@flowriki/agent 0.1.0-beta.57 → 0.1.0-beta.59

package/dist/{chunk-JFMQJDBI.js → chunk-HXUFFE2E.js}

@@ -68,7 +68,7 @@ import {
   textIndependent,
   validateActors,
   webCloudOf
-} from "./chunk-O34WXNH6.js";
+} from "./chunk-MJXAVVGP.js";
 
 // src/engine/machine.ts
 import fs from "fs";
@@ -76,7 +76,7 @@ import os from "os";
 import path from "path";
 import { createRequire } from "module";
 import { spawnSync } from "child_process";
-var AGENT_VERSION = "0.1.0-beta.57";
+var AGENT_VERSION = "0.1.0-beta.59";
 async function prepareBrowsers() {
   if (browsersReady()) {
     return { ok: true, already: true, message: "chromium is already installed" };
@@ -136,7 +136,7 @@ async function checkMachine() {
     }
   } catch {
   }
-  const { mobileDoctor: mobileDoctor2 } = await import("./dist-RDJIIBQV.js");
+  const { mobileDoctor: mobileDoctor2 } = await import("./dist-BB5LTG3H.js");
   return {
     agentVersion: AGENT_VERSION,
     node: process.versions.node,
@@ -9806,6 +9806,8 @@ function assistantHandlers({ deps }) {
         ...linked.length ? { folders: linked } : {},
         ...mode ? { mode } : {},
         ...brief ? { brief } : {},
+        ...command.args["platform"] === "web" || command.args["platform"] === "android" || command.args["platform"] === "ios" ? { platform: command.args["platform"] } : {},
+        ...command.args["reuse"] === false ? { reuse: false } : {},
         ...typeof sessionId === "string" ? { sessionId } : {},
         emit
       });
@@ -10175,7 +10177,8 @@ function buildAskArgs(opts) {
   const brief = opts.brief?.trim();
   const guidance = [
     `You are the Flowriki assistant for the project "${opts.projectName}" (id ${opts.projectId}). Use the Flowriki tools, and pass projectId "${opts.projectId}" wherever a tool takes one. ` + (brief ? "The project brief below is this project's memory for this job: read it first." : "Start with project_brief: it is this project's memory."),
-    ask ? "ASK MODE: answer the question; change nothing. Cite what you rely on as [[testCase:<its id, tc_\u2026>]], [[knowledge:features/checkout.md]] or [[file:path/in/repo.ts]]." : "DO MODE: do what is asked, on Claude's working branch, for a person to review. Ask in the chat and wait before publishing, deleting, or sending anything to another system. End by writing a short summary of what you did with write_knowledge (memory/sessions/<date>-<topic>.md, append) so the next session knows.",
+    ask ? "ASK MODE: answer the question; change nothing. Cite what you rely on as [[testCase:<its id, tc_\u2026>]], [[knowledge:features/checkout.md]] or [[file:path/in/repo.ts]]." : "DO MODE: do what is asked, on Claude's working branch, for a person to review. Ask in the chat and wait before publishing, deleting, or sending anything to another system. Work quietly: the person sees what you build and your final answer, so keep words between tool calls to a minimum and end with a short result \u2014 what was saved, and anything that needs them. When you saved something, write a short summary with write_knowledge (memory/sessions/<date>-<topic>.md, append) so the next session knows.",
+    "Use only the tools you have: a tool is used by calling it, not by writing the call out as text. If Flowriki's tools are missing, say so in one sentence and stop.",
     folders.length ? `The project's code is in ${folders.join(" and ")}. Read it with Read, Grep and Glob when it helps; you cannot change it.` : "",
     brief ? `
 ${brief}` : ""
@@ -10207,6 +10210,16 @@ ${brief}` : ""
     ...opts.sessionId ? ["--resume", opts.sessionId] : []
   ];
 }
+function clearOld(dir) {
+  try {
+    for (const name of fs43.readdirSync(dir)) {
+      if (!/^mcp-.+\.json$/.test(name)) continue;
+      const file2 = path42.join(dir, name);
+      if (Date.now() - fs43.statSync(file2).mtimeMs > 24 * 60 * 6e4) fs43.rmSync(file2, { force: true });
+    }
+  } catch {
+  }
+}
 function eventsFromLine(askId, line) {
   let parsed;
   try {
@@ -10215,8 +10228,10 @@ function eventsFromLine(askId, line) {
     return [];
   }
   const out = [];
-  if (parsed.type === "system" && parsed.subtype === "init" && parsed.session_id) {
-    out.push({ name: "assistant.session", data: { askId, sessionId: parsed.session_id } });
+  if (parsed.type === "system" && parsed.subtype === "init") {
+    const ours = (parsed.mcp_servers ?? []).find((s2) => s2.name === MCP_SERVER_NAME);
+    if (ours) out.push({ name: "assistant.tools", data: { askId, status: ours.status ?? "unknown" } });
+    if (parsed.session_id) out.push({ name: "assistant.session", data: { askId, sessionId: parsed.session_id } });
   }
   if (parsed.type === "assistant") {
     for (const block of parsed.message?.content ?? []) {
@@ -10299,6 +10314,35 @@ function createClaudeCli(options) {
       child.on("close", (code) => finish(code, false));
     });
   }
+  const LANE_FRESH_MS = 60 * 6e4;
+  const LANE_MAX_TOKENS = 15e4;
+  const laneFile = (dir) => path42.join(dir, "lane-sessions.json");
+  const readLanes = (dir) => {
+    try {
+      return JSON.parse(fs43.readFileSync(laneFile(dir), "utf8"));
+    } catch {
+      return {};
+    }
+  };
+  const writeLane = (dir, platform, record) => {
+    const all = readLanes(dir);
+    if (record) all[platform] = record;
+    else delete all[platform];
+    try {
+      fs43.writeFileSync(laneFile(dir), JSON.stringify(all), { mode: 384 });
+    } catch {
+    }
+  };
+  function laneSession(ask, dir) {
+    if (ask.sessionId || !ask.platform) return ask.sessionId;
+    if (ask.reuse === false) {
+      writeLane(dir, ask.platform, null);
+      return void 0;
+    }
+    const kept = readLanes(dir)[ask.platform];
+    if (!kept || Date.now() - kept.updatedAt > LANE_FRESH_MS || kept.tokens > LANE_MAX_TOKENS) return void 0;
+    return kept.sessionId;
+  }
   return {
     async status() {
       const version = await capture(["--version"]);
@@ -10344,13 +10388,14 @@ function createClaudeCli(options) {
       fs43.mkdirSync(scratchRoot2, { recursive: true, mode: 448 });
       const dir = path42.join(scratchRoot2, `project-${ask.projectId.replace(/[^A-Za-z0-9_-]/g, "_")}`);
       fs43.mkdirSync(dir, { recursive: true, mode: 448 });
-      const mcpConfigPath = path42.join(dir, "mcp.json");
+      clearOld(dir);
+      const mcpConfigPath = path42.join(dir, `mcp-${ask.askId.replace(/[^A-Za-z0-9_-]/g, "_").slice(0, 80)}.json`);
       fs43.writeFileSync(
         mcpConfigPath,
-        JSON.stringify({ mcpServers: { [MCP_SERVER_NAME]: { command: options.mcp.command, args: options.mcp.args } } }),
+        JSON.stringify({ mcpServers: { [MCP_SERVER_NAME]: { command: options.mcp.command, args: options.mcp.args, env: { FLOWRIKI_ASK_ID: ask.askId } } } }),
         { mode: 384 }
       );
-      return launch(ask, dir, mcpConfigPath, ask.sessionId);
+      return launch(ask, dir, mcpConfigPath, laneSession(ask, dir));
     },
     cancel(askId) {
       const child = running.get(askId);
@@ -10366,7 +10411,7 @@ function createClaudeCli(options) {
       return running.size >= MAX_AT_ONCE;
     }
   };
-  function launch(ask, dir, mcpConfigPath, sessionId) {
+  function launch(ask, dir, mcpConfigPath, sessionId, attempt = 1) {
     let child;
     try {
       child = spawn2(
@@ -10390,8 +10435,21 @@ function createClaudeCli(options) {
     running.set(ask.askId, child);
     let buffer = "", stderr = "", sawDone = false, ended = false;
     let silentFailure = false;
+    let toolsMissing = false;
+    let session;
+    let spent = 0;
     const handleLine = (line) => {
+      if (toolsMissing) return;
       for (const event of eventsFromLine(ask.askId, line)) {
+        if (event.name === "assistant.tools") {
+          if (event.data["status"] !== "connected") {
+            toolsMissing = true;
+            child.kill("SIGTERM");
+            return;
+          }
+          continue;
+        }
+        if (event.name === "assistant.session") session = String(event.data["sessionId"] ?? "") || session;
         if (event.name === "assistant.done" && event.data["ok"] === false && !String(event.data["result"] ?? "").trim()) {
           silentFailure = true;
           continue;
@@ -10399,6 +10457,14 @@ function createClaudeCli(options) {
         if (event.name === "assistant.done") sawDone = true;
         ask.emit(event.name, event.data);
       }
+      try {
+        const parsed = JSON.parse(line);
+        if (parsed.type === "result" && parsed.usage) {
+          const u = parsed.usage;
+          spent += (u.input_tokens ?? 0) + (u.output_tokens ?? 0) + (u.cache_read_input_tokens ?? 0) + (u.cache_creation_input_tokens ?? 0);
+        }
+      } catch {
+      }
     };
     const end = (ok, detail) => {
       if (ended) return;
@@ -10406,10 +10472,26 @@ function createClaudeCli(options) {
       clearTimeout(timer);
       if (buffer.trim()) handleLine(buffer);
       running.delete(ask.askId);
+      if (toolsMissing && cancelled !== ask.askId) {
+        if (attempt < 3) {
+          setTimeout(() => {
+            const again = launch(ask, dir, mcpConfigPath, sessionId, attempt + 1);
+            if (!again.started) ask.emit("assistant.done", { askId: ask.askId, ok: false, result: again.detail ?? "The assistant could not be started." });
+          }, 1500 * attempt + Math.floor(Math.random() * 1e3)).unref?.();
+          return;
+        }
+        ask.emit("assistant.done", { askId: ask.askId, ok: false, result: "Flowriki's tools did not connect to the assistant on this computer after three tries, so nothing was done. Send it again in a moment; if it keeps happening, restart the agent." });
+        return;
+      }
       if (silentFailure && sessionId && /no conversation found/i.test(stderr) && cancelled !== ask.askId) {
+        if (ask.platform) writeLane(dir, ask.platform, null);
         const again = launch(ask, dir, mcpConfigPath, void 0);
         if (again.started) return;
       }
+      if (ask.platform && session && ok && !silentFailure) {
+        const before = sessionId && readLanes(dir)[ask.platform]?.sessionId === sessionId ? readLanes(dir)[ask.platform].tokens : 0;
+        writeLane(dir, ask.platform, { sessionId: session, updatedAt: Date.now(), tokens: Math.max(before, spent) });
+      }
       if (!sawDone) {
         const why = detail || stderr.trim().split("\n").filter((l) => l && !/^Warning:/.test(l)).slice(-3).join(" ") || "The assistant stopped without saying why.";
         ask.emit("assistant.done", { askId: ask.askId, ok: ok && !silentFailure, result: ok && !silentFailure ? "" : why });

package/dist/{chunk-O34WXNH6.js → chunk-MJXAVVGP.js}

@@ -687,7 +687,13 @@ var ExecutionConfigSchema = z3.object({
 var BUSINESS_KEYWORDS = ["Given", "When", "Then", "And", "But"];
 var BusinessStepSchema = z22.object({
   keyword: z22.enum(BUSINESS_KEYWORDS),
-  text: z22.string().min(1).max(2e3)
+  text: z22.string().min(1).max(2e3),
+  /** THE CASE STEPS THAT CARRY IT OUT, once it is automated — each as
+   *  `<occurrenceId>/<stepId>`, so a person reads the case line by line with
+   *  the steps under the words they prove. Said by whoever automated it (the
+   *  assistant, when it builds the case). A step whose row is gone is simply
+   *  not found, and shows with the steps no line claims. */
+  steps: z22.array(z22.string().min(1).max(200)).max(500).optional()
 }).strict();
 var TestCaseOriginSchema = z22.object({
   kind: z22.enum(["import", "assistant"]),

package/dist/cli.js

@@ -24,7 +24,7 @@ import {
   runtimeValue,
   start,
   status
-} from "./chunk-JFMQJDBI.js";
+} from "./chunk-HXUFFE2E.js";
 import {
   CLOUD_PROVIDERS,
   CLOUD_PROVIDER_NAME,
@@ -51,7 +51,7 @@ import {
   slugify,
   stepRefusedOn,
   targetFor
-} from "./chunk-O34WXNH6.js";
+} from "./chunk-MJXAVVGP.js";
 
 // src/cli.ts
 import { parseArgs } from "util";
@@ -137,6 +137,7 @@ function serve(tools, info, io = {}) {
 }
 
 // src/mcp/service-client.ts
+var ASK_ID = (process.env["FLOWRIKI_ASK_ID"] ?? "").trim().slice(0, 200);
 var ServiceClient = class {
   constructor(connection, fetcher = fetch) {
     this.connection = connection;
@@ -159,7 +160,10 @@ var ServiceClient = class {
       headers: {
         "x-flowriki-device": this.connection.deviceId,
         "x-flowriki-device-secret": this.connection.secret,
-        "content-type": "application/json"
+        "content-type": "application/json",
+        // The dashboard's question these tools are answering, when they are:
+        // what they build is filed on it, for the person to watch.
+        ...ASK_ID ? { "x-flowriki-ask": ASK_ID } : {}
       }
     });
     const text2 = await res.text();
@@ -255,6 +259,12 @@ var ServiceClient = class {
     );
   }
   /** Build a test case from flows — as a dry run, or saved. */
+  /** The flow being recorded, as it stands, for the dashboard's job. Only
+   *  when these tools answer a question asked in the dashboard. */
+  async reportDraft(projectId, draft) {
+    if (!ASK_ID) return;
+    await this.call(`/api/agent/jobs/draft?projectId=${encodeURIComponent(projectId)}`, { method: "POST", body: JSON.stringify({ draft }) });
+  }
   buildCase(projectId, spec, dryRun) {
     return this.call(`/api/agent/cases?projectId=${encodeURIComponent(projectId)}`, { method: "POST", body: JSON.stringify({ ...spec, dryRun }) });
   }
@@ -497,11 +507,21 @@ function deviceFrom(said, platform, devices, mobile) {
 async function devicesSummary() {
   return listDevices().catch(() => []);
 }
-function describeRows(rows) {
-  return rows.map((r) => [
-    `Row ${r.row}: "${r.flow}" (${r.browser})`,
-    ...r.checks.length ? r.checks.map((c) => `  ${c.on ? "[on] " : "[off]"} ${c.n}. ${c.check}`) : ["  (no checks in this flow)"]
-  ].join("\n")).join("\n");
+function describeRows(rows, testCase) {
+  return rows.map((r, i) => {
+    const steps = testCase?.occurrences[i]?.steps;
+    if (r.steps?.length && steps?.length === r.steps.length) {
+      const on = new Map(r.checks.map((c) => [c.n, c.on]));
+      return [
+        `Row ${r.row}: "${r.flow}" (${r.browser})`,
+        ...r.steps.map((st, j) => `  step ${st.n}. ${describe(steps[j])}${st.check !== void 0 ? `  \u2014 check ${st.check} ${on.get(st.check) ? "[on]" : "[off]"}` : ""}`)
+      ].join("\n");
+    }
+    return [
+      `Row ${r.row}: "${r.flow}" (${r.browser})`,
+      ...r.checks.length ? r.checks.map((c) => `  ${c.on ? "[on] " : "[off]"} ${c.n}. ${c.check}`) : ["  (no checks in this flow)"]
+    ].join("\n");
+  }).join("\n");
 }
 function describe(step) {
   const target = (step.name ? `${step.role || "element"} "${step.name}"` : step.role || "") + (step.nth !== void 0 ? ` (nth ${step.nth})` : "");
@@ -634,6 +654,7 @@ var AuthoringSession = class _AuthoringSession {
     const left = opts.fresh ? null : this.readDraft(input.projectId, input.name);
     if (opts.fresh) this.dropDraft(input.projectId, input.name);
     const opened = isMobile(input.platform) ? await this.startMobile(input.platform, catalog) : await this.startWeb(input, catalog);
+    this.report();
     if (!left?.steps.length) return opened;
     return this.resume(left.steps, opened);
   }
@@ -686,6 +707,7 @@ ${view}`;
     const draft = this.draft;
     if (!draft) return;
     this.verified = null;
+    this.report();
     try {
       const file = _AuthoringSession.draftFile(draft.projectId, draft.name);
       fs.mkdirSync(path.dirname(file), { recursive: true, mode: 448 });
@@ -694,6 +716,36 @@ ${view}`;
     } catch {
     }
   }
+  /** The draft as it stands, to the dashboard question this answers — at
+   *  most every half second, a recording keeps a step a second. */
+  reportTimer = null;
+  report() {
+    if (this.reportTimer) return;
+    this.reportTimer = setTimeout(() => void this.sendReport(), 500);
+    this.reportTimer.unref?.();
+  }
+  /** Whatever is waiting to be reported, now — before a save or a close, so
+   *  the dashboard ends on the flow as it was saved. */
+  async sendReport() {
+    if (this.reportTimer) {
+      clearTimeout(this.reportTimer);
+      this.reportTimer = null;
+    }
+    {
+      const draft = this.draft;
+      if (!draft) return;
+      const featureName = draft.featureName ?? this.catalog?.features?.find((f) => f.id === draft.featureId)?.name ?? "";
+      if (typeof this.service.reportDraft !== "function") return;
+      await this.service.reportDraft(draft.projectId, {
+        name: draft.name,
+        feature: featureName,
+        platform: draft.platform ?? "web",
+        after: (draft.after ?? []).map((a) => a.name),
+        signsIn: Boolean(draft.signIn),
+        steps: withoutTyped(draft.steps).map((st) => st.secret ? { ...st, value: "" } : st)
+      }).catch(() => void 0);
+    }
+  }
   readDraft(projectId, name) {
     try {
       const raw = JSON.parse(fs.readFileSync(_AuthoringSession.draftFile(projectId, name), "utf8"));
@@ -1019,6 +1071,7 @@ Fix the step that failed (remove_step, then act again) and save again \u2014 or
       replayed = " It was replayed from the start first and passed.";
     }
     const checksOff = draft.steps.map((step) => step.type === "assert" ? { ...step, proposals: step.proposals.map((p) => ({ ...p, enabled: false })) } : step);
+    if (this.reportTimer) await this.sendReport();
     const checkCount = checksOff.filter((step) => step.type === "assert").length;
     const secretValues = {};
     for (const step of checksOff) {
@@ -1052,7 +1105,7 @@ Review and publish it here: ${review}${note}`;
   async previewCase(projectId, spec) {
     const built = await this.service.buildCase(projectId, spec, true);
     return `"${spec.name}" in ${spec.folder} would be:
-${describeRows(built.rows)}`;
+${describeRows(built.rows, built.testCase)}`;
   }
   /** RUN THE CASE, THEN SAVE IT. Built as a dry run, replayed in a fresh
    *  browser exactly as a test run does it, and saved onto Claude's working
@@ -1102,7 +1155,7 @@ ${describeRows(built.rows)}`;
       return `The case did not pass, so it was NOT saved.
 ${lines.join("\n")}
 
-${describeRows(built.rows)}`;
+${describeRows(built.rows, testCase)}`;
     }
     const saved = await this.service.buildCase(projectId, spec, false);
     const review = `${this.service.origin}${saved.path}`;
@@ -1111,6 +1164,7 @@ ${lines.join("\n")}
 Review and publish it here: ${review}`;
   }
   async close() {
+    if (this.reportTimer) await this.sendReport();
     const run = this.run;
     const mobile = this.mobile;
     this.run = null;
@@ -1445,10 +1499,18 @@ HOW TO WORK:
    catalog, then type {{its saveAs name}} into the field that needs it.
 5. Add checks for what each step proves (a heading, a message, a value, the
    URL). replay_flow, then save_flow. A flow must record at least one check.
-6. Build each test case: preview_case lists its rows and every check, numbered;
-   save_case with the numbers to switch on runs the whole case and saves it only
-   if it passed. Give the person the review link.
-7. Saving the same name again replaces it on the working branch. Old branches
+6. Build each test case: preview_case lists its rows and every step, numbered
+   across the case, with each check's number in its row; save_case with the
+   checks to switch on runs the whole case and saves it only if it passed.
+   Pass businessSteps to both: the case's Given / When / Then lines (its own,
+   from read_test_cases, when it has them) each with the step numbers that
+   carry it out, so the person reads every line with its steps under it.
+   Give the person the review link.
+7. Saving the same name again replaces it on the working branch. ONE VERSION
+   of each scenario per platform: redo a flow or case under its existing name
+   (project_catalog lists them) rather than saving another beside it. When the
+   project tests several platforms, name flows "<scenario> \u2014 Web", "\u2014 Android",
+   "\u2014 iOS". Old branches
    of yours? gather_branches puts them into one; discard_branch deletes one \u2014
    only after the person asks or agrees.
 Test data you invent (a new user's email or password for a sign-up) is fine to
@@ -1480,6 +1542,19 @@ var CASE_INPUT = {
         additionalProperties: false
       }
     },
+    businessSteps: {
+      type: "array",
+      description: 'The case in Given / When / Then lines, in order, each with the step numbers from preview_case that carry it out \u2014 e.g. {"step": "Given the user is on the sign-in page", "steps": [1, 2]}. A case that already has business steps keeps its own words: give the same lines.',
+      items: {
+        type: "object",
+        properties: {
+          step: { type: "string", description: "One line, starting Given, When, Then, And or But." },
+          steps: { type: "array", items: { type: "number" }, description: "The step numbers (across the case) that carry it out." }
+        },
+        required: ["step"],
+        additionalProperties: false
+      }
+    },
     branch: { type: "string", description: "Optional: save onto this branch (chg_\u2026) instead of the working one." },
     projectId: { type: "string" }
   },
@@ -1491,6 +1566,7 @@ var caseSpec = (args) => ({
   folder: String(args["folder"] ?? ""),
   ...args["description"] ? { description: String(args["description"]) } : {},
   rows: Array.isArray(args["rows"]) ? args["rows"] : [],
+  ...Array.isArray(args["businessSteps"]) ? { businessSteps: args["businessSteps"] } : {},
   ...args["branch"] ? { changeId: String(args["branch"]) } : {}
 });
 var text = (value, isError = false) => ({ content: [{ type: "text", text: value }], ...isError ? { isError } : {} });

package/dist/{dist-RDJIIBQV.js → dist-BB5LTG3H.js}

@@ -46,7 +46,7 @@ import {
   textOf,
   toolEnv,
   typesSecrets
-} from "./chunk-O34WXNH6.js";
+} from "./chunk-MJXAVVGP.js";
 export {
   AGENT_HOME,
   APPIUM_DIR,

package/dist/engine.js

@@ -9,8 +9,8 @@ import {
   resolveCheckouts,
   start,
   status
-} from "./chunk-JFMQJDBI.js";
-import "./chunk-O34WXNH6.js";
+} from "./chunk-HXUFFE2E.js";
+import "./chunk-MJXAVVGP.js";
 export {
   AGENT_VERSION,
   browsersReady,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@flowriki/agent",
-  "version": "0.1.0-beta.57",
+  "version": "0.1.0-beta.59",
   "description": "Run Flowriki UI tests on your own computer. Connects this machine to your Flowriki account.",
   "license": "SEE LICENSE IN LICENSE",
   "type": "module",