toolnexus 0.20.0 → 0.22.0

package/src/index.ts

@@ -14,5 +14,36 @@ export * from "./adapters.js"
 export * from "./translate.js"
 export * from "./toolkit.js"
 export * from "./client.js"
+export * from "./store.js"
 export * from "./classifier.js"
 export * as agents from "./agents/index.js"
+// Simple judgments (§8B, add-judge-adapters). `judge.noul/choice/score` are the NAMED builders
+// (the bare `noul/choice/score` above stay the §8B wire builders); the rest is also top-level.
+export * as judge from "./judge.js"
+// Judge batteries (§8B "Batteries", add-judge-batteries): eight `*Classifier` values + latestUserText.
+export * from "./batteries.js"
+export {
+  State,
+  context,
+  questionMap,
+  ask,
+  gate,
+  decide,
+  applyGate,
+  applyPolicy,
+  answersOf,
+  staticClassifier,
+  Tape,
+  DEFAULT_BANDS,
+  type NamedQuestion,
+  type Band,
+  type Bands,
+  type Answers,
+  type Rule,
+  type GateOutcome,
+  type Policy,
+} from "./judge.js"
+export * from "./promptcache.js"
+export * from "./tooltiers.js"
+export * from "./models.js"
+export * from "./builtin_extras.js"

package/src/judge.ts

@@ -0,0 +1,305 @@
+/**
+ * Simple judgments (SPEC.md §8B "Simple judgments — ask / gate", change `add-judge-adapters`).
+ * A thin layer over any `Classifier`: named question builders, a role-carrying `State`,
+ * `ask` (answers by name with bands) and `gate` / `Policy` (rules that escalate instead of
+ * deciding when unsure). The wire is unchanged — the builders produce exactly the §8B
+ * `evaluate(state, questions)` inputs.
+ *
+ * A gate never authorises; it only declines to decide.
+ */
+import {
+  Classifier,
+  Decision,
+  type DecisionAnswer,
+  type NoulCriteria,
+  type Question,
+  type RecordedDecision,
+  noul as noulQ,
+  choice as choiceQ,
+  score as scoreQ,
+} from "./classifier.js"
+import type { Request } from "./types.js"
+
+// ---------------------------------------------------------------- questions
+
+/** One named question of an ordered question list. */
+export interface NamedQuestion {
+  name: string
+  question: Question
+}
+
+/** `noul` question with a name. Name the state field it judges in `instructions`. */
+export function noul(name: string, instructions: string, criteria?: NoulCriteria): NamedQuestion {
+  return { name, question: noulQ(instructions, criteria) }
+}
+/** `choice` question with a name; `options` maps option -> what picking it MEANS. */
+export function choice(name: string, instructions: string, options: Record<string, string>): NamedQuestion {
+  return { name, question: choiceQ(instructions, options) }
+}
+/** `score` question with a name; `levels` is the ordered rubric. */
+export function score(name: string, instructions: string, levels: string[]): NamedQuestion {
+  return { name, question: scoreQ(instructions, levels) }
+}
+
+/** Ordered list -> the §8B questions map. A duplicate name is an error naming it. */
+export function questionMap(questions: readonly NamedQuestion[]): Record<string, Question> {
+  const out: Record<string, Question> = {}
+  for (const { name, question } of questions) {
+    if (Object.prototype.hasOwnProperty.call(out, name)) throw new Error(`duplicate question name ${JSON.stringify(name)}`)
+    out[name] = question
+  }
+  return out
+}
+
+// ---------------------------------------------------------------- state
+
+/** `role` next to the data's fields; a non-object value goes under `data`. */
+export function State(role: string, data: unknown): Record<string, unknown> {
+  if (data !== null && typeof data === "object" && !Array.isArray(data)) return { role, ...(data as object) }
+  return { role, data }
+}
+
+/** Sugar: `{context, message, ...extra}`. */
+export function context(ctx: string, message: string, extra: Record<string, unknown> = {}): Record<string, unknown> {
+  return { context: ctx, message, ...extra }
+}
+
+// ---------------------------------------------------------------- bands + answers
+
+export type Band = "yes" | "no" | "uncertain"
+export interface Bands {
+  low: number
+  high: number
+}
+/** Default cut-points; exclusive on the confident side. */
+export const DEFAULT_BANDS: Readonly<Bands> = Object.freeze({ low: 0.3, high: 0.7 })
+
+/** One answer from `ask`: the §8B answer plus `band` (noul) or `sure` (choice/score). */
+export type Answer = DecisionAnswer & {
+  /** noul only. */
+  band?: Band
+  /** choice / score only. */
+  sure?: boolean
+  /** The one number: noul probability, score value, or choice confidence. */
+  value(): number
+  /** The picked option of a choice answer ("" otherwise). JS keeps §8B's `choice` string FIELD on a
+   * choice answer (a method of the same name would shadow it), so this is `pick()`. */
+  pick(): string
+}
+export type Answers = Record<string, Answer>
+
+function noulBand(p: number, b: Bands): Band {
+  return p < b.low ? "no" : p > b.high ? "yes" : "uncertain"
+}
+
+function wrap(a: DecisionAnswer, b: Bands): Answer {
+  const extra =
+    a.type === "noul"
+      ? { band: noulBand(a.noul, b) }
+      : { sure: a.confidence > b.high && !(a.type === "choice" && a.nearUniform) }
+  const value = () => (a.type === "noul" ? a.noul : a.type === "score" ? a.score : a.confidence)
+  const pick = () => (a.type === "choice" ? a.choice : "")
+  const out = { ...a, ...extra } as Answer
+  Object.defineProperty(out, "value", { value, enumerable: false })
+  Object.defineProperty(out, "pick", { value: pick, enumerable: false })
+  return out
+}
+
+/** Band / sure every answer of a decision. */
+export function answersOf(decision: Decision, bands: Bands = DEFAULT_BANDS): Answers {
+  return Object.fromEntries(Object.entries(decision.answers).map(([k, a]) => [k, wrap(a, bands)]))
+}
+
+/** Evaluate once; answers by name, each with a band (noul) or `sure` (choice/score). */
+export async function ask(
+  classifier: Classifier,
+  state: unknown,
+  questions: readonly NamedQuestion[],
+  bands: Bands = DEFAULT_BANDS,
+): Promise<Answers> {
+  return answersOf(await classifier.evaluate(state, questionMap(questions)), bands)
+}
+
+// ---------------------------------------------------------------- gate / policy
+
+export interface Rule {
+  question: string
+  below?: number
+  at_least?: number
+  is?: string
+  action: string
+  target?: string
+}
+
+export interface GateOutcome {
+  action: string
+  target: string
+  escalated: boolean
+  /** Present iff escalated: a §10 `input` Request. */
+  request?: Request
+  answers: Answers
+}
+
+export interface Policy {
+  rules: Rule[]
+  /** The no-rule-fired action; empty escalates with reason "no rule fired". */
+  default?: string
+  bands?: Bands
+  /** Skip a rule whose answer is uncertain instead of escalating on it. */
+  skipUncertain?: boolean
+}
+
+function unsure(a: Answer): boolean {
+  return a.type === "noul" ? a.band === "uncertain" : !a.sure
+}
+
+function escalate(answers: Answers, id: string, question: string, reason: string, prompt: string): GateOutcome {
+  return {
+    action: "needs_input",
+    target: "",
+    escalated: true,
+    answers,
+    request: { id, kind: "input", prompt, data: { question, reason, answers } },
+  }
+}
+
+function applyRules(answers: Answers, rules: readonly Rule[], skipUncertain: boolean): GateOutcome | undefined {
+  for (const [i, r] of rules.entries()) {
+    const a = answers[r.question]
+    if (!a || unsure(a)) {
+      if (a && skipUncertain) continue
+      const reason = `${a ? "uncertain" : "missing"} answer "${r.question}"`
+      return escalate(
+        answers,
+        `gate:${i}:${r.question}`,
+        r.question,
+        reason,
+        `Classifier is unsure about "${r.question}" (${reason}). Decide rule ${i} (${r.action}).`,
+      )
+    }
+    const key = JSON.stringify(r.question)
+    let fired = false
+    let misfit = ""
+    if (r.is !== undefined) {
+      if (a.type === "choice") fired = a.choice === r.is
+      else misfit = `is-rule on ${a.type} answer ${key}`
+    } else if (a.type === "choice") {
+      misfit = `numeric rule on choice answer ${key}`
+    } else {
+      const v = a.type === "noul" ? a.noul : a.score
+      if (r.below !== undefined) fired = v < r.below
+      else if (r.at_least !== undefined) fired = v >= r.at_least
+      else misfit = `rule on ${key} has no below / at_least / is`
+    }
+    // A rule that does not fit its answer escalates; it is never skipped.
+    if (misfit)
+      return escalate(
+        answers,
+        `gate:${i}:${r.question}`,
+        r.question,
+        misfit,
+        `Rule ${i} (${r.action}) does not fit its answer (${misfit}). Decide.`,
+      )
+    if (fired) return { action: r.action, target: r.target ?? "", escalated: false, answers }
+  }
+  return undefined
+}
+
+/** Pure half of `gate`: first-match rules; an unsure or missing answer on rule i escalates and wins. */
+export function applyGate(answers: Answers, rules: readonly Rule[]): GateOutcome {
+  return applyRules(answers, rules, false) ?? { action: "", target: "", escalated: false, answers }
+}
+
+/** Pure half of a `Policy`. */
+export function applyPolicy(answers: Answers, policy: Policy): GateOutcome {
+  const hit = applyRules(answers, policy.rules, policy.skipUncertain ?? false)
+  if (hit) return hit
+  if (policy.default) return { action: policy.default, target: "", escalated: false, answers }
+  return escalate(answers, "gate:default", "", "no rule fired", "No rule fired. Decide the action.")
+}
+
+/** Ask, then apply rules; no rule fired -> empty action (use a `Policy` for a declared default). */
+export async function gate(
+  classifier: Classifier,
+  state: unknown,
+  questions: readonly NamedQuestion[],
+  rules: readonly Rule[],
+  bands: Bands = DEFAULT_BANDS,
+): Promise<GateOutcome> {
+  return applyGate(await ask(classifier, state, questions, bands), rules)
+}
+
+/** Ask, then apply a `Policy` (rules, default, bands, skipUncertain). */
+export async function decide(
+  classifier: Classifier,
+  state: unknown,
+  questions: readonly NamedQuestion[],
+  policy: Policy,
+): Promise<GateOutcome> {
+  return applyPolicy(await ask(classifier, state, questions, policy.bands ?? DEFAULT_BANDS), policy)
+}
+
+// ---------------------------------------------------------------- static + tape
+
+/** One-line `static` classifier from recorded decisions. */
+export function staticClassifier(decisions: RecordedDecision[] | RecordedDecision, model?: string): Classifier {
+  return new Classifier({ style: "static", decisions: Array.isArray(decisions) ? decisions : [decisions], model })
+}
+
+/** The recorded-body form of a Decision, decodable by `static`. */
+function toRaw(d: Decision): Record<string, unknown> {
+  const answers: Record<string, unknown> = {}
+  for (const [k, a] of Object.entries(d.answers)) {
+    if (a.type === "choice") {
+      const { nearUniform: _derived, ...rest } = a
+      answers[k] = rest
+    } else answers[k] = { ...a }
+  }
+  const usage: Record<string, number> = { input_tokens: d.usage.inputTokens, output_tokens: d.usage.outputTokens }
+  if (d.usage.cost !== undefined) usage.cost = d.usage.cost
+  return { model: d.model, answers, usage, calibrated: d.calibrated }
+}
+
+/**
+ * Records live decisions keyed by a caller-given call name, and replays them with no network.
+ * `tape.classifier(name)` returns a Classifier: recording (live given) or replaying (no live).
+ */
+export class Tape {
+  readonly entries: Record<string, RecordedDecision>
+
+  constructor(
+    private readonly live?: Classifier,
+    entries: Record<string, RecordedDecision> = {},
+  ) {
+    this.entries = { ...entries }
+  }
+
+  /** A replay-only tape from saved entries. */
+  static replay(entries: Record<string, RecordedDecision>): Tape {
+    return new Tape(undefined, entries)
+  }
+
+  classifier(name: string): Classifier {
+    const live = this.live
+    if (live) {
+      return new Classifier({
+        style: "custom",
+        evaluate: async (state, questions, signal) => {
+          const d = await live.evaluate(state, questions, signal)
+          this.entries[name] = { state, questions, response: toRaw(d) }
+          return d
+        },
+      })
+    }
+    // Replay: resolved when evaluated (a miss fails then, naming the key, and sends nothing).
+    return new Classifier({
+      style: "custom",
+      evaluate: async (state, questions, signal) => {
+        const rec = this.entries[name]
+        if (!rec) throw new Error(`tape: no recorded decision for call "${name}"`)
+        const raw = rec.response as { model?: string }
+        return staticClassifier(rec, raw?.model).evaluate(state, questions, signal)
+      },
+    })
+  }
+}

package/src/mcp.ts

@@ -10,7 +10,7 @@ import { StreamableHTTPClientTransport } from "@modelcontextprotocol/sdk/client/
 import { SSEClientTransport } from "@modelcontextprotocol/sdk/client/sse.js"
 import type { Tool as McpToolDef, ElicitRequest, ElicitResult } from "@modelcontextprotocol/sdk/types.js"
 import { ElicitRequestSchema } from "@modelcontextprotocol/sdk/types.js"
-import { type JSONSchema, type Tool, type ToolContext, type ToolResult, type McpStatus, type Request, type Answer, sanitize } from "./types.js"
+import { type JSONSchema, type Tool, type ToolContext, type ToolHints, type ToolResult, type McpStatus, type Request, type Answer, sanitize } from "./types.js"
 import { type ContentPart, summarizePart } from "./content.js"
 
 // ---------------------------------------------------------------------------
@@ -88,12 +88,38 @@ function isEnabled(cfg: ServerConfig): boolean {
 export function parseMcpConfig(input: string | object): McpConfig {
   const raw: any = typeof input === "string" ? JSON.parse(readFileSync(input, "utf8")) : input
   const wrapped = raw.mcpServers ?? raw.servers ?? raw.mcp
-  if (wrapped !== undefined) return wrapped as McpConfig
+  if (wrapped !== undefined) return normalizeServers(wrapped)
   // Bare map of servers — but strip sibling top-level config keys (builtins/
   // agents/a2a/mcpServer) so they are not mistaken for MCP servers when no
   // wrapper key is present.
   const { builtins: _builtins, agents: _agents, a2a: _a2a, mcpServer: _mcpServer, ...servers } = raw
-  return servers as McpConfig
+  return normalizeServers(servers)
+}
+
+/**
+ * §2 `command` + `args`. A local server's `command` may be a non-empty string or a non-empty
+ * string array; an optional `args` string array is appended. The result is always the argv form
+ * (`command: string[]`, no `args`). Spawned with no shell. A remote server ignores `args`.
+ */
+function normalizeServers(servers: Record<string, any>): McpConfig {
+  const out: McpConfig = {}
+  for (const [name, cfg] of Object.entries(servers ?? {})) {
+    if (cfg === null || typeof cfg !== "object" || isRemote(cfg as ServerConfig)) {
+      out[name] = cfg
+      continue
+    }
+    const { args, command, ...rest } = cfg as any
+    const head: unknown[] | undefined =
+      typeof command === "string" ? [command] : Array.isArray(command) ? command : undefined
+    if (!head || head.length === 0 || head[0] === "" || head.some((c) => typeof c !== "string")) {
+      throw new Error(`mcp server "${name}": invalid command`)
+    }
+    if (args !== undefined && (!Array.isArray(args) || args.some((a: unknown) => typeof a !== "string"))) {
+      throw new Error(`mcp server "${name}": invalid args`)
+    }
+    out[name] = { ...rest, command: [...(head as string[]), ...((args as string[] | undefined) ?? [])] }
+  }
+  return out
 }
 
 /**
@@ -240,12 +266,16 @@ function convertTool(server: string, def: McpToolDef, client: Client, timeout: n
     description: def.description ?? "",
     inputSchema,
     source: "mcp",
+    ...(mcpHints((def as any).annotations) ? { hints: mcpHints((def as any).annotations) } : {}),
     async execute(args: Record<string, unknown>, ctx?: ToolContext): Promise<ToolResult> {
       try {
         const result: any = await client.callTool(
           { name: def.name, arguments: args ?? {} },
           undefined,
-          { resetTimeoutOnProgress: true, signal: ctx?.signal, timeout: ctx?.timeout ?? timeout },
+          // The SDK only attaches `_meta.progressToken` when `onprogress` is supplied; without it
+          // `resetTimeoutOnProgress` never fires (add-mcp-progress-timeout). Notifications for a
+          // token that matches no in-flight request are dropped by the SDK, so they never reset.
+          { onprogress: () => {}, resetTimeoutOnProgress: true, signal: ctx?.signal, timeout: ctx?.timeout ?? timeout },
         )
         // Parts are collected BEFORE any branch returns: the isError and structuredContent
         // short-circuits both read `content[]` on the way past, so neither reintroduces the
@@ -271,6 +301,129 @@ export interface McpSource {
   tools: Tool[]
   status: Record<string, McpStatus>
   close(): Promise<void>
+  /**
+   * `list_mcp_resources` and `read_mcp_resource` bound to this source's connected servers
+   * (SPEC §4C). Both read-only. Opt-in: register them on a toolkit. They stop working once the
+   * source is closed.
+   */
+  resourceTools(): Tool[]
+}
+
+/** One server's outcome, reported by `loadMcp`'s `onServer` the moment that server finishes. */
+export interface McpServerResult {
+  name: string
+  status: McpStatus
+  /** This server's converted tools (empty unless connected). Usable until the source closes. */
+  tools: Tool[]
+  /** Why the server failed (absent when connected or disabled). */
+  error?: Error
+}
+
+/**
+ * MCP tool annotations → ToolHints (SPEC §4C). No hint fields at all ⇒ undefined (unknown).
+ * Defaults follow the MCP spec: readOnlyHint false, destructiveHint and openWorldHint true.
+ */
+export function mcpHints(a: { readOnlyHint?: boolean; destructiveHint?: boolean; openWorldHint?: boolean } | undefined): ToolHints | undefined {
+  if (!a || (a.readOnlyHint === undefined && a.destructiveHint === undefined && a.openWorldHint === undefined)) return undefined
+  const h: ToolHints = { readOnly: a.readOnlyHint ?? false, destructive: a.destructiveHint ?? true, openWorld: a.openWorldHint ?? true }
+  if (h.readOnly) h.destructive = false
+  return h
+}
+
+interface ServerConn {
+  client: Client
+  timeout: number
+}
+
+function resourceTools(servers: Map<string, ServerConn>): Tool[] {
+  const ro: ToolHints = { readOnly: true, destructive: false, openWorld: false }
+  const names = () => [...servers.keys()].sort()
+  const conn = (server: string): ServerConn => {
+    const c = servers.get(server)
+    if (!c) throw new Error(`no connected MCP server "${server}" (connected: ${names().join(", ")})`)
+    return c
+  }
+  const list: Tool = {
+    name: "list_mcp_resources",
+    description: "List the resources MCP servers offer (uri, name, description, mimeType). Optionally for one server.",
+    inputSchema: {
+      type: "object",
+      properties: { server: { type: "string", description: "Only this server (default: all)" } },
+      additionalProperties: false,
+    },
+    source: "mcp",
+    hints: ro,
+    async execute(args, ctx) {
+      let targets = names()
+      const s = typeof args?.server === "string" ? args.server : ""
+      if (s) {
+        try {
+          conn(s)
+        } catch (e) {
+          return { output: `list_mcp_resources: ${(e as Error).message}`, isError: true }
+        }
+        targets = [s]
+      }
+      const lines: string[] = []
+      const items: unknown[] = []
+      for (const name of targets) {
+        const c = conn(name)
+        try {
+          const res: any = await c.client.listResources({}, { timeout: c.timeout, signal: ctx?.signal })
+          for (const r of res.resources ?? []) {
+            let line = `${name} ${r.uri}`
+            if (r.name) line += ` — ${r.name}`
+            if (r.mimeType) line += ` (${r.mimeType})`
+            if (r.description) line += `: ${r.description}`
+            lines.push(line)
+            items.push({ server: name, uri: r.uri, name: r.name ?? "", mimeType: r.mimeType ?? "", description: r.description ?? "" })
+          }
+        } catch (e) {
+          lines.push(`${name}: error: ${e instanceof Error ? e.message : String(e)}`)
+        }
+      }
+      if (!lines.length) return { output: "No resources.", isError: false, metadata: { count: 0 } }
+      return { output: lines.join("\n"), isError: false, metadata: { count: items.length, resources: items } }
+    },
+  }
+  const read: Tool = {
+    name: "read_mcp_resource",
+    description: "Read one MCP resource by server and uri. Text comes back as text; binary content is described, not inlined.",
+    inputSchema: {
+      type: "object",
+      properties: {
+        server: { type: "string", description: "The MCP server name" },
+        uri: { type: "string", description: "The resource uri" },
+      },
+      required: ["server", "uri"],
+      additionalProperties: false,
+    },
+    source: "mcp",
+    hints: ro,
+    async execute(args, ctx) {
+      const server = typeof args?.server === "string" ? args.server : ""
+      const uri = typeof args?.uri === "string" ? args.uri : ""
+      let c: ServerConn
+      try {
+        c = conn(server)
+      } catch (e) {
+        return { output: `read_mcp_resource: ${(e as Error).message}`, isError: true }
+      }
+      if (!uri) return { output: "read_mcp_resource: uri is required", isError: true }
+      try {
+        const res: any = await c.client.readResource({ uri }, { timeout: c.timeout, signal: ctx?.signal })
+        const parts: string[] = []
+        for (const rc of res.contents ?? []) {
+          if (typeof rc.text === "string") parts.push(rc.text)
+          else if (typeof rc.blob === "string") parts.push(`[binary ${rc.mimeType ?? ""}, ${rc.blob.length} base64 chars]`)
+        }
+        return { output: parts.join("\n"), isError: false, metadata: { server, uri } }
+      } catch (e) {
+        return { output: `read_mcp_resource: ${e instanceof Error ? e.message : String(e)}`, isError: true }
+      }
+    },
+  }
+  return [list, read]
 }
 
 /** Create a fresh (not-yet-connected) client, with the elicitation bridge attached if a waitFor
@@ -318,19 +471,44 @@ async function connectTransport(client: Client, cfg: ServerConfig): Promise<void
  *  whole load promptly on parent cancellation. */
 export async function loadMcp(
   input: string | object,
-  opts?: { waitFor?: (request: Request) => Promise<Answer>; signal?: AbortSignal },
+  opts?: {
+    waitFor?: (request: Request) => Promise<Answer>
+    signal?: AbortSignal
+    /**
+     * Called once per server the moment it is connected, failed or found disabled (SPEC §4C),
+     * so a host can register a fast server's tools while a slow one is still connecting.
+     * Calls are serialized (an async callback is awaited before the next one runs); the load
+     * still resolves only after every server and every callback has finished.
+     */
+    onServer?: (result: McpServerResult) => void | Promise<void>
+  },
 ): Promise<McpSource> {
   const config = parseMcpConfig(input)
   const tools: Tool[] = []
   const status: Record<string, McpStatus> = {}
   const clients: Client[] = []
+  const servers = new Map<string, ServerConn>()
   const waitFor = opts?.waitFor
   const signal = opts?.signal
+  const onServer = opts?.onServer
+  let reports: Promise<void> = Promise.resolve()
+  let reportError: unknown
+  const report = (r: McpServerResult) => {
+    if (!onServer) return
+    reports = reports.then(async () => {
+      try {
+        await onServer(r)
+      } catch (e) {
+        reportError ??= e
+      }
+    })
+  }
 
-  await Promise.all(
+  const loading = Promise.all(
     Object.entries(config).map(async ([name, cfg]) => {
       if (!isEnabled(cfg)) {
         status[name] = "disabled"
+        report({ name, status: "disabled", tools: [] })
         return
       }
       const timeout = cfg.timeout ?? DEFAULT_TIMEOUT
@@ -341,19 +519,37 @@ export async function loadMcp(
         await raceTimeout(connectTransport(client, cfg), timeout, signal)
         const defs = await raceTimeout(paginateTools(client, timeout), timeout, signal)
         const kept = applyToolsFilter(name, defs, (cfg as { tools?: Record<string, boolean> }).tools)
-        for (const def of kept) tools.push(convertTool(name, def, client, timeout))
+        const converted = kept.map((def) => convertTool(name, def, client, timeout))
+        tools.push(...converted)
         clients.push(client)
+        servers.set(name, { client, timeout })
         status[name] = "connected"
+        report({ name, status: "connected", tools: converted })
       } catch (e) {
         status[name] = "failed"
         await client.close().catch(() => {})
         // Parent cancellation/deadline aborts the whole load (§2 Gap 3, A1); a per-server timeout
         // within budget isolates just this server.
         if (signal?.aborted) throw e
+        report({ name, status: "failed", tools: [], error: e instanceof Error ? e : new Error(String(e)) })
         console.warn(`[toolnexus] MCP server "${name}" failed: ${e instanceof Error ? e.message : String(e)}`)
       }
     }),
   )
+  try {
+    await loading
+  } catch (e) {
+    // Parent cancellation: close every client, including ones whose tools onServer handed out.
+    await reports
+    await Promise.all(clients.map((c) => c.close().catch(() => {})))
+    throw e
+  }
+
+  await reports
+  if (reportError) {
+    await Promise.all(clients.map((c) => c.close().catch(() => {})))
+    throw reportError
+  }
 
   return {
     tools,
@@ -361,6 +557,7 @@ export async function loadMcp(
     async close() {
       await Promise.all(clients.map((c) => c.close().catch(() => {})))
     },
+    resourceTools: () => resourceTools(servers),
   }
 }
 

package/src/models.ts

@@ -0,0 +1,56 @@
+/**
+ * listModels (SPEC §4C): the models a provider offers, for model switching. One GET per page,
+ * no retries. Same key rules as the client (apiKey, apiKeyEnv, or the host's own env var).
+ */
+import { resolveKey, type ClientOptions } from "./client.js"
+
+/** One model a provider lists. */
+export interface ModelInfo {
+  id: string
+  /** The provider's display name, when it gives one. */
+  name?: string
+  /** Context window in tokens when the provider reports it (OpenRouter does). */
+  contextLength?: number
+}
+
+async function getJson(f: typeof fetch, url: string, headers: Record<string, string>, signal?: AbortSignal): Promise<any> {
+  const res = await f(url, { method: "GET", headers, signal })
+  const text = await res.text()
+  if (!res.ok) throw new Error(`list models: HTTP ${res.status}: ${text.slice(0, 300)}`)
+  return JSON.parse(text)
+}
+
+/**
+ * List the models at `opts.baseUrl` in `opts.style`: openai `GET {base}/models`; anthropic
+ * `GET {base}/v1/models`, following `has_more`/`last_id`. Sorted by id.
+ */
+export async function listModels(
+  opts: Pick<ClientOptions, "baseUrl" | "style" | "apiKey" | "apiKeyEnv" | "headers" | "fetch">,
+  ctx?: { signal?: AbortSignal },
+): Promise<ModelInfo[]> {
+  const key = resolveKey(opts)
+  const f = opts.fetch ?? fetch
+  const base = opts.baseUrl.replace(/\/+$/, "")
+  const out: ModelInfo[] = []
+  if (opts.style === "anthropic") {
+    const endpoint = base.endsWith("/v1") ? `${base}/models` : `${base}/v1/models`
+    let after = ""
+    for (let page = 0; page < 100; page++) {
+      const url = `${endpoint}?limit=1000${after ? `&after_id=${encodeURIComponent(after)}` : ""}`
+      const body = await getJson(f, url, { "x-api-key": key, "anthropic-version": "2023-06-01", ...opts.headers }, ctx?.signal)
+      for (const m of body?.data ?? []) out.push({ id: m.id, ...(m.display_name ? { name: m.display_name } : {}) })
+      if (!body?.has_more || !body?.last_id) break
+      after = body.last_id
+    }
+  } else {
+    const body = await getJson(f, `${base}/models`, { Authorization: `Bearer ${key}`, ...opts.headers }, ctx?.signal)
+    for (const m of body?.data ?? []) {
+      out.push({
+        id: m.id,
+        ...(m.name ? { name: m.name } : {}),
+        ...(typeof m.context_length === "number" && m.context_length > 0 ? { contextLength: m.context_length } : {}),
+      })
+    }
+  }
+  return out.sort((a, b) => (a.id < b.id ? -1 : a.id > b.id ? 1 : 0))
+}