scanruler-mcp 0.1.0 → 0.3.0

package/src/mcp.js

@@ -1,14 +1,22 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // The MCP side: the tools an agent lists and calls. Each of the page's
-// commands is a tool (tools.js); two more are this server's own —
+// commands is a tool (tools.js); four more are this server's own —
 // scanruler_status, to see whether a ScanRuler page is connected and how to
 // connect one; scanruler_open, to open one in the user's browser, paired;
-// and scanruler_wait, to wait for it or for it to be idle.
+// scanruler_wait, to wait for it or for it to be idle; and scanruler_batch,
+// to run several of the page's tools in one call. In compact mode, for a
+// model with a small context window, the page's tools are not listed but
+// for those that return a picture, which a batch gives only as text: a
+// fifth of the server's own, scanruler_tools, lists them a line each and
+// gives the definitions of those asked for, and scanruler_batch runs them.
+// Headless (engine.js), the tools are the commands ScanRuler runs in this
+// process with no page, listed in either mode.
 
 import { readFileSync } from 'node:fs'
 import { Server } from '@modelcontextprotocol/sdk/server/index.js'
 import { CallToolRequestSchema, ListToolsRequestSchema } from '@modelcontextprotocol/sdk/types.js'
-import { imageContent, inputFor, saveFileResult, toolFor, toolName } from './tools.js'
+import { agentJson, imageContent, inputFor, saveFileResult, toolFor, toolName, waitIn } from './tools.js'
+import { jsonBytes, replyMeasure } from './trace.js'
 
 /** The app's own commands as of this release, so the tools are there before
  *  a page connects — unless a page has connected before, whose list was
@@ -17,16 +25,79 @@ import { imageContent, inputFor, saveFileResult, toolFor, toolName } from './too
  *  (tests/mcpCommands.test.ts). */
 export const SNAPSHOT = JSON.parse(readFileSync(new URL('./commands.json', import.meta.url), 'utf8'))
 
-export const INSTRUCTIONS = [
+/** How the page's commands are listed: `full`, each its own tool, or
+ *  `compact`, read through scanruler_tools. */
+export const TOOL_MODES = ['full', 'compact']
+
+/** The most definitions scanruler_tools gives in one answer: twenty are
+ *  some ten thousand tokens, a small window's fair share. */
+const MAX_DEFINITIONS = 20
+
+const ABOUT = [
   'These tools drive ScanRuler, a measuring app for 3D scans, in the browser tab the user has open: the user watches every step in the viewport and can undo it there.',
   'Start with scanruler_status. If no ScanRuler page is connected, call scanruler_open: it opens ScanRuler in the user’s browser, paired with this server, and waits for it. If that leaves it unconnected, tell the user what howToConnect says, then call scanruler_wait.',
-  'Read session_state before acting and after: the scan’s bounding box says where the part lies, the elements and dimensions what is measured. Lengths are millimetres, angles degrees.',
+]
+const WORKING = [
+  'Read session_state before acting and after: the scan’s bounding box says where the part lies, the elements and dimensions what is measured. Lengths are millimetres, angles degrees; answers give numbers to six significant digits.',
   'A place on the scan is { point: [x, y, z] } (the nearest scan vertex is taken) or { vertex: n }. To fit a feature, aim at its middle, not at an edge.',
   'Every tool that changes the session is one step of the undo history; history_undo takes the last back. report_get gives the report as JSON (format "json") or as the text the panel copies (format "text").',
-  'view_render shows you the 3D view as a picture; view_set turns it to a standard view or fits it.',
-].join(' ')
+]
+const INSTRUCTIONS = {
+  full: [
+    ...ABOUT,
+    ...WORKING,
+    'view_render shows you the 3D view as a picture; view_set turns it to a standard view or fits it.',
+    'scanruler_batch runs several tools in one call, in order — many scan_nearest or element_fit calls, say — and stops at the first that is refused unless told to go on.',
+  ].join(' '),
+  compact: [
+    ...ABOUT,
+    'ScanRuler’s own tools are not listed one by one, but for those that return a picture. scanruler_tools lists them all, a name and a line each; scanruler_tools with names gives those tools’ full definitions. Read a tool’s definition before its first use, and give it the arguments its inputSchema asks for.',
+    'Run the tools through scanruler_batch — { "calls": [{ "tool": "session_state", "arguments": {} }] }; a single tool is a batch of one, and several run in order. Each result is { tool, ok, result } or { tool, ok: false, error: { code, message } }: the code is the page’s own (invalid_input, no_scan, busy, invalid_state, not_found …) and the message says what was wrong, so retry a refused call with its arguments corrected.',
+    ...WORKING,
+    'The tools that return a picture — view_render, which pictures the 3D view, among them — are listed as tools of their own: call them directly to see the picture. In a batch a picture comes back as text only, and is written to a file when its call gives path. view_set turns the view.',
+  ].join(' '),
+}
 
-const text = (value) => ({ content: [{ type: 'text', text: typeof value === 'string' ? value : JSON.stringify(value) }] })
+/** A command in a line, for scanruler_tools' list: its title, or the first
+ *  sentence of its description. */
+function summary(command) {
+  if (command.title) return command.title
+  const description = String(command.description ?? '').replace(/\s+/g, ' ').trim()
+  return /^.*?[.!?](?= |$)/.exec(description)?.[0] ?? description
+}
+
+/** What the server says it is when it runs ScanRuler itself, headless. */
+const HEADLESS_ABOUT = [
+  'These tools drive ScanRuler headless: the app runs inside this server, with no browser, no scan and no viewport — nobody watches it, and nothing is drawn. scanruler_status says so (backend: "headless").',
+]
+const HEADLESS_WORKING = [
+  'Lengths are millimetres, angles degrees; answers give numbers to six significant digits. Every tool that changes the session is one step of the undo history; history_undo takes the last back. Read session_state before acting and after.',
+]
+const HEADLESS_INSTRUCTIONS = {
+  full: [
+    ...HEADLESS_ABOUT,
+    'The tools listed are those that run so: modelling solids feature by feature with parameters, describing and checking the bodies, notes, and STEP and STL files read and written on this computer. A tool that needs the page — a scan, a picture, the browser — is not listed, and answers unavailable if called; it needs this server started in page mode, paired with a ScanRuler tab.',
+    ...HEADLESS_WORKING,
+    'scanruler_batch runs several tools in one call, in order, and stops at the first that is refused unless told to go on.',
+  ].join(' '),
+  compact: [
+    ...HEADLESS_ABOUT,
+    'ScanRuler’s tools are those that run so — modelling solids feature by feature with parameters, describing and checking the bodies, notes, and STEP and STL files read and written on this computer — and are not listed one by one. scanruler_tools lists them, a name and a line each; scanruler_tools with names gives those tools’ full definitions. Read a tool’s definition before its first use, and give it the arguments its inputSchema asks for. A tool that needs the page — a scan, a picture, the browser — is not among them, and answers unavailable if called; it needs this server started in page mode, paired with a ScanRuler tab.',
+    'Run the tools through scanruler_batch — { "calls": [{ "tool": "session_state", "arguments": {} }] }; a single tool is a batch of one, and several run in order. Each result is { tool, ok, result } or { tool, ok: false, error: { code, message } }: the message says what was wrong, so retry a refused call with its arguments corrected.',
+    ...HEADLESS_WORKING,
+  ].join(' '),
+}
+
+/** How many jobs the server keeps answers of, the oldest finished let go
+ *  first. */
+const JOBS_KEPT = 50
+
+/** A change answered directly that took longer than this has likely
+ *  outlived its client's wait (Claude Desktop's is a minute): its answer
+ *  counts as not yet seen, for scanruler_wait { for: "idle" } to give. */
+const SEEN_WITHIN_MS = 55_000
+
+const text = (value, exact = false) => ({ content: [{ type: 'text', text: typeof value === 'string' ? value : agentJson(value, exact) }] })
 const failure = (e) => {
   const code = e?.code ?? 'failed'
   const message = e?.message ?? String(e)
@@ -46,7 +117,11 @@ const failure = (e) => {
  *   openPage?: () => Promise<boolean>,
  *   openGraceMs?: number,
  *   openWaitMs?: number,
- * }} options — `openPage` opens the pairing link in the user's browser
+ *   tools?: 'full' | 'compact',
+ *   trace?: ReturnType<typeof import('./trace.js').createTrace>,
+ * }} options — `tools` is how the page's commands are listed (TOOL_MODES).
+ *   `trace`, when given, gets a line per call (trace.js).
+ *   `openPage` opens the pairing link in the user's browser
  *   (browser.js), true when it could; scanruler_open first gives a tab that
  *   is already paired `openGraceMs` to connect by itself — the page tries
  *   again at most ten seconds apart — and then waits `openWaitMs` for the
@@ -65,10 +140,14 @@ export function createMcpServer({
   openPage = async () => false,
   openGraceMs = 11_000,
   openWaitMs = 40_000,
+  tools: toolMode = 'full',
+  trace = null,
 }) {
+  const headless = link.backend === 'headless'
+  const compact = toolMode === 'compact'
   const server = new Server(
     { name: 'scanruler', title: 'ScanRuler', version },
-    { capabilities: { tools: { listChanged: true } }, instructions: INSTRUCTIONS },
+    { capabilities: { tools: { listChanged: true } }, instructions: (headless ? HEADLESS_INSTRUCTIONS : INSTRUCTIONS)[compact ? 'compact' : 'full'] },
   )
 
   // The last connected page's commands, for listing before a page is back:
@@ -88,7 +167,70 @@ export function createMcpServer({
     }
   }
 
+  // ---- jobs ---------------------------------------------------------------
+  // Every call that changes the session is a job, its answer kept: one
+  // started with wait: false answers at once with the job, and one the
+  // client gave up waiting for keeps running and keeps its reply — the
+  // status lists both, and scanruler_wait { for: "job" } gives the answer.
+  const jobs = new Map()
+  let nextJob = 1
+  // `ctx` is the call's own, for the trace: the job is named on the call's
+  // line, and one answered at once (queued) gets a line when it is done.
+  const startJob = (tool, run, ctx = null, queued = false) => {
+    const job = { job: `j${nextJob++}`, tool, state: 'running', started: Date.now(), ms: null, reply: null }
+    if (ctx) Object.assign(ctx.fields, { job: job.job, ...(queued ? { queued: true } : {}) })
+    const finished = (reply) => {
+      job.state = reply.isError ? 'failed' : 'done'
+      job.ms = Date.now() - job.started
+      job.reply = reply
+      // Answered to the client: seen, unless queued or likely timed out.
+      job.told = !queued && job.ms < SEEN_WITHIN_MS
+      if (queued && trace) trace.write({ ev: 'job', job: job.job, call: ctx?.call ?? null, tool, t: job.started, ms: job.ms, ...replyMeasure(reply) })
+      return reply
+    }
+    job.done = run().then(finished, (e) => finished(failure(e)))
+    jobs.set(job.job, job)
+    // The newest kept, but never one still running.
+    for (const [id, j] of jobs) {
+      if (jobs.size <= JOBS_KEPT) break
+      if (j.state !== 'running') jobs.delete(id)
+    }
+    return job
+  }
+  const jobBrief = (j) => ({ job: j.job, tool: j.tool, state: j.state, ...(j.ms !== null ? { ms: j.ms } : { runningMs: Date.now() - j.started }) })
+  /** A job's answer as scanruler_wait gives it: the tool's own answer
+   *  under result, or its refusal under error. */
+  const jobAnswer = (j) => {
+    if (!j.reply) return jobBrief(j)
+    const body = j.reply.content.find((x) => x.type === 'text')?.text ?? ''
+    if (j.reply.isError) {
+      const m = /^([a-z_]+): ([\s\S]*)$/.exec(body)
+      return { ...jobBrief(j), error: m ? { code: m[1], message: m[2] } : { code: 'failed', message: body } }
+    }
+    let result
+    try {
+      result = JSON.parse(body)
+    } catch {
+      result = body
+    }
+    return { ...jobBrief(j), result }
+  }
+
+  /** Whether the tab is hidden now — null with no page, or headless. */
+  const hiddenNow = () => (link.connected && link.page && !headless ? (link.page.hidden ?? false) : null)
+  const runningJobs = () => [...jobs.values()].filter((j) => j.state === 'running').length
+
+  /** A finished job as scanruler_wait { for: "idle" } lists it: a failed
+   *  one with its error, a batch with the call it stopped at. */
+  const finishedOut = (j) => {
+    const answer = jobAnswer(j)
+    if (j.state === 'failed') return { ...jobBrief(j), error: answer.error }
+    const failed = j.tool === 'scanruler_batch' ? answer.result?.failed : undefined
+    return { ...jobBrief(j), ...(failed ? { failed } : {}) }
+  }
+
   const status = () => ({
+    backend: headless ? 'headless' : 'page',
     connected: link.connected,
     page: link.page
       ? {
@@ -97,18 +239,21 @@ export function createMcpServer({
           origin: link.page.origin,
           plugins: link.page.plugins ?? [],
           viewport: link.page.viewport ?? null,
+          hidden: link.page.hidden ?? null,
           commands: link.page.commands.length,
         }
       : null,
     server: serverInfo(),
-    ...(link.connected ? {} : { howToConnect: connectHelp() }),
+    ...(link.connected || headless ? {} : { howToConnect: connectHelp() }),
+    ...(link.page?.hidden ? { hiddenNote: 'The ScanRuler tab is hidden — in the background, its window covered or minimised, or the screen locked. Changes still run; pictures and anything that waits on the screen are slower, and a browser that unloads background tabs (Chrome’s Memory Saver) may close it: add the site to "Always keep these sites active".' } : {}),
+    ...(jobs.size > 0 ? { jobs: [...jobs.values()].reverse().map(jobBrief) } : {}),
   })
 
   const own = [
     {
       name: 'scanruler_status',
       title: 'ScanRuler connection status',
-      description: 'Whether a ScanRuler page is connected to this agent, which version and plugins it has, and — when none is — how the user connects one.',
+      description: 'Whether a ScanRuler page is connected to this agent, which version and plugins it has, whether its tab is hidden (page.hidden), and — when none is — how the user connects one. backend says whether ScanRuler runs in a tab (page) or in this server (headless). jobs lists the latest calls that change the session, newest first, each running, done or failed — a call whose answer you did not get is among them; scanruler_wait { for: "job", job } gives its answer.',
       inputSchema: { type: 'object', properties: {}, additionalProperties: false },
       annotations: { readOnlyHint: true, openWorldHint: false },
       call: async () => text(status()),
@@ -120,6 +265,7 @@ export function createMcpServer({
       inputSchema: { type: 'object', properties: {}, additionalProperties: false },
       annotations: { readOnlyHint: false, openWorldHint: false },
       call: async () => {
+        if (headless) return text({ ...status(), opened: false, note: 'This server runs ScanRuler headless, in this process: there is no page to open.' })
         // A tab that was paired before is likely open and on its way back.
         if (link.connected || (await link.waitForPage(openGraceMs))) return text({ ...status(), opened: false })
         const opened = await openPage().catch(() => false)
@@ -130,25 +276,48 @@ export function createMcpServer({
     {
       name: 'scanruler_wait',
       title: 'Wait for ScanRuler',
-      description: 'Wait until a ScanRuler page is connected (for: "page", the default), or until it is also done with what it is busy with (for: "idle"). Returns the status; connected is false if the time ran out.',
+      description: 'Wait until a ScanRuler page is connected (for: "page", the default), or until it is also done with what it is busy with and every change still running has finished (for: "idle" — then finished lists the changes done or failed since the last such answer, a failed one with its error), or until a job is done (for: "job" with job — a call started with wait: false, or one whose answer did not come; scanruler_status lists them). Returns the status, or for a job its answer: { job, tool, state, ms, result } or { …, error: { code, message } } — state running if the time ran out first. connected is false if the time ran out waiting for a page.',
       inputSchema: {
         type: 'object',
         properties: {
-          for: { type: 'string', enum: ['page', 'idle'], description: 'page (the default) or idle.' },
+          for: { type: 'string', enum: ['page', 'idle', 'job'], description: 'page (the default), idle, or job.' },
+          job: { type: 'string', minLength: 1, description: 'The job to wait for, with for: "job" — as a call with wait: false or scanruler_status gave it.' },
           timeoutSeconds: { type: 'number', exclusiveMinimum: 0, maximum: 600, description: 'How long to wait at most, seconds — 120 by default.' },
         },
         additionalProperties: false,
       },
       annotations: { readOnlyHint: true, openWorldHint: false },
-      call: async (args) => {
+      call: async (args, ctx) => {
+        if (ctx) ctx.fields.wait = args?.for ?? (args?.job !== undefined ? 'job' : 'page')
         const until = Date.now() + 1000 * (args?.timeoutSeconds ?? 120)
+        if (args?.for === 'job' || args?.job !== undefined) {
+          const j = jobs.get(String(args?.job ?? ''))
+          if (!j) return failure({ code: 'not_found', message: `There is no job "${args?.job ?? ''}" — scanruler_status lists the latest.` })
+          let timer
+          await Promise.race([j.done, new Promise((r) => (timer = setTimeout(r, Math.max(0, until - Date.now()))))])
+          clearTimeout(timer)
+          if (j.state !== 'running') j.told = true
+          return text(jobAnswer(j))
+        }
         if (!(await link.waitForPage(Math.max(0, until - Date.now())))) return text(status())
         if (args?.for === 'idle') {
+          // The changes still running first — their answers are what idle
+          // waits for — then the page's own work.
+          for (const j of [...jobs.values()].filter((x) => x.state === 'running')) {
+            let timer
+            await Promise.race([j.done, new Promise((r) => (timer = setTimeout(r, Math.max(0, until - Date.now()))))])
+            clearTimeout(timer)
+          }
           while (Date.now() < until) {
             const state = await link.run('session.state', {}).catch(() => null)
             if (state && state.busy === null && state.running === null) break
             await new Promise((r) => setTimeout(r, 500))
           }
+          // What finished since it was last said: each job once, a failed
+          // one with its error.
+          const finished = [...jobs.values()].filter((j) => j.state !== 'running' && !j.told)
+          for (const j of finished) j.told = true
+          return text({ ...status(), finished: finished.map(finishedOut) })
         }
         return text(status())
       },
@@ -156,11 +325,17 @@ export function createMcpServer({
   ]
 
   let listed = ''
-  const listNow = () => [...own.map(({ call: _call, ...tool }) => tool), ...commands().map((c) => toolFor(c, { outDir }))]
+  const definition = ({ call: _call, ...tool }) => tool
+  // Compact, the list is the server's own tools and the page's that return
+  // a picture — a batch gives a picture as its text, so those are called
+  // directly; scanruler_tools reads commands() when it is called.
+  const shown = (c) => !compact || Boolean(c.returnsImage)
+  const listNow = () => [...own.map(definition), ...commands().filter(shown).map((c) => toolFor(c, { outDir }))]
 
   server.setRequestHandler(ListToolsRequestSchema, async () => {
     const tools = listNow()
     listed = JSON.stringify(tools)
+    trace?.write({ ev: 'list', t: trace.now(), tools: tools.length, bytes: Buffer.byteLength(listed, 'utf8') })
     return { tools }
   })
 
@@ -178,15 +353,242 @@ export function createMcpServer({
     })
   }
 
-  server.setRequestHandler(CallToolRequestSchema, async (request, extra) => {
-    const { name, arguments: args } = request.params
-    const mine = own.find((t) => t.name === name)
-    if (mine) return mine.call(args)
+  /** One of the page's tools run on the page: the result as MCP content,
+   *  or a failure. */
+  const runPageTool = async (name, args, onProgress, { timings = false, ms = false } = {}) => {
     const command = commands().find((c) => toolName(c) === name)
+    const needsPage = !command && headless ? link.pageCommand(name) : null
+    if (needsPage) return failure({ code: 'unavailable', message: `${name} needs the ScanRuler page — a scan, the viewport or the browser — and this server runs ScanRuler headless, with none. Start scanruler-mcp in page mode (without --engine or --headless) and pair it with a ScanRuler tab for it.` })
     if (!command) return failure({ code: 'unknown_command', message: `There is no tool ${name}.` })
     if (!link.connected && !(await link.waitForPage(pageTimeoutMs))) {
       return failure({ code: 'unavailable', message: `No ScanRuler page is connected; scanruler_open opens one. ${connectHelp()}` })
     }
+    try {
+      const { input, binaryKey } = await inputFor(command, args, { cwd })
+      const started = Date.now()
+      let result = await link.run(command.name, input, binaryKey, onProgress)
+      if (ms && result && typeof result === 'object' && !Array.isArray(result)) result = { ...result, ms: Date.now() - started }
+      const image = command.returnsImage ? await imageContent(result, args, { cwd, outDir }) : null
+      if (image) return { content: image }
+      // Which steps a change's build played, and how long each took, only
+      // for a batch asked for its timings.
+      if (!timings && result?.steps?.played) {
+        const { played: _played, ...steps } = result.steps
+        result = { ...result, steps }
+      }
+      return text(await saveFileResult(result, args, { cwd, outDir }), Boolean(command.exact))
+    } catch (e) {
+      return failure(e)
+    }
+  }
+
+  // Several of the page's tools in one call — what cuts a run of twenty
+  // scan_nearest calls to one round trip. Each answer is the tool's own; a
+  // picture comes back as the text beside it (where it was written, if a
+  // path was given), not as an image.
+  own.push({
+    name: 'scanruler_batch',
+    title: 'Run several ScanRuler tools',
+    description:
+      'Run several of the ScanRuler tools in one call, in order, each with its own arguments — a run of scan_nearest, element_fit or section_get calls, say. Returns results, one per call in order: { tool, ok, result } or { tool, ok: false, error: { code, message } }. The first refused stops the rest (stopped says which index) unless stopOnError is false — and then, with rollback (the default), every change the batch made before it is undone, so the batch stands or falls whole: the answer names failed: { index, tool, error } and undone (the steps taken back, newest first). Each call that changes the session is still its own undo step. timings: true gives each call its ms and the batch total_ms, and a change\'s build the steps it played with their ms (result.steps.played). wait: false answers at once with { queued, job }; scanruler_wait { for: "job", job } gives the batch\'s answer. A tool that returns a picture gives its text here, the picture written to a file only if its call gave path. The server’s own tools (scanruler_status, _open, _wait, _batch) cannot be batched.',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        calls: {
+          type: 'array',
+          minItems: 1,
+          maxItems: 200,
+          description: 'The calls, in order.',
+          items: {
+            type: 'object',
+            properties: {
+              tool: { type: 'string', minLength: 1, description: 'A tool’s name, as listed — scan_nearest, element_fit.' },
+              arguments: { type: 'object', description: 'Its arguments, as the tool takes them; {} for none.' },
+            },
+            required: ['tool'],
+            additionalProperties: false,
+          },
+        },
+        stopOnError: { type: 'boolean', description: 'Stop at the first refused call (the default), or run every call and report each.' },
+        rollback: { type: 'boolean', description: 'When the batch stops at a refused call, undo what it changed before it (the default) — or leave those changes standing (false).' },
+        timings: { type: 'boolean', description: 'Each call’s ms, the batch’s total_ms, and the steps a change’s build played with their ms.' },
+        wait: waitIn,
+      },
+      required: ['calls'],
+      additionalProperties: false,
+    },
+    annotations: { readOnlyHint: false, openWorldHint: false },
+    call: async (args, ctx) => {
+      const calls = Array.isArray(args?.calls) ? args.calls : []
+      if (calls.length === 0) return failure({ code: 'invalid_input', message: 'calls: give at least one call.' })
+      if (ctx) ctx.fields.calls = calls.length
+      const job = startJob('scanruler_batch', () => runBatch(calls, args, ctx), ctx, args?.wait === false)
+      return args?.wait === false ? text({ queued: true, job: job.job, tool: 'scanruler_batch' }) : job.done
+    },
+  })
+
+  /** The batch's calls run in order — see scanruler_batch. */
+  const runBatch = async (calls, args, ctx = null) => {
+    const stopOnError = args?.stopOnError !== false
+    const timings = args?.timings === true
+    const commandOf = (tool) => commands().find((x) => toolName(x) === tool)
+    // Where the history stood, to go back to when a call is refused — only
+    // when a call may change something, and a page that cannot mark its
+    // history simply has its batch left as it went.
+    let checkpoint = null
+    if (args?.rollback !== false && stopOnError && calls.some((c) => commandOf(String(c?.tool ?? '')) && !commandOf(String(c?.tool ?? '')).readOnly)) {
+      checkpoint = await link
+        .run('history.checkpoint', {})
+        .then((r) => r?.checkpoint ?? null)
+        .catch(() => null)
+    }
+    const results = []
+    let stopped
+    let failed
+    // A batch holding an exact answer (a plane's frame, say) keeps
+    // every answer exact: the rounding is one pass over the whole.
+    let exact = false
+    const started = Date.now()
+    for (let i = 0; i < calls.length; i++) {
+      const c = calls[i] ?? {}
+      const tool = String(c.tool ?? '')
+      const at = Date.now()
+      const { wait: _wait, timings: _timings, ...callArgs } = c.arguments ?? {}
+      const hiddenAt = hiddenNow()
+      const one = own.some((t) => t.name === tool)
+        ? failure({ code: 'invalid_input', message: `${tool} is the server’s own and is not batched — call it on its own.` })
+        : await runPageTool(tool, callArgs, undefined, { timings })
+      if (ctx && trace) {
+        trace.write({ ev: 'entry', call: ctx.call, index: i, t: at, tool, ms: Date.now() - at, in: jsonBytes(callArgs), ...replyMeasure(one), hidden: [hiddenAt, hiddenNow()], ...(trace.args ? { args: callArgs } : {}) })
+      }
+      const ms = timings ? { ms: Date.now() - at } : {}
+      if (!one.isError && commandOf(tool)?.exact) exact = true
+      const body = one.content.find((x) => x.type === 'text')?.text ?? ''
+      if (one.isError) {
+        const m = /^([a-z_]+): ([\s\S]*)$/.exec(body)
+        const error = m ? { code: m[1], message: m[2] } : { code: 'failed', message: body }
+        results.push({ tool, ok: false, error, ...ms })
+        if (stopOnError) {
+          stopped = i
+          failed = { index: i, tool, error }
+          break
+        }
+      } else {
+        let parsed
+        try {
+          parsed = JSON.parse(body)
+        } catch {
+          parsed = body
+        }
+        results.push({ tool, ok: true, result: parsed, ...ms })
+      }
+    }
+    // Refused: what the batch did before it taken back, newest first.
+    let undone
+    let rollbackError
+    if (failed && checkpoint) {
+      try {
+        undone = (await link.run('history.restore', { checkpoint }))?.undone ?? []
+      } catch (e) {
+        rollbackError = `${e?.code ?? 'failed'}: ${e?.message ?? String(e)}`
+      }
+    }
+    return text(
+      {
+        results,
+        ran: results.length,
+        of: calls.length,
+        ...(stopped !== undefined ? { stopped } : {}),
+        ...(failed ? { failed, ...(checkpoint ? { undone: undone ?? [] } : {}), ...(rollbackError ? { rollbackError } : {}) } : {}),
+        ...(timings ? { total_ms: Date.now() - started } : {}),
+      },
+      exact,
+    )
+  }
+
+  // The page's tools for a client that does not list them: a line each,
+  // or the definitions of some — each as the full list would give it.
+  if (compact) {
+    own.push({
+      name: 'scanruler_tools',
+      title: 'List the ScanRuler tools',
+      description: `The ScanRuler tools, which this server lists one by one only when they return a picture. With no arguments: every tool’s name and what it does, a line each. With names: those tools’ full definitions — description and inputSchema — up to ${MAX_DEFINITIONS} at a time; read a tool’s before its first use. Run them with scanruler_batch, a single tool as a batch of one. Names it does not know come back in unknown. The list is the connected page’s, so it can change when a page connects.`,
+      inputSchema: {
+        type: 'object',
+        properties: {
+          names: {
+            type: 'array',
+            items: { type: 'string', minLength: 1 },
+            maxItems: MAX_DEFINITIONS,
+            description: 'Tool names, as the list gives them — element_fit, scan_nearest. Leave out for the list.',
+          },
+        },
+        additionalProperties: false,
+      },
+      annotations: { readOnlyHint: true, openWorldHint: false },
+      call: async (args) => {
+        const asked = typeof args?.names === 'string' ? args.names.split(/[\s,]+/) : Array.isArray(args?.names) ? args.names.map(String) : []
+        const names = [...new Set(asked.filter(Boolean))]
+        const list = commands()
+        if (names.length === 0) {
+          return text({
+            count: list.length,
+            tools: Object.fromEntries(list.map((c) => [toolName(c), summary(c)])),
+            next: 'scanruler_tools with names gives a tool’s definition; scanruler_batch runs it.',
+          })
+        }
+        if (names.length > MAX_DEFINITIONS) {
+          return failure({ code: 'invalid_input', message: `names: at most ${MAX_DEFINITIONS} at a time — ask for the rest in another call.` })
+        }
+        const found = []
+        const unknown = []
+        for (const name of names) {
+          const command = list.find((c) => toolName(c) === name)
+          const mine = own.find((t) => t.name === name)
+          if (command) found.push(toolFor(command, { outDir }))
+          else if (mine) found.push(definition(mine))
+          else unknown.push(name)
+        }
+        // Exact: a schema's numbers are limits and defaults, not readings.
+        return text({ tools: found, unknown }, true)
+      },
+    })
+  }
+
+  // Each call traced as one line when it is answered — what the client
+  // waited for — with the tab's hidden state at its start and end and the
+  // jobs running when it came; a batch's entries and a queued job's end get
+  // lines of their own.
+  let callNo = 0
+  server.setRequestHandler(CallToolRequestSchema, async (request, extra) => {
+    if (!trace) return dispatch(request, extra, null)
+    const { name, arguments: args } = request.params
+    const ctx = { call: ++callNo, fields: {} }
+    const t = trace.now()
+    const hiddenAt = hiddenNow()
+    const running = runningJobs()
+    const reply = await dispatch(request, extra, ctx)
+    trace.write({
+      ev: 'call',
+      call: ctx.call,
+      t,
+      tool: name,
+      ms: trace.now() - t,
+      in: jsonBytes(args ?? {}),
+      ...replyMeasure(reply),
+      hidden: [hiddenAt, hiddenNow()],
+      ...(running ? { running } : {}),
+      ...ctx.fields,
+      ...(trace.args ? { args: args ?? {} } : {}),
+    })
+    return reply
+  })
+
+  /** A tool call answered — see the handler above. */
+  const dispatch = async (request, extra, ctx) => {
+    const { name, arguments: args } = request.params
+    const mine = own.find((t) => t.name === name)
+    if (mine) return mine.call(args, ctx)
     // Progress lines from the page, for a client that asked for them.
     const token = request.params._meta?.progressToken
     let step = 0
@@ -196,16 +598,18 @@ export function createMcpServer({
         : (line) => {
             extra.sendNotification({ method: 'notifications/progress', params: { progressToken: token, progress: ++step, message: line } }).catch(() => {})
           }
-    try {
-      const { input, binaryKey } = await inputFor(command, args, { cwd })
-      const result = await link.run(command.name, input, binaryKey, onProgress)
-      const image = command.returnsImage ? await imageContent(result, args, { cwd, outDir }) : null
-      if (image) return { content: image }
-      return text(await saveFileResult(result, args, { cwd, outDir }))
-    } catch (e) {
-      return failure(e)
-    }
-  })
+    // A change runs as a job: answered when done, or at once with the job
+    // for wait: false — and kept either way, for a client that stops
+    // waiting before it is done.
+    const command = commands().find((c) => toolName(c) === name)
+    if (!command || command.readOnly) return runPageTool(name, args, onProgress)
+    const { wait, timings, ...rest } = args ?? {}
+    const job = startJob(name, () => runPageTool(name, rest, onProgress, { timings: timings === true, ms: timings === true }), ctx, wait === false)
+    return wait === false ? text({ queued: true, job: job.job, tool: name }) : job.done
+  }
+
+  trace?.write({ ev: 'start', t: trace.now(), version, backend: headless ? 'headless' : 'page', tools: toolMode, pid: process.pid, args: trace.args })
+  link.on?.('visibility', (hidden) => trace?.write({ ev: 'visibility', t: trace.now(), hidden }))
 
   return server
 }

package/src/relay.js

@@ -80,6 +80,7 @@ export class PeerLink extends Asker {
     this.page = page
     if (page && !was) this.emit('connected', page)
     else if (!page && was) this.emit('disconnected')
+    else if (page && (page.hidden ?? false) !== (was.hidden ?? false)) this.emit('visibility', page.hidden === true)
     else if (page) this.emit('commands', page.commands)
   }
 
@@ -159,7 +160,7 @@ export class SharedLink extends EventEmitter {
 
   /** Pass on a link's page events while `active()` says it is the one in use. */
   #forward(link, active) {
-    for (const event of ['connected', 'disconnected', 'commands']) {
+    for (const event of ['connected', 'disconnected', 'commands', 'visibility']) {
       link.on(event, (...args) => {
         if (active()) this.emit(event, ...args)
       })

package/src/serve.js

@@ -31,7 +31,7 @@ const TYPES = {
 }
 
 /** The file a URL path names under `root`, or null for one outside it. */
-export function fileFor(root, urlPath) {
+function fileFor(root, urlPath) {
   const path = decodeURIComponent(urlPath.split('?')[0].split('#')[0])
   const file = resolve(root, normalize('.' + (path.endsWith('/') ? `${path}index.html` : path)))
   return file === root || file.startsWith(root + sep) ? file : null