@flowriki/agent 0.1.0-beta.60 → 0.1.0-beta.62

package/dist/{chunk-BYFYSV5H.js → chunk-GPSLXDGQ.js}

@@ -68,7 +68,7 @@ import {
   textIndependent,
   validateActors,
   webCloudOf
-} from "./chunk-WQ22OUIQ.js";
+} from "./chunk-YCFTNJQU.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.60";
+var AGENT_VERSION = "0.1.0-beta.62";
 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-IBNVN2BD.js");
+  const { mobileDoctor: mobileDoctor2 } = await import("./dist-EGC7VEBG.js");
   return {
     agentVersion: AGENT_VERSION,
     node: process.versions.node,
@@ -3139,6 +3139,15 @@ async function checkAssertion(page, kind, value, timeoutMs = 3e3, assertion) {
   }
   fail(`expected ${sentence}`, last);
 }
+function whereOn(url) {
+  try {
+    const u = new URL(url);
+    if (u.protocol === "about:") return "";
+    return `${u.pathname}${u.hash.startsWith("#/") ? u.hash.split("?")[0] : ""}`;
+  } catch {
+    return "";
+  }
+}
 async function applyAction(loc, step, value, ctx = {}) {
   switch (step.type) {
     case "fill":
@@ -3620,6 +3629,10 @@ async function runStepsIn(session, steps, opts) {
             result.foundBy = foundBy;
           }
         });
+        try {
+          result.where = whereOn(actionPage.url());
+        } catch {
+        }
         for (const a of step.proposals ?? []) {
           if (a.enabled === false) continue;
           const label = describeAssertion(a, step.name);
@@ -10169,7 +10182,8 @@ var ASK_MODE_TOOLS = [
   "knowledge_search",
   "knowledge_read",
   "read_test_cases",
-  "list_my_branches"
+  "list_my_branches",
+  "run_insights"
 ];
 function buildAskArgs(opts) {
   const ask = opts.mode === "ask";
@@ -10177,7 +10191,7 @@ function buildAskArgs(opts) {
   const brief = opts.brief?.trim();
   const guidance = [
     `You are the Flowriki assistant for the project "${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. 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.",
+    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 learned something about the app the next job should know (a screen, a rule, a quirk, what failed and why), write it with write_knowledge (memory/sessions/<date>-<topic>.md, append) \u2014 Flowriki keeps it and gives it to the next job.",
     "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 ? `

package/dist/{chunk-WQ22OUIQ.js → chunk-YCFTNJQU.js}

@@ -656,6 +656,8 @@ import { z as z7 } from "zod";
 import { z as z8 } from "zod";
 import { z as z9 } from "zod";
 import { z as z10 } from "zod";
+import { z as z11 } from "zod";
+import { z as z12 } from "zod";
 var EngineSchema = z3.enum(["chromium", "firefox", "webkit", "android", "ios"]);
 var CapabilitySchema = z3.object({
   id: z3.string(),
@@ -1820,6 +1822,42 @@ function decodeEntities(text) {
   return out;
 }
 var nameKey = (name) => decodeEntities(name).toLowerCase().replace(/&/g, " and ").replace(/[^\p{L}\p{N}]+/gu, " ").trim();
+var PLATFORM_PART = "(web|android|ios)";
+var LEARNED_PATH = new RegExp(`^(explore\\/${PLATFORM_PART}\\.md|memory\\/sessions\\/[a-z0-9][a-z0-9-]{0,120}\\.md)$`);
+var isLearnedPath = (path8) => LEARNED_PATH.test(path8);
+var LearnedSchema = z11.object({
+  id: z11.string().regex(/^lrn_[a-z0-9_-]+$/),
+  projectId: z11.string().min(1),
+  /** explore: the app as the assistant walked it, one per platform.
+   *  session: what one job learned. */
+  kind: z11.enum(["explore", "session"]),
+  /** Its page path — what it is read and searched by. */
+  path: z11.string().refine(isLearnedPath, "not a learned page"),
+  platform: z11.enum(["web", "android", "ios"]).optional(),
+  content: z11.string().max(4e5),
+  by: z11.string().max(200).default(""),
+  /** new until a person acts on it: promoted into a written page, or
+   *  dismissed — kept out of the assistant's brief and search. */
+  status: z11.enum(["new", "promoted", "dismissed"]).default("new"),
+  createdAt: z11.string(),
+  updatedAt: z11.string()
+}).strict();
+var RUN_MEMORY_WEEKS = 5;
+var Bucket = z12.object({
+  week: z12.number().int(),
+  runs: z12.number().int().nonnegative(),
+  steps: z12.record(z12.string(), z12.unknown()),
+  screens: z12.record(z12.string(), z12.unknown()),
+  links: z12.record(z12.string(), z12.unknown())
+}).strict();
+var RunMemorySchema = z12.object({
+  id: z12.string().regex(/^rm_/),
+  projectId: z12.string().min(1),
+  /** Filled from the runs already kept, once — after that, run by run. */
+  backfilled: z12.boolean().default(false),
+  buckets: z12.array(Bucket).max(RUN_MEMORY_WEEKS + 1),
+  updatedAt: z12.number()
+}).strict();
 
 // ../engine-appium/dist/index.js
 import path5 from "path";
@@ -3765,6 +3803,7 @@ async function performSignIn(session, step, opts) {
   }
   return { ok: true, label, detail: "" };
 }
+var MOVES = /* @__PURE__ */ new Set(["click", "tap", "longPress", "press", "back", "navigate", "launchApp", "resetApp", "signIn", "signOut", "useHelper"]);
 async function runMobileSteps(session, steps, opts) {
   const results = [];
   let status = "passed";
@@ -3830,6 +3869,10 @@ async function runMobileSteps(session, steps, opts) {
         result.failureKind = e.kind ?? failureKind(e);
         result.detail = firstLine(e);
       }
+      if (result.status === "passed" && MOVES.has(step.type)) {
+        const at = await session.currentScreenName().catch(() => "");
+        if (at) result.where = at;
+      }
       const shots = opts.screenshots ?? "on-failure";
       if (opts.evidenceDir && (shots === "always" || shots === "on-failure" && result.status === "failed")) {
         const file = `step-${String((opts.stepOffset ?? 0) + results.length + 1).padStart(2, "0")}.png`;

package/dist/cli.js

@@ -24,7 +24,7 @@ import {
   runtimeValue,
   start,
   status
-} from "./chunk-BYFYSV5H.js";
+} from "./chunk-GPSLXDGQ.js";
 import {
   CLOUD_PROVIDERS,
   CLOUD_PROVIDER_NAME,
@@ -52,7 +52,7 @@ import {
   slugify,
   stepRefusedOn,
   targetFor
-} from "./chunk-WQ22OUIQ.js";
+} from "./chunk-YCFTNJQU.js";
 
 // src/cli.ts
 import { parseArgs } from "util";
@@ -266,6 +266,11 @@ var ServiceClient = class {
     if (!ASK_ID) return;
     await this.call(`/api/agent/jobs/draft?projectId=${encodeURIComponent(projectId)}`, { method: "POST", body: JSON.stringify({ draft }) });
   }
+  /** What runs taught us: failing steps, and the app's map. */
+  insights(projectId, document) {
+    const q = new URLSearchParams({ projectId, ...document ? { document } : {} });
+    return this.call(`/api/agent/insights?${q}`);
+  }
   /** The test case being built, as it stands, for the dashboard's job. */
   async reportCase(projectId, live) {
     if (!ASK_ID) return;
@@ -1780,6 +1785,33 @@ Helpers:${c.helpers.length ? "" : " none \u2014 create_helper makes one"}`,
         ].filter(Boolean).join("\n"));
       }
     },
+    {
+      name: "run_insights",
+      title: "What runs taught us",
+      description: "What the project's runs of the last 30 days show: the steps that fail most often \u2014 how often, whether they also pass sometimes (unreliable rather than broken), the last error \u2014 and the app's map: each page (web) or screen (Android) runs reached, what was done there, and which step leads where. Read it before healing a test, planning new ones, or answering what the project knows. document: one flow or test case, by name or id.",
+      inputSchema: { type: "object", properties: { projectId: { type: "string" }, document: { type: "string" } }, additionalProperties: false },
+      annotations: { readOnlyHint: true },
+      run: async (args) => {
+        const projectId = await projectFor(args["projectId"]);
+        const r = await service.insights(projectId, str(args["document"]));
+        if (!r.runs) return text("No runs in the last 30 days \u2014 nothing learned from runs yet.");
+        const lines = [`From the last ${r.runs} run${r.runs === 1 ? "" : "s"} (30 days).`];
+        lines.push(r.failing.length ? "\nSteps that failed, most often first:" : "\nNo step failed.");
+        for (const f of r.failing.slice(0, 30)) {
+          lines.push(`- ${f.document} (${f.documentKind}): ${f.step} \u2014 failed ${f.failures} of ${f.runs}${f.flaky ? ", passes sometimes (unreliable)" : ""}${f.lastKind ? `; last: ${f.lastKind}` : ""}${f.lastDetail ? ` \u2014 ${f.lastDetail}` : ""}`);
+        }
+        for (const m of r.maps) {
+          lines.push(`
+${m.platform === "web" ? "Web pages" : "Android screens"} runs reached:`);
+          for (const sc of m.screens.slice(0, 30)) lines.push(`- ${sc.where} (${sc.visits}\xD7)${sc.actions.length ? `: ${sc.actions.join(", ")}` : ""}`);
+          if (m.links.length) {
+            lines.push("Which leads where:");
+            for (const l of m.links.slice(0, 40)) lines.push(`- ${l.from} \u2192 ${l.to} by ${l.via} (${l.count}\xD7)`);
+          }
+        }
+        return text(lines.join("\n"));
+      }
+    },
     {
       name: "start_case",
       title: "Start building a test case",
@@ -2458,7 +2490,7 @@ ${pages.filter((p) => !shown.has(p.path)).map((p) => `- ${p.path}`).join("\n") |
     {
       name: "write_knowledge",
       title: "Write a knowledge page",
-      description: `Write what the project should remember, on Claude's working branch for review: product.md (About the product \u2014 the person's own description; write it only when they ask) and product/<web|android|ios>.md (what is particular to that platform's app), explore/<web|android|ios>.md (what you saw walking that platform: "# Explored: <Platform>", "Explored: <yyyy-mm-dd>", "Environment: <name>", "Screens: <n>" lines, then Screens, Journeys, Rules observed, Vocabulary, Suggested features), codebase/<folder>.md (what a linked folder holds \u2014 starting with "# <folder>", "Commit: <sha>" and "Learned: <yyyy-mm-dd>" lines, then routes, screens, forms and their validation, roles, APIs, flags; summarised, never copied code), features/<name>.md (screens, rules, quirks), memory/decisions.md (decisions and conventions), memory/sessions/<yyyy-mm-dd>-<topic>.md (what a job did). append: true adds to the end instead of replacing. Never a password, key or token \u2014 those are refused.`,
+      description: `Write what the project should remember, on Claude's working branch for review: product.md (About the product \u2014 the person's own description; write it only when they ask) and product/<web|android|ios>.md (what is particular to that platform's app), explore/<web|android|ios>.md (what you saw walking that platform: "# Explored: <Platform>", "Explored: <yyyy-mm-dd>", "Environment: <name>", "Screens: <n>" lines, then Screens, Journeys, Rules observed, Vocabulary, Suggested features), codebase/<folder>.md (what a linked folder holds \u2014 starting with "# <folder>", "Commit: <sha>" and "Learned: <yyyy-mm-dd>" lines, then routes, screens, forms and their validation, roles, APIs, flags; summarised, never copied code), features/<name>.md (screens, rules, quirks), memory/decisions.md (decisions and conventions), memory/sessions/<yyyy-mm-dd>-<topic>.md (what a job did \u2014 what you learned about the app that the next job should know). explore/\u2026 and memory/sessions/\u2026 are KEPT BY FLOWRIKI at once \u2014 no branch, nothing to publish \u2014 and given to the next job; the rest go on Claude's working branch for review. append: true adds to the end instead of replacing. Never a password, key or token \u2014 those are refused.`,
       inputSchema: {
         type: "object",
         properties: { projectId: { type: "string" }, path: { type: "string" }, content: { type: "string" }, append: { type: "boolean" }, branch: { type: "string" } },
@@ -2473,7 +2505,7 @@ ${pages.filter((p) => !shown.has(p.path)).map((p) => `- ${p.path}`).join("\n") |
           ...args["append"] === true ? { append: true } : {},
           ...str(args["branch"]) ? { changeId: str(args["branch"]) } : {}
         });
-        return text(`Wrote ${saved.path} (${saved.length} characters) on Claude's working branch ${saved.branch}.`);
+        return text(saved.keptBy === "flowriki" ? `Kept ${saved.path} (${saved.length} characters) in the project's memory \u2014 the next job is given it; nothing to publish.` : `Wrote ${saved.path} (${saved.length} characters) on Claude's working branch ${saved.branch}.`);
       }
     },
     {

package/dist/{dist-IBNVN2BD.js → dist-EGC7VEBG.js}

@@ -46,7 +46,7 @@ import {
   textOf,
   toolEnv,
   typesSecrets
-} from "./chunk-WQ22OUIQ.js";
+} from "./chunk-YCFTNJQU.js";
 export {
   AGENT_HOME,
   APPIUM_DIR,

package/dist/engine.js

@@ -9,8 +9,8 @@ import {
   resolveCheckouts,
   start,
   status
-} from "./chunk-BYFYSV5H.js";
-import "./chunk-WQ22OUIQ.js";
+} from "./chunk-GPSLXDGQ.js";
+import "./chunk-YCFTNJQU.js";
 export {
   AGENT_VERSION,
   browsersReady,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@flowriki/agent",
-  "version": "0.1.0-beta.60",
+  "version": "0.1.0-beta.62",
   "description": "Run Flowriki UI tests on your own computer. Connects this machine to your Flowriki account.",
   "license": "SEE LICENSE IN LICENSE",
   "type": "module",