# Conflicts: # .agents/notes/implemented/feature/2026-06-30-interception-seams.md # docs/config-catalog.md # docs/cookbook/adding-a-tool.i18n.yaml # docs/cookbook/adding-a-tool.md # docs/cookbook/adding-a-tool.zh.md # docs/cordis-catalog/events.md # docs/cordis-catalog/services.md # docs/core-data-structures/tools.md # docs/event-producer-consumer.md # docs/persistence-catalog.md # examples/acp-agent/tests/snapshots/cordis-inspect-jsdoc/session.jsonl # examples/acp-agent/tests/snapshots/cordis-inspect-jsdoc/stdout.expected.jsonl # packages/bash/tool-bash/src/index.ts # packages/core/agent-loop/src/tool-calls.ts # packages/core/agent-loop/tests/cancel.spec.ts # packages/core/agent-loop/tests/contract-regressions.spec.ts # packages/core/agent-loop/tests/tool-calls.spec.ts # packages/core/tools/README.md # packages/core/tools/src/index.ts # packages/core/tools/tests/code-mode.spec.ts # packages/core/tools/tests/tools.spec.ts # packages/fs/tool-fs-search/tests/integration.spec.ts # packages/fs/tool-fs-search/tests/tools.spec.ts # packages/fs/tool-fs/tests/integration.spec.ts # packages/mcp/mcp-client/src/tools.ts # packages/timeout/timeout-policy/tests/timeout-policy.spec.ts # packages/web/tool-web/tests/integration.spec.ts # packages/web/tool-web/tests/tool-web.spec.ts
315 lines
12 KiB
TypeScript
315 lines
12 KiB
TypeScript
/**
|
|
* Tool bridge: discovers MCP tools, registers them on the harness ToolRegistry
|
|
* under deterministic server-qualified public names, and handles re-sync when
|
|
* the server's tool list changes.
|
|
*
|
|
* Naming contract (see the mcp-client Agent Note "Naming invariants"): every MCP tool
|
|
* has the stable identity `(serverName, rawName)`; the model-facing public name
|
|
* is `mcp__<serverName>__<rawName>`, normalized to the DeepSeek function-name
|
|
* constraints. The raw name is only ever sent on the wire (`tools/call`); the
|
|
* public name is never parsed to recover it.
|
|
*
|
|
* @module
|
|
*/
|
|
|
|
import { createHash } from 'node:crypto'
|
|
import type { Client } from '@modelcontextprotocol/sdk/client/index.js'
|
|
import { ListToolsResultSchema } from '@modelcontextprotocol/sdk/types.js'
|
|
import { z } from 'zod'
|
|
import type { Context } from 'cordis'
|
|
import type { ToolDefinition, ToolExecution } from '@deepseek-ai/dsh-tools'
|
|
import { assertSupportedJsonSchema } from '@deepseek-ai/dsh-tools'
|
|
import type { JsonSchemaNode, JsonValue } from '@deepseek-ai/dsh-tools'
|
|
|
|
/** Resolved options relevant to tool bridging. */
|
|
export interface ToolBridgeOptions {
|
|
serverName: string
|
|
toolCallTimeoutMs: number
|
|
}
|
|
|
|
/** State for one sync generation: the current set of disposers keyed by public name. */
|
|
export type ToolDisposers = Map<string, () => void>
|
|
|
|
/** Canonical MCP result exposed to Code Mode without discarding protocol blocks. */
|
|
export type McpResult<Structured extends JsonValue = JsonValue> = {
|
|
content: JsonValue[]
|
|
structuredContent?: Structured
|
|
}
|
|
|
|
/**
|
|
* DeepSeek function-name contract: at most 64 characters. Wire-protocol
|
|
* constant, not configuration.
|
|
*/
|
|
const MAX_PUBLIC_NAME_LENGTH = 64
|
|
|
|
/** DeepSeek function-name contract: only `[A-Za-z0-9_-]` is allowed. */
|
|
const INVALID_NAME_CHARS = /[^A-Za-z0-9_-]/g
|
|
|
|
/** Hex chars of the SHA-256 identity hash appended on lossy normalization. */
|
|
const HASH_LENGTH = 12
|
|
|
|
/** Raw result record: the bridge owns JSON-value validation after transport. */
|
|
const RawCallToolResultSchema = z.record(z.string(), z.unknown())
|
|
|
|
/** List without mutating the SDK's per-page output-validator cache. */
|
|
function listToolsUncached(client: Client, cursor?: string) {
|
|
return client.request(
|
|
{ method: 'tools/list', ...cursor === undefined ? {} : { params: { cursor } } },
|
|
ListToolsResultSchema,
|
|
)
|
|
}
|
|
|
|
/** Call without the SDK pre-validating an output schema the bridge may not support. */
|
|
function callToolUncached(
|
|
client: Client,
|
|
rawName: string,
|
|
args: Record<string, unknown>,
|
|
exec: ToolExecution,
|
|
opts: ToolBridgeOptions,
|
|
) {
|
|
return client.request(
|
|
{ method: 'tools/call', params: { name: rawName, arguments: args } },
|
|
RawCallToolResultSchema,
|
|
{
|
|
signal: exec.signal,
|
|
timeout: opts.toolCallTimeoutMs,
|
|
},
|
|
)
|
|
}
|
|
|
|
/**
|
|
* Derive the model-facing public name for one MCP tool.
|
|
*
|
|
* Deterministic pure function of `(serverName, rawName)`: the clean case is
|
|
* `mcp__<serverName>__<rawName>` verbatim. When character replacement or
|
|
* truncation to the DeepSeek function-name contract (64 chars,
|
|
* `[A-Za-z0-9_-]`) changes the name, a 12-hex-char SHA-256 hash of the
|
|
* identity is appended so distinct MCP identities never collapse into the
|
|
* same public name.
|
|
*
|
|
* @param serverName - Stable local namespace from plugin config.
|
|
* @param rawName - The MCP server's own tool name.
|
|
* @returns The globally unique, model-facing ToolRegistry name.
|
|
*/
|
|
export function publicToolName(serverName: string, rawName: string): string {
|
|
const joined = `mcp__${serverName}__${rawName}`
|
|
const normalized = joined.replace(INVALID_NAME_CHARS, '_')
|
|
if (normalized === joined && normalized.length <= MAX_PUBLIC_NAME_LENGTH) return normalized
|
|
const hash = createHash('sha256').update(`${serverName}\0${rawName}`).digest('hex').slice(0, HASH_LENGTH)
|
|
return `${normalized.slice(0, MAX_PUBLIC_NAME_LENGTH - HASH_LENGTH - 1)}_${hash}`
|
|
}
|
|
|
|
/**
|
|
* Sync the MCP server's tool list into the harness ToolRegistry.
|
|
*
|
|
* Two phases keep the swap safe:
|
|
*
|
|
* 1. Fetch: drain uncached `tools/list` pagination and build the full next
|
|
* generation of `ToolDefinition`s under public names. Any failure here
|
|
* (network error, duplicate raw name in the server's list) rejects and
|
|
* leaves the previous generation registered untouched.
|
|
* 2. Swap: dispose the previous generation, register the new one. A registry
|
|
* conflict here can only mean a foreign registration squats on this
|
|
* server's `mcp__<serverName>__` namespace — the partial generation is
|
|
* rolled back (zero tools from this server), the error is logged, and an
|
|
* empty map is returned.
|
|
*
|
|
* @param client - Connected MCP Client instance used to list and call tools.
|
|
* @param ctx - Cordis context providing the `tools` service for registration.
|
|
* @param opts - Bridge options: server namespace and per-call timeout.
|
|
* @param previous - Disposer map from the prior sync generation; disposed
|
|
* during the swap phase (only after the fetch phase succeeded).
|
|
* @returns A map of registered public tool names to their unregister
|
|
* disposers — the exact set of live registrations owned by this server.
|
|
*/
|
|
export async function syncTools(
|
|
client: Client,
|
|
ctx: Context,
|
|
opts: ToolBridgeOptions,
|
|
previous: ToolDisposers,
|
|
): Promise<ToolDisposers> {
|
|
// Phase 1: fetch and build the next generation without touching the registry.
|
|
const definitions = new Map<string, ToolDefinition>()
|
|
let cursor: string | undefined
|
|
do {
|
|
const response = await listToolsUncached(client, cursor)
|
|
for (const tool of response.tools) {
|
|
const publicName = publicToolName(opts.serverName, tool.name)
|
|
if (definitions.has(publicName)) {
|
|
throw new Error(
|
|
`mcp-client(${opts.serverName}): server listed tool "${tool.name}" more than once — invalid tool list`,
|
|
)
|
|
}
|
|
definitions.set(publicName, {
|
|
name: publicName,
|
|
description: tool.description ?? '',
|
|
parameters: tool.inputSchema,
|
|
output: createOutput(tool.name, supportedOutputSchema(tool.outputSchema)),
|
|
execute: createExecutor(client, tool.name, tool.execution?.taskSupport === 'required', opts),
|
|
})
|
|
}
|
|
cursor = response.nextCursor
|
|
} while (cursor)
|
|
|
|
// Phase 2: swap generations.
|
|
for (const dispose of previous.values()) dispose()
|
|
const disposers: ToolDisposers = new Map()
|
|
try {
|
|
for (const [publicName, definition] of definitions) {
|
|
disposers.set(publicName, ctx.tools.register(definition))
|
|
}
|
|
} catch (error) {
|
|
// A conflict on an `mcp__<serverName>__`-qualified name means a foreign
|
|
// registration occupies this server's namespace. Roll back so the model
|
|
// sees either the full generation or none of it — never a partial set.
|
|
for (const dispose of disposers.values()) dispose()
|
|
ctx.logger.error(`mcp-client(${opts.serverName}): tool registration failed, no tools registered: ${String(error)}`)
|
|
return new Map()
|
|
}
|
|
return disposers
|
|
}
|
|
|
|
/**
|
|
* The shape we read from each MCP content block. Intentionally looser than the
|
|
* SDK's `ContentBlock` type: we're at a network trust boundary (data arrives
|
|
* from an external MCP server process via JSON-RPC), so fields that the SDK
|
|
* declares required may be absent at runtime if the server is buggy.
|
|
*/
|
|
interface McpContentBlock {
|
|
type: string
|
|
text?: string
|
|
mimeType?: string
|
|
}
|
|
|
|
/** Keep a supported advertised schema; unsupported MCP vocabulary falls back to JsonValue. */
|
|
function supportedOutputSchema(candidate: unknown): JsonSchemaNode | undefined {
|
|
if (candidate === undefined) return undefined
|
|
try {
|
|
assertSupportedJsonSchema(candidate)
|
|
return candidate
|
|
} catch {
|
|
return undefined
|
|
}
|
|
}
|
|
|
|
/** Build the canonical result schema and existing Native text projection. */
|
|
function createOutput(rawName: string, structuredSchema: JsonSchemaNode | undefined): ToolDefinition['output'] {
|
|
return {
|
|
schema: {
|
|
type: 'object',
|
|
properties: {
|
|
content: { type: 'array', items: {} },
|
|
structuredContent: structuredSchema ?? {},
|
|
},
|
|
required: structuredSchema === undefined ? ['content'] : ['content', 'structuredContent'],
|
|
additionalProperties: false,
|
|
},
|
|
render(_args, value) {
|
|
const result = value as unknown as McpResult
|
|
return [{ type: 'text', text: extractText(result.content, rawName) }]
|
|
},
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Create an execute function for one MCP tool. The executor closes over the
|
|
* raw MCP tool name and sends an uncached `tools/call` request with it (never
|
|
* the public name), with abort signal and timeout, then maps the result to
|
|
* harness ContentBlocks. Owning the raw request prevents the SDK's internal
|
|
* per-page schema cache from pre-validating a different contract.
|
|
*
|
|
* When the MCP server returns `isError: true`, the executor throws so that
|
|
* the ToolRegistry's catch path produces an `isError` result for the model.
|
|
*/
|
|
function createExecutor(
|
|
client: Client,
|
|
rawName: string,
|
|
taskRequired: boolean,
|
|
opts: ToolBridgeOptions,
|
|
): ToolDefinition['execute'] {
|
|
return async (args: unknown, exec: ToolExecution) => {
|
|
if (taskRequired) {
|
|
throw new Error(`Tool "${rawName}" requires task-based execution, which this bridge does not support`)
|
|
}
|
|
// The agent loop passes `JSON.parse(model_arguments)` which is usually an
|
|
// object, but can be any JSON value if the model misbehaves (outputs a bare
|
|
// string/number/null). Fallback to {} lets the MCP server produce a
|
|
// specific "missing required param" error the model can learn from.
|
|
const argsObj = (typeof args === 'object' && args !== null ? args : {}) as Record<string, unknown>
|
|
const result = await callToolUncached(client, rawName, argsObj, exec, opts)
|
|
|
|
// The SDK may return a legacy `toolResult` shape; normalize to content array.
|
|
if (!Array.isArray(result.content)) {
|
|
const rendered: unknown = 'toolResult' in result
|
|
? JSON.stringify(result.toolResult)
|
|
: '(no output)'
|
|
const text = typeof rendered === 'string' ? rendered : '(no output)'
|
|
if (result.isError === true) throw new Error(text)
|
|
return {
|
|
content: [{ type: 'text', text }],
|
|
...result.structuredContent !== undefined
|
|
? { structuredContent: result.structuredContent as JsonValue }
|
|
: {},
|
|
}
|
|
}
|
|
|
|
// Trust boundary: the SDK's return type erases to `any[]` due to the
|
|
// union of CallToolResult | CompatibilityCallToolResult. We process each
|
|
// element defensively in extractText (reading only .type/.text/.mimeType
|
|
// with optional fallbacks).
|
|
const content = result.content as unknown as JsonValue[]
|
|
const text = extractText(content, rawName)
|
|
|
|
// MCP isError → throw so ToolRegistry produces an isError result for the model.
|
|
if (result.isError === true) {
|
|
throw new Error(text)
|
|
}
|
|
|
|
return {
|
|
content,
|
|
...result.structuredContent !== undefined
|
|
? { structuredContent: result.structuredContent as JsonValue }
|
|
: {},
|
|
}
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Extract text from an MCP content array into a single string.
|
|
* - text blocks: join with '\n'
|
|
* - image/audio/resource blocks: replaced with a placeholder
|
|
*
|
|
* Defensive: fields that the MCP spec declares required (mimeType, text) are
|
|
* guarded with fallbacks because this is a network trust boundary.
|
|
*/
|
|
function extractText(mcpContent: JsonValue[], toolName: string): string {
|
|
const parts: string[] = []
|
|
|
|
for (const value of mcpContent) {
|
|
if (typeof value !== 'object' || value === null || Array.isArray(value)) {
|
|
parts.push('[unsupported content type: unknown]')
|
|
continue
|
|
}
|
|
const block = value as unknown as McpContentBlock
|
|
switch (block.type) {
|
|
case 'text':
|
|
if (block.text !== undefined) parts.push(block.text)
|
|
break
|
|
case 'image':
|
|
parts.push(`[image: ${block.mimeType ?? 'unknown'}, content discarded]`)
|
|
break
|
|
case 'audio':
|
|
parts.push(`[audio: ${block.mimeType ?? 'unknown'}, content discarded]`)
|
|
break
|
|
case 'resource':
|
|
case 'resource_link':
|
|
parts.push('[resource: content discarded]')
|
|
break
|
|
default:
|
|
parts.push(`[unsupported content type: ${block.type}]`)
|
|
}
|
|
}
|
|
|
|
return parts.join('\n') || `(${toolName} returned no text content)`
|
|
}
|