name)
if (prefixMatches.length > 0) {
return prefixMatches
}
}
const queryTerms = queryLower.split(/\s+/).filter(term => term.length > 0)
// Partition into required (+prefixed) and optional terms
const requiredTerms: string[] = []
const optionalTerms: string[] = []
for (const term of queryTerms) {
if (term.startsWith('+') && term.length > 1) {
requiredTerms.push(term.slice(1))
} else {
optionalTerms.push(term)
}
}
const allScoringTerms =
requiredTerms.length > 0 ? [...requiredTerms, ...optionalTerms] : queryTerms
const termPatterns = compileTermPatterns(allScoringTerms)
// Pre-filter to tools matching ALL required terms in name or description
let candidateTools = deferredTools
if (requiredTerms.length > 0) {
const matches = await Promise.all(
deferredTools.map(async tool => {
const parsed = parseToolName(tool.name)
const description = await getToolDescriptionMemoized(tool.name, tools)
const descNormalized = description.toLowerCase()
const hintNormalized = tool.searchHint?.toLowerCase() ?? ''
const matchesAll = requiredTerms.every(term => {
const pattern = termPatterns.get(term)!
return (
parsed.parts.includes(term) ||
parsed.parts.some(part => part.includes(term)) ||
pattern.test(descNormalized) ||
(hintNormalized && pattern.test(hintNormalized))
)
})
return matchesAll ? tool : null
}),
)
candidateTools = matches.filter((t): t is Tool => t !== null)
}
const scored = await Promise.all(
candidateTools.map(async tool => {
const parsed = parseToolName(tool.name)
const description = await getToolDescriptionMemoized(tool.name, tools)
const descNormalized = description.toLowerCase()
const hintNormalized = tool.searchHint?.toLowerCase() ?? ''
let score = 0
for (const term of allScoringTerms) {
const pattern = termPatterns.get(term)!
// Exact part match (high weight for MCP server names, tool name parts)
if (parsed.parts.includes(term)) {
score += parsed.isMcp ? 12 : 10
} else if (parsed.parts.some(part => part.includes(term))) {
score += parsed.isMcp ? 6 : 5
}
// Full name fallback (for edge cases)
if (parsed.full.includes(term) && score === 0) {
score += 3
}
// searchHint match — curated capability phrase, higher signal than prompt
if (hintNormalized && pattern.test(hintNormalized)) {
score += 4
}
// Description match - use word boundary to avoid false positives
if (pattern.test(descNormalized)) {
score += 2
}
}
return { name: tool.name, score }
}),
)
return scored
.filter(item => item.score > 0)
.sort((a, b) => b.score - a.score)
.slice(0, maxResults)
.map(item => item.name)
}
export const ToolSearchTool = buildTool({
isEnabled() {
return isToolSearchEnabledOptimistic()
},
isConcurrencySafe() {
return true
},
isReadOnly() {
return true
},
name: TOOL_SEARCH_TOOL_NAME,
maxResultSizeChars: 100_000,
async description() {
return getPrompt()
},
async prompt() {
return getPrompt()
},
get inputSchema(): InputSchema {
return inputSchema()
},
get outputSchema(): OutputSchema {
return outputSchema()
},
async call(input, { options: { tools }, getAppState }) {
const { query, max_results = 5 } = input
const deferredTools = tools.filter(isDeferredTool)
maybeInvalidateCache(deferredTools)
// Check for MCP servers still connecting
function getPendingServerNames(): string[] | undefined {
const appState = getAppState()
const pending = appState.mcp.clients.filter(c => c.type === 'pending')
return pending.length > 0 ? pending.map(s => s.name) : undefined
}
// Helper to log search outcome
function logSearchOutcome(
matches: string[],
queryType: 'select' | 'keyword',
): void {
logEvent('tengu_tool_search_outcome', {
query:
query as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
queryType:
queryType as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
matchCount: matches.length,
totalDeferredTools: deferredTools.length,
maxResults: max_results,
hasMatches: matches.length > 0,
})
}
// Check for select: prefix — direct tool selection.
// Supports comma-separated multi-select: `select:A,B,C`.
// If a name isn't in the deferred set but IS in the full tool set,
// we still return it — the tool is already loaded, so "selecting" it
// is a harmless no-op that lets the model proceed without retry churn.
const selectMatch = query.match(/^select:(.+)$/i)
if (selectMatch) {
const requested = selectMatch[1]!
.split(',')
.map(s => s.trim())
.filter(Boolean)
const found: string[] = []
const missing: string[] = []
for (const toolName of requested) {
const tool =
findToolByName(deferredTools, toolName) ??
findToolByName(tools, toolName)
if (tool) {
if (!found.includes(tool.name)) found.push(tool.name)
} else {
missing.push(toolName)
}
}
if (found.length === 0) {
logForDebugging(
`ToolSearchTool: select failed — none found: ${missing.join(', ')}`,
)
logSearchOutcome([], 'select')
const pendingServers = getPendingServerNames()
return buildSearchResult(
[],
query,
deferredTools.length,
pendingServers,
)
}
if (missing.length > 0) {
logForDebugging(
`ToolSearchTool: partial select — found: ${found.join(', ')}, missing: ${missing.join(', ')}`,
)
} else {
logForDebugging(`ToolSearchTool: selected ${found.join(', ')}`)
}
logSearchOutcome(found, 'select')
return buildSearchResult(found, query, deferredTools.length)
}
// Keyword search
const matches = await searchToolsWithKeywords(
query,
deferredTools,
tools,
max_results,
)
logForDebugging(
`ToolSearchTool: keyword search for "${query}", found ${matches.length} matches`,
)
logSearchOutcome(matches, 'keyword')
// Include pending server info when search finds no matches
if (matches.length === 0) {
const pendingServers = getPendingServerNames()
return buildSearchResult(
matches,
query,
deferredTools.length,
pendingServers,
)
}
return buildSearchResult(matches, query, deferredTools.length)
},
renderToolUseMessage() {
return null
},
userFacingName: () => '',
/**
* Returns a tool_result with tool_reference blocks.
* This format works on 1P/Foundry. Bedrock/Vertex may not support
* client-side tool_reference expansion yet.
*/
mapToolResultToToolResultBlockParam(
content: Output,
toolUseID: string,
): ToolResultBlockParam {
if (content.matches.length === 0) {
let text = 'No matching deferred tools found'
if (
content.pending_mcp_servers &&
content.pending_mcp_servers.length > 0
) {
text += `. Some MCP servers are still connecting: ${content.pending_mcp_servers.join(', ')}. Their tools will become available shortly — try searching again.`
}
return {
type: 'tool_result',
tool_use_id: toolUseID,
content: text,
}
}
return {
type: 'tool_result',
tool_use_id: toolUseID,
content: content.matches.map(name => ({
type: 'tool_reference' as const,
tool_name: name,
})),
} as unknown as ToolResultBlockParam
},
} satisfies ToolDef<InputSchema, Output>)
|