@flowriki/agent 0.1.0-beta.33 → 0.1.0-beta.34

package/dist/{chunk-3UUCI66O.js → chunk-7OODSC7P.js}

@@ -6,7 +6,7 @@ import os from "os";
 import path from "path";
 import { createRequire } from "module";
 import { spawnSync } from "child_process";
-var AGENT_VERSION = "0.1.0-beta.33";
+var AGENT_VERSION = "0.1.0-beta.34";
 async function prepareBrowsers() {
   if (browsersReady()) {
     return { ok: true, already: true, message: "chromium is already installed" };

package/dist/cli.js

@@ -21,7 +21,7 @@ import {
   runtimeValue,
   start,
   status
-} from "./chunk-3UUCI66O.js";
+} from "./chunk-7OODSC7P.js";
 
 // src/cli.ts
 import { parseArgs } from "util";
@@ -39,7 +39,7 @@ async function handle(message, tools, info) {
       const version = PROTOCOL_VERSIONS.includes(asked) ? asked : PROTOCOL_VERSIONS[0];
       return reply({
         protocolVersion: version,
-        capabilities: { tools: { listChanged: false } },
+        capabilities: { tools: { listChanged: true } },
         serverInfo: { name: info.name, version: info.version },
         instructions: info.instructions
       });
@@ -95,6 +95,10 @@ function serve(tools, info, io = {}) {
     const task = handle(message, tools, info).then((answer) => {
       if (answer) output.write(`${JSON.stringify(answer)}
 `);
+      if (message.method === "notifications/initialized") {
+        output.write(`${JSON.stringify({ jsonrpc: "2.0", method: "notifications/tools/list_changed" })}
+`);
+      }
     });
     pending.add(task);
     void task.finally(() => pending.delete(task));
@@ -168,16 +172,36 @@ var ServiceClient = class {
       { method: "DELETE" }
     );
   }
-  saveFlow(projectId, title, flow) {
+  saveFlow(projectId, title, flow, where = {}) {
     return this.call(
       `/api/agent/flows?projectId=${encodeURIComponent(projectId)}`,
-      { method: "POST", body: JSON.stringify({ title, flow }) }
+      { method: "POST", body: JSON.stringify({ title, flow, ...where }) }
+    );
+  }
+  /** Gather Claude's open branches here into one. */
+  gatherChanges(projectId, opts = {}) {
+    return this.call(
+      `/api/agent/changes/gather?projectId=${encodeURIComponent(projectId)}`,
+      { method: "POST", body: JSON.stringify(opts) }
     );
   }
 };
 
 // src/mcp/session.ts
 var MAX_TREE = 14e3;
+var MAX_STEP_VIEW = 6e3;
+function topLevelBlocks(tree) {
+  const blocks = [];
+  for (const line of tree.split("\n")) {
+    if (line.startsWith("- ") || blocks.length === 0) blocks.push([line]);
+    else blocks[blocks.length - 1].push(line);
+  }
+  const named = (b, kinds) => kinds.some((k) => b[0].startsWith(`- ${k}`));
+  return {
+    dialogs: blocks.filter((b) => named(b, ["dialog", "alertdialog"])).map((b) => b.join("\n")),
+    navigation: blocks.find((b) => named(b, ["navigation", "banner"]))?.join("\n") ?? null
+  };
+}
 var STEP_TIMEOUT_MS = 1e4;
 var ACTIONS = ["click", "dblclick", "rightclick", "hover", "fill", "select", "check", "press", "navigate", "back", "forward"];
 var isAction = (value) => ACTIONS.includes(String(value));
@@ -238,6 +262,7 @@ var AuthoringSession = class {
   page = null;
   values = /* @__PURE__ */ new Map();
   resolvers = null;
+  lastTree = null;
   draft = null;
   catalog = null;
   get open() {
@@ -272,6 +297,7 @@ var AuthoringSession = class {
     this.catalog = catalog;
     this.draft = { ...input, steps: [] };
     this.values = /* @__PURE__ */ new Map();
+    this.lastTree = null;
     const base = this.baseUrl();
     this.run = await openRunSession({ steps: [], headless: this.opts.headless ?? false });
     this.page = this.run.page;
@@ -304,19 +330,52 @@ var AuthoringSession = class {
   }
   /** The page as Claude reads it: where it is, and what is on it by role and name. */
   async snapshot() {
+    return this.view(true);
+  }
+  /** WHAT CHANGED, after a step — not the whole page again.
+   *
+   * Every step used to answer with the full tree, which on a long storefront
+   * is twenty thousand characters repeated forty times to say "the box now
+   * has text in it". So a step answers with what it is about: an open dialog
+   * (with the navigation, which says who is signed in), or the lines that
+   * appeared and disappeared — and only after a real navigation, when most of
+   * the page is new, the page itself, capped. read_page is always the whole
+   * thing. */
+  async view(full) {
     const { page, draft } = this.requireOpen();
     await page.waitForLoadState("domcontentloaded", { timeout: 1e4 }).catch(() => void 0);
-    let tree = labelTree(
+    const tree = labelTree(
       await page.locator("body").ariaSnapshot({ timeout: 8e3 }).catch((e) => `(could not read the page: ${e.message})`),
       await fieldNames(page)
     );
-    if (tree.length > MAX_TREE) tree = `${tree.slice(0, MAX_TREE)}
-\u2026 (cut \u2014 the page is long; act on what is above, or scroll with press "PageDown")`;
     const title = await page.title().catch(() => "");
-    return `Page: ${page.url()}${title ? ` \u2014 "${title}"` : ""}
+    const head = `Page: ${page.url()}${title ? ` \u2014 "${title}"` : ""}
 Steps so far: ${draft.steps.length}
 
-${tree}`;
+`;
+    const before = this.lastTree;
+    this.lastTree = tree;
+    const cap = (t, max) => t.length > max ? `${t.slice(0, max)}
+\u2026 (cut \u2014 read_page shows the whole page)` : t;
+    if (full || before === null) return head + cap(tree, MAX_TREE);
+    const { dialogs, navigation } = topLevelBlocks(tree);
+    if (dialogs.length) {
+      return head + cap([...navigation ? [navigation] : [], ...dialogs].join("\n"), MAX_STEP_VIEW) + "\n(only the open dialog and the navigation \u2014 read_page shows the whole page)";
+    }
+    const was = new Set(before.split("\n"));
+    const now = new Set(tree.split("\n"));
+    const added = tree.split("\n").filter((l) => !was.has(l));
+    const removed = before.split("\n").filter((l) => !now.has(l));
+    if (added.length === 0 && removed.length === 0) return head + "(the page did not change \u2014 read_page shows it)";
+    if (added.length + removed.length > 120) return head + cap(tree, MAX_STEP_VIEW);
+    return head + [
+      ...navigation ? [navigation, ""] : [],
+      added.length ? `Appeared:
+${added.join("\n")}` : "",
+      removed.length ? `Gone:
+${removed.join("\n")}` : "",
+      "(changes since the last step \u2014 read_page shows the whole page)"
+    ].filter(Boolean).join("\n");
   }
   /** Do one thing on the page, and keep it as a step if it worked. */
   async act(input) {
@@ -354,13 +413,13 @@ ${tree}`;
     } catch (error) {
       throw new Error(`That did not work, and was not kept: ${error.message.split("\n")[0]}
 
-${await this.snapshot()}`);
+${await this.view(false)}`);
     }
     await this.settle();
     draft.steps.push(step);
     return `Step ${draft.steps.length} kept: ${describe(step)}.
 
-${await this.snapshot()}`;
+${await this.view(false)}`;
   }
   /** Check something about the page now, and keep the check if it holds. */
   async check(input) {
@@ -437,7 +496,7 @@ ${await this.snapshot()}`;
     draft.steps.push(step);
     return `Step ${draft.steps.length} kept: ${describe(step)}. It got "${got.value}" (${got.detail}). Use {{${step.helper.saveAs}}} as the value of a later step.` + (step.helper.then === "open" ? `
 
-${await this.snapshot()}` : "");
+${await this.view(false)}` : "");
   }
   list() {
     const draft = this.draft;
@@ -473,24 +532,26 @@ ${this.list()}`;
 ${lines.join("\n")}`;
   }
   /** Save the draft as a flow on a new branch, for a person to review. */
-  async save(title) {
+  async save(title, where = {}) {
     const { draft } = this.requireOpen();
     if (draft.steps.length === 0) throw new Error("there are no steps to save yet");
     if (!draft.steps.some((s) => s.type === "assert" || s.proposals.length > 0)) {
       throw new Error("the flow checks nothing \u2014 add at least one check (add_check) that proves it worked, then save");
     }
-    const saved = await this.service.saveFlow(draft.projectId, title || `Write "${draft.name}" with Claude`, {
+    const saved = await this.service.saveFlow(draft.projectId, title || "Tests written with Claude", {
       name: draft.name,
       ...draft.featureId ? { featureId: draft.featureId } : { featureName: draft.featureName },
       startPath: draft.startPath,
       steps: draft.steps,
       ...draft.signIn ? { signIn: draft.signIn } : {}
-    });
+    }, where);
     const review = `${this.service.origin}/projects/${draft.projectId}/changes/${saved.changeId}`;
+    const what = saved.replaced ? `Replaced the earlier "${draft.name}"` : `Saved "${draft.name}"`;
+    const where_ = saved.newBranch ? `on a new branch, ${saved.branch}` : `on Claude's working branch, ${saved.branch}`;
     const others = await this.service.agentChanges(draft.projectId).then((r) => r.changes.filter((c) => c.id !== saved.changeId)).catch(() => []);
     const note = others.length ? `
-Claude also has ${others.length} other open branch${others.length === 1 ? "" : "es"} on this project \u2014 list_my_branches shows them, if one is now superseded.` : "";
-    return `Saved "${draft.name}" on a new branch, ${saved.branch}. Nothing is published yet.
+Claude has ${others.length} other open branch${others.length === 1 ? "" : "es"} on this project \u2014 gather_branches puts them into this one, for one review.` : "";
+    return `${what} ${where_}. Nothing is published yet.
 Review and publish it here: ${review}${note}`;
   }
   async close() {
@@ -537,11 +598,15 @@ How to write a good Flowriki test:
    a heading, a message, the URL. Saving is refused for a flow that checks nothing.
 7. Before saving, call replay_flow to run the whole flow from the start. If a
    step fails there, fix it (remove_step, act again) and replay again.
-8. Save with save_flow and give the person the review link it returns.
-9. Redoing a flow? Each save is a new branch, so clear the superseded one with
-   list_my_branches and discard_branch \u2014 only after the person asks for it, or
-   agrees, because it deletes the branch. Only Claude's own unedited branches
-   can go this way; anything a person touched is theirs to delete.
+8. Save with save_flow and give the person the review link it returns. Saves go
+   to ONE working branch per project \u2014 everything you write here lands on it,
+   for one review. Saving a flow with the same name again replaces it there.
+   The catalog lists what is already on that branch: reuse its features, and
+   do not write a flow twice.
+9. Old branches of yours from before (one per flow)? gather_branches puts them
+   into one. discard_branch deletes one \u2014 only after the person asks for it or
+   agrees. Only Claude's own unedited branches can be gathered or deleted;
+   anything a person touched is theirs.
 Test data you invent (a new user's email or password for a sign-up) is fine to
 type. For anything that must be NEW on every run \u2014 a sign-up's email or
 username \u2014 use {{unique}}: it is different on every run and the same within
@@ -621,6 +686,12 @@ It is now the project these tools work on. Open it: ${service.origin}/projects/$
 Features:${c.features.length ? "" : " none yet"}`,
           ...c.features.map((f) => `  - ${f.name} (id ${f.id})${f.description ? ` \u2014 ${f.description}` : ""}
 ${flowsBy(f.id)}`),
+          ...c.draft ? [
+            `
+On Claude's working branch "${c.draft.title}" (${c.draft.changeId}), not yet published:`,
+            ...c.draft.features.length ? [`  New features: ${c.draft.features.map((f) => `${f.name} (id ${f.id})`).join(", ")}`] : [],
+            ...c.draft.flows.length ? c.draft.flows.map((f) => `    - ${f.name} (${f.steps} steps, feature ${f.featureId})`) : ["  (no flows yet)"]
+          ] : [],
           `
 Sign-in profiles:${c.signIns.length ? "" : " none"}`,
           ...c.signIns.map((s) => `  - ${s.name} (profileId ${s.profileId}): ${s.accounts.map((a) => `${a.name} (accountId ${a.accountId})`).join(", ") || "no accounts yet"}`),
@@ -783,9 +854,20 @@ ${page}`);
     {
       name: "save_flow",
       title: "Save the flow to a branch",
-      description: "Save the flow to a new branch in the project's repository, for a person to review and publish. Returns the review link. Refused if the flow checks nothing.",
-      inputSchema: { type: "object", properties: { title: { type: "string", description: "The branch's title \u2014 what this change adds." } }, additionalProperties: false },
-      run: async (args) => text(await session.save(str(args["title"])))
+      description: "Save the flow onto Claude's working branch for this project (made on the first save), for a person to review and publish. Saving the same name again replaces the earlier version there. Returns the review link. Refused if the flow checks nothing.",
+      inputSchema: {
+        type: "object",
+        properties: {
+          title: { type: "string", description: "The branch's title, used only when a new branch is started." },
+          branch: { type: "string", description: "Optional: a branch id (chg_\u2026) from list_my_branches to save onto instead." },
+          newBranch: { type: "boolean", description: "Start a separate branch instead of the working one. Only when the person asks for separate reviews." }
+        },
+        additionalProperties: false
+      },
+      run: async (args) => text(await session.save(str(args["title"]), {
+        ...str(args["branch"]) ? { changeId: str(args["branch"]) } : {},
+        ...args["newBranch"] === true ? { newBranch: true } : {}
+      }))
     },
     {
       name: "list_my_branches",
@@ -800,6 +882,30 @@ ${page}`);
         return text(changes.map((c) => `- ${c.id} \u2014 "${c.title}" (${new Date(c.createdAt).toISOString().slice(0, 16).replace("T", " ")}) ${service.origin}${c.path}`).join("\n"));
       }
     },
+    {
+      name: "gather_branches",
+      title: "Gather Claude's branches into one",
+      description: "Move the flows from Claude's open branches on this project into one branch, and delete the emptied ones \u2014 so the person reviews and publishes once. Only Claude's own, unedited branches; anything a person touched stops it.",
+      inputSchema: {
+        type: "object",
+        properties: {
+          into: { type: "string", description: "The branch to gather into (chg_\u2026). Default: Claude's oldest open branch." },
+          title: { type: "string", description: 'A title for the gathered branch, e.g. "Sign in and sign up tests".' },
+          projectId: { type: "string" }
+        },
+        additionalProperties: false
+      },
+      annotations: { destructiveHint: true },
+      run: async (args) => {
+        const projectId = await projectFor(args["projectId"]);
+        const done = await service.gatherChanges(projectId, {
+          ...str(args["into"]) ? { into: str(args["into"]) } : {},
+          ...str(args["title"]) ? { title: str(args["title"]) } : {}
+        });
+        catalogs.delete(projectId);
+        return text(`Gathered ${done.gathered.length} branch${done.gathered.length === 1 ? "" : "es"} into "${done.title}" (${done.changeId}). Review and publish it here: ${service.origin}${done.path}`);
+      }
+    },
     {
       name: "discard_branch",
       title: "Delete one of Claude's branches",

package/dist/engine.js

@@ -9,7 +9,7 @@ import {
   resolveCheckouts,
   start,
   status
-} from "./chunk-3UUCI66O.js";
+} from "./chunk-7OODSC7P.js";
 export {
   AGENT_VERSION,
   browsersReady,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@flowriki/agent",
-  "version": "0.1.0-beta.33",
+  "version": "0.1.0-beta.34",
   "description": "Run Flowriki UI tests on your own computer. Connects this machine to your Flowriki account.",
   "license": "SEE LICENSE IN LICENSE",
   "type": "module",