@aarwitz/tapp 0.17.3 → 0.17.4

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "tapp",
   "description": "Give Claude hands and eyes on iOS, Android, and web apps, with exploration, replayable flows, evidence, and deterministic CI gates.",
-  "version": "0.17.3",
+  "version": "0.17.4",
   "author": {
     "name": "Aaron Horowitz",
     "url": "https://github.com/aarwitz"
@@ -24,7 +24,7 @@
       "command": "npx",
       "args": [
         "-y",
-        "@aarwitz/tapp@0.17.3",
+        "@aarwitz/tapp@0.17.4",
         "mcp"
       ],
       "cwd": "${CLAUDE_PROJECT_DIR}"

package/AGENTS.md

@@ -106,7 +106,13 @@ Rules that prevent 90% of failures:
    secret-templated replay step and avoids native secure-field refocus behavior.
 5. Tap results: `ok` (landed), `not_hittable` (exists but disabled/covered — the harness
    auto-dismisses keyboards and retries), `not_found` (nothing matches — re-read the tree).
-6. One session at a time. `session_start` always begins from a fresh app launch.
+6. One session at a time. `session_start` always begins from a fresh **cold** launch
+   (terminate + relaunch, for a deterministic starting screen). Persisted app data such as
+   Keychain credentials survives, but the app opens on its launch screen, not a resumed
+   foreground state — an app that gates each cold start behind sign-in WILL show its login
+   wall, so plan `login` (or a bypass launch argument) as the first act. Plain `tapp tree` /
+   `tapp screenshot` warm-resume the currently foregrounded app instead, which is why they can
+   look signed-in when a fresh session does not.
 
 ## Autonomous exploration (`tapp_explore`)
 
@@ -145,6 +151,11 @@ without a coding agent, model, subscription, or API key. AI generation and `asse
   a final screen assertion auto-inserted; typed credentials are templated to `$TEST_EMAIL`/`$TEST_PASSWORD`.
 - **Replay:** `tapp_flow_run { flowPath: ".tapp/flows/checkout.yml" }` — exact steps,
   deterministic assertions, same result every time. A failed assertion is a finding.
+- **Credentials at replay:** pass real values (`testEmail`/`testPassword`, CLI `--email`/
+  `--password`), or name a configured actor (`actor: "coach"`, CLI `--actor coach`) and Tapp
+  resolves `$TEST_EMAIL`/`$TEST_PASSWORD` from the env vars that actor binds. Actors store
+  env-var **names** only — never values. `tapp actor set` refuses to overwrite an existing
+  actor unless you pass `--replace`, so idempotent setup scripts must include it.
 - **Generate:** `tapp_flow_generate { goal: "log in and add the first item to cart" }` —
   grounded in the app's actually-explored screens, so it can't invent steps.
 - **Discover the file format without MCP:** `npx -y @aarwitz/tapp@latest flow example` prints a
@@ -161,6 +172,15 @@ without a coding agent, model, subscription, or API key. AI generation and `asse
   and installs from an Xcode project/workspace; or the user's normal build).
 - A simulator must be booted (`tapp_list_simulators` → `tapp_boot_simulator`).
 - Screenshots/captures land in `~/.tapp/captures/`.
+- Driving `tapp mcp` from a raw stdio client: **consume or discard stderr** — the server logs
+  progress there, and an unread stderr pipe can deadlock a naive client. The first
+  `tapp_session_start` on a cold machine includes the one-time harness build, so it can take
+  minutes before the first result arrives; that is startup cost, not a hang.
+- Managed web targets always bind `127.0.0.1` and prefer the repository's declared or framework
+  default port (vite → 5173, next → 3000). If the app's backend uses a CORS allowlist, pin the
+  origin with `"web": { "port": 5173 }` in `.tapp/project.json` — an unexpected port surfaces as
+  misleading fetch/CORS findings, and a busy pinned port is a hard error, never a silent
+  ephemeral fallback.
 
 ## Honesty rules
 

package/README.md

@@ -348,7 +348,7 @@ jobs:
     timeout-minutes: 45
     steps:
       - uses: actions/checkout@v4
-      - uses: aarwitz/tapp@v0.17.3 # or pin the reviewed release commit SHA
+      - uses: aarwitz/tapp@v0.17.4 # or pin the reviewed release commit SHA
         with:
           project: MyApp.xcodeproj # or MyApp.xcworkspace
           scheme: MyApp
@@ -398,7 +398,7 @@ Android CI runs on Linux with an emulator/device already connected. The Action c
 or accept a prebuilt one:
 
 ```yaml
-- uses: aarwitz/tapp@v0.17.3 # or pin the reviewed release commit SHA
+- uses: aarwitz/tapp@v0.17.4 # or pin the reviewed release commit SHA
   with:
     platform: android
     android-app-id: com.acme.app

package/bin/tapp.js

@@ -270,7 +270,7 @@ function safeCommandUsage(verb) {
     shot: "tapp shot [--out FILE]",
     apps: "tapp apps",
     build: "tapp build [repo] [--scheme NAME] [--configuration NAME]",
-    flow: "tapp flow example\ntapp flow validate FILE [--platform PLATFORM] [--map FILE]\ntapp flow run FILE [--email VALUE] [--password VALUE]",
+    flow: "tapp flow example\ntapp flow validate FILE [--platform PLATFORM] [--map FILE]\ntapp flow run FILE [--actor NAME] [--email VALUE] [--password VALUE]",
     task: "tapp task validate FILE [--platform PLATFORM] [--map FILE]\ntapp task compile FILE --platform PLATFORM [--inputs JSON] [--out FILE]\ntapp task run FILE --platform PLATFORM [--url URL|--bundle-id ID|--app-id ID] [--inputs JSON]",
     contract: "tapp contract validate FILE [--platform PLATFORM] [--map FILE]\ntapp contract compile FILE --platform PLATFORM [--out FILE]\ntapp contract run FILE --platform PLATFORM [--url URL|--bundle-id ID|--app-id ID]",
     scenario: "tapp scenario validate FILE [--project-dir DIR]\ntapp scenario run FILE --platform web --url URL [--project-dir DIR]",
@@ -279,7 +279,7 @@ function safeCommandUsage(verb) {
     plan: "tapp plan show [FILE]\ntapp plan review [FILE] --approve NAME[,NAME] --reject NAME[,NAME] --defer NAME[,NAME]\ntapp plan generate|validate|promote [FILE] [options]",
     baseline: "tapp baseline create [repo] [--platform PLATFORM] [--target NAME] [--from GATE.json] [--replace]",
     ci: "tapp ci ...\ntapp ci install [repo] [--out FILE] [--manifest FILE] [--dry-run] [--replace]",
-    actor: "tapp actor set NAME --email-env ENV --password-env ENV [--project-dir DIR]\ntapp actor list [repo]",
+    actor: "tapp actor set NAME --email-env ENV --password-env ENV [--replace] [--project-dir DIR]\ntapp actor list [repo]",
     app: "tapp app [repo] [--no-open] [--port PORT]",
     report: "tapp report [captureId|latest]",
     doctor: "tapp doctor",
@@ -1111,7 +1111,7 @@ switch (command) {
       break;
     }
     if (!["run", "validate"].includes(verb) || !flowPath) {
-      console.error("usage: tapp flow example\n       tapp flow run <flow.yml> [--platform ios|android|web] [--url URL] [--app-id ID] [--apk FILE] [--serial ID]\n       tapp flow validate <flow.yml>");
+      console.error("usage: tapp flow example\n       tapp flow run <flow.yml> [--platform ios|android|web] [--actor NAME] [--email VALUE] [--password VALUE] [--url URL] [--app-id ID] [--apk FILE] [--serial ID]\n       tapp flow validate <flow.yml>");
       process.exit(2);
     }
     const absolute = path.resolve(flowPath);
@@ -1146,6 +1146,28 @@ switch (command) {
     const env = { ...process.env, FLOW_LOG: flowLog, TAPP_FLOW_EVIDENCE_DIR: evidenceDir };
     if (typeof flags.email === "string") env.OCQA_TEST_EMAIL = flags.email;
     if (typeof flags.password === "string") env.OCQA_TEST_PASSWORD = flags.password;
+    if (typeof flags.actor === "string" && flags.actor) {
+      const { readProjectConfig } = await import(path.join(packageRoot, "mcp-server", "src", "project-config.js"));
+      const loaded = readProjectConfig(process.cwd());
+      if (loaded.errors.length) { console.error(`❌ Invalid ${loaded.relativePath}: ${loaded.errors.join("; ")}`); process.exit(2); }
+      const actor = loaded.config.actors?.[flags.actor];
+      if (!actor) {
+        console.error(`❌ Actor '${flags.actor}' is not configured in ${loaded.relativePath}. Run from the repository root, or configure it: tapp actor set ${flags.actor} --email-env ENV --password-env ENV`);
+        process.exit(2);
+      }
+      // Actors store env-var NAMES only; resolve the values here. Explicit --email/--password win.
+      for (const [credential, flagName, envKey] of [["email", "email", "OCQA_TEST_EMAIL"], ["password", "password", "OCQA_TEST_PASSWORD"]]) {
+        if (typeof flags[flagName] === "string") continue;
+        const binding = actor.credentials?.[credential];
+        if (!binding) continue;
+        const value = process.env[binding.env];
+        if (!value) {
+          console.error(`❌ Actor '${flags.actor}' binds ${credential} to $${binding.env}, but that environment variable is not set.`);
+          process.exit(2);
+        }
+        env[envKey] = value;
+      }
+    }
     let invocation;
     if (platform === "web") {
       const url = typeof flags.url === "string" ? flags.url : flow.url || flow.app;

package/docs/application-model.md

@@ -95,7 +95,11 @@ derived lockfile-backed install command, runs its declared build script when pre
 `start`, `dev`, `serve`, or `preview` package script with argument-array process execution (never
 generated shell source). A static site with no script uses Tapp's local read-only static server. The
 runtime binds to an available loopback port, writes its log under the Tapp runtime directory, and is
-terminated after exploration even when QA fails. Multiple web targets, an unlocked dependency
+terminated after exploration even when QA fails. Port precedence is: an explicit
+`"web": { "port": N }` pin in `.tapp/project.json`, then a port declared by the start script, then
+the framework default (vite 5173, next 3000), then an ephemeral port. Pin the port when a backend
+CORS allowlist expects a fixed origin; a busy or contradicted pin is a hard startup error rather
+than a silent fallback that would resurface as fetch/CORS findings. The host is always `127.0.0.1`. Multiple web targets, an unlocked dependency
 graph, an unrecognized start path, or backend-specific configuration produce explicit remediation;
 provide `--target` and/or an already-running owned `--url` in those cases. Running repository build
 scripts executes repository code and should only be used for a checkout the customer trusts.

package/docs/scenarios.md

@@ -74,7 +74,7 @@ tapp ci --platform web --url http://127.0.0.1:4180 \
 GitHub Action:
 
 ```yaml
-- uses: aarwitz/tapp@v0.17.3 # or pin the reviewed release commit SHA
+- uses: aarwitz/tapp@v0.17.4 # or pin the reviewed release commit SHA
   with:
     platform: web
     url: http://127.0.0.1:4180

package/mcp-server/src/application-model.js

@@ -91,8 +91,15 @@ function applyRuntimeTargetValidation(root, targets, validation) {
   if (!container || !scheme || !bundleId) return targets;
   const captureId = String(validation.evidence?.captureId || "").trim();
   const explored = !!captureId;
+  // The build records the container findXcodeContainer resolved (shallowest workspace-first),
+  // which can differ from the modeled sourcePath in a repo exposing both a workspace and a
+  // project. With exactly one iOS target there is no ambiguity — apply the validation rather
+  // than silently dropping it and re-blocking on scheme confirmation after every refresh.
+  const iosTargets = targets.filter((target) => target.platform === "ios");
+  const pathMatched = iosTargets.some((target) => posix(target.sourcePath) === container);
   return targets.map((target) => {
-    if (target.platform !== "ios" || posix(target.sourcePath) !== container) return target;
+    if (target.platform !== "ios") return target;
+    if (posix(target.sourcePath) !== container && (pathMatched || iosTargets.length !== 1)) return target;
     return {
       ...target,
       status: "configured",
@@ -917,12 +924,25 @@ function mergePlanDecisions(next, prior, { invalidateValidation = false } = {})
     priorByNameScope.set(key, list);
   }
   const consumedPriorIds = new Set();
+  const derivedScopes = new Set(next.items.map((item) => item.scope || "."));
   const carried = next.items.map((item) => {
     const exact = priorItems.get(item.id);
     const lineage = item.origin === "committed"
       ? (priorByNameScope.get(`${item.scope || "."}|${item.name}`) || []).find((candidate) => candidate.origin === "promoted-validated" || candidate.generation?.status === "promoted")
       : null;
-    const previous = exact || lineage;
+    // Proposal ids hash the node's targetId, which can drift between refreshes when target
+    // detection or map attribution shifts. Without name+scope regrounding, that drift strands
+    // the reviewed item as stale AND re-adds the same proposal as pending — a duplicate the
+    // customer already decided.
+    const sameNameScope = exact || lineage ? [] : (priorByNameScope.get(`${item.scope || "."}|${item.name}`) || []).filter((candidate) => !consumedPriorIds.has(candidate.id));
+    const regrounded = sameNameScope.find((candidate) => candidate.decision && candidate.decision !== "pending") || sameNameScope[0] || null;
+    // When a scope stops being derived entirely (its target no longer grounds any proposals — the
+    // mis-attribution case), a decided same-name item from that vanished scope carries the
+    // customer's decision onto the surviving surface instead of lingering as a stale duplicate
+    // beside a re-added pending twin. The migration is recorded via regroundedFromScope.
+    const crossScope = exact || lineage || regrounded ? null
+      : (prior.items || []).find((candidate) => !consumedPriorIds.has(candidate.id) && candidate.name === item.name && !derivedScopes.has(candidate.scope || ".") && candidate.decision && candidate.decision !== "pending") || null;
+    const previous = exact || lineage || regrounded || crossScope;
     if (!previous) return item;
     consumedPriorIds.add(previous.id);
     if (exact && lineage && exact.id !== lineage.id) consumedPriorIds.add(lineage.id);
@@ -936,10 +956,20 @@ function mergePlanDecisions(next, prior, { invalidateValidation = false } = {})
     // A reviewed proposal becomes a committed contract without becoming a different
     // customer decision. Preserve its original plan identity so browser links, CLI
     // item selectors, and review history remain stable across promotion refreshes.
-    return { ...item, id: previous.id, ...human };
+    return { ...item, id: previous.id, ...human, ...(crossScope ? { regroundedFromScope: previous.scope || "." } : {}) };
   });
   const currentIds = new Set(carried.map((item) => item.id));
-  for (const previous of prior.items || []) if (!currentIds.has(previous.id) && !consumedPriorIds.has(previous.id) && previous.decision && previous.decision !== "pending") carried.push({ ...previous, stale: true, status: "not-derived-on-refresh" });
+  const survivorNames = new Set(carried.map((item) => item.name));
+  for (const previous of prior.items || []) {
+    if (currentIds.has(previous.id) || consumedPriorIds.has(previous.id) || !previous.decision || previous.decision === "pending") continue;
+    const committed = previous.origin === "committed" || previous.origin === "promoted-validated" || previous.generation?.status === "promoted";
+    // A decided proposal whose scope is no longer derived duplicates the surviving same-name
+    // decision — drop the duplicate rather than carrying it as stale forever. If that scope is
+    // re-derived later, its proposals return as pending review; the failure direction is more
+    // human review, never a silently granted decision. Committed/promoted items always carry.
+    if (!committed && !derivedScopes.has(previous.scope || ".") && survivorNames.has(previous.name)) continue;
+    carried.push({ ...previous, stale: true, status: "not-derived-on-refresh" });
+  }
   let generation = prior.generation;
   if (invalidateValidation && generation) generation = {
     ...generation,

package/mcp-server/src/index.js

@@ -634,9 +634,9 @@ export function agentFacingElements(elements, limit = 160) {
   return result;
 }
 
-function agentScreenProjection(screen) {
+function agentScreenProjection(screen, { full = false } = {}) {
   const all = screen?.elements || [];
-  const elements = agentFacingElements(all);
+  const elements = full ? all : agentFacingElements(all);
   return {
     screenTitle:screen?.screenTitle ?? null,
     elementCount:elements.length,
@@ -843,22 +843,32 @@ export function isStableFlowCheckpoint(value) {
   const text = String(value || "").replace(/\s+/g, " ").trim();
   if (!text || /^(loading|fetching|please wait|preparing|connecting|syncing|signing in)(?:[.…!]*|\s.*)$/i.test(text)) return false;
   if (/^(?:mon|tues?|wed(?:nes)?|thu(?:rs)?|fri|sat(?:ur)?|sun)(?:day)?\b/i.test(text)) return false;
-  if (/^(?:jan(?:uary)?|feb(?:ruary)?|mar(?:ch)?|apr(?:il)?|may|jun(?:e)?|jul(?:y)?|aug(?:ust)?|sep(?:tember)?|oct(?:ober)?|nov(?:ember)?|dec(?:ember)?)\s+\d{1,2}(?:,\s+\d{4})?$/i.test(text)) return false;
+  // Single dates ("Aug 16, 2026") and date ranges ("AUG 16 - AUG 22, 2026") both roll over on
+  // the calendar, so neither is a stable replay checkpoint.
+  const monthDay = "(?:jan(?:uary)?|feb(?:ruary)?|mar(?:ch)?|apr(?:il)?|may|jun(?:e)?|jul(?:y)?|aug(?:ust)?|sep(?:tember)?|oct(?:ober)?|nov(?:ember)?|dec(?:ember)?)\\s+\\d{1,2}";
+  if (new RegExp(`^${monthDay}(?:,\\s+\\d{4})?(?:\\s*[-–—]\\s*(?:${monthDay}|\\d{1,2})(?:,\\s+\\d{4})?)?$`, "i").test(text)) return false;
   if (/^\d{4}-\d{2}-\d{2}(?:[ T].*)?$/.test(text)) return false;
   return true;
 }
 
 export function semanticTargetAtPoint(elements, x, y) {
   if (!Number.isFinite(x) || !Number.isFinite(y)) return "";
-  return (elements || [])
+  const containing = (elements || [])
     .filter((element) => {
       const frame = element.frame || {};
-      return element.hittable !== false && Number.isFinite(frame.x) && Number.isFinite(frame.y) && Number.isFinite(frame.width) && Number.isFinite(frame.height)
+      return Number.isFinite(frame.x) && Number.isFinite(frame.y) && Number.isFinite(frame.width) && Number.isFinite(frame.height)
         && x >= frame.x && y >= frame.y && x <= frame.x + frame.width && y <= frame.y + frame.height
         && String(element.id || element.identifier || element.label || "").trim();
     })
-    .sort((a, b) => (a.frame.width * a.frame.height) - (b.frame.width * b.frame.height))
-    .map((element) => String(element.id || element.identifier || element.label || "").trim())[0] || "";
+    .sort((a, b) => (a.frame.width * a.frame.height) - (b.frame.width * b.frame.height));
+  // Prefer a hittable match. A labelled control that reports hittable=false at its own
+  // tree-reported center (custom tab bars over safe-area insets do this) still responds to an
+  // element tap, which retries in the harness — so fall back to it rather than firing a raw
+  // coordinate tap that lands on nothing. Containers stay excluded: redirecting a coordinate
+  // tap to the center of a large labelled group would change where the tap lands.
+  const best = containing.find((element) => element.hittable !== false)
+    || containing.find((element) => /button|link|cell|tab|switch|checkbox|toggle|textfield|securetext|textarea|edittext|segmented|menuitem/i.test(`${element.role || ""} ${element.type || ""}`));
+  return best ? String(best.id || best.identifier || best.label || "").trim() : "";
 }
 
 /** Append a Flow step for an act (record-by-doing). Inserts wait_for on screen change for
@@ -2258,8 +2268,21 @@ export async function startManagedWebTarget({ root, requestedTarget = "", timeou
   const dependencies = { ...(pkg.dependencies || {}), ...(pkg.devDependencies || {}) };
   const declaredPort = startMatch ? declaredPortFromStartScript(pkg.scripts?.[startMatch[1]]) : 0;
   const frameworkPort = declaredPort ? 0 : managedWebDefaultPort(dependencies);
-  const port = declaredPort || (frameworkPort && await localPortAvailable(frameworkPort) ? frameworkPort : await openLocalPort());
-  const portBasis = declaredPort ? "repository-declared" : port === frameworkPort ? "framework-default" : "available-ephemeral";
+  // An explicit `.tapp/project.json` web.port pin outranks everything: backends with a CORS
+  // allowlist need the managed origin to be exact, so a busy pinned port is a hard error, never
+  // a silent ephemeral fallback that would resurface as fetch/CORS findings.
+  const { readProjectConfig } = await import("./project-config.js");
+  const projectConfig = readProjectConfig(root);
+  if (projectConfig.exists && projectConfig.errors.length) return { error: `Invalid ${projectConfig.relativePath}: ${projectConfig.errors.join("; ")}` };
+  const pinnedPort = Number(projectConfig.config?.web?.port) || 0;
+  if (pinnedPort && declaredPort && pinnedPort !== declaredPort) {
+    return { error: `${projectConfig.relativePath} pins the managed web port to ${pinnedPort}, but the repository start script declares port ${declaredPort}`, details: { remediation: "Align web.port with the start script (or remove one of them) so the managed origin is unambiguous." } };
+  }
+  if (pinnedPort && !(await localPortAvailable(pinnedPort))) {
+    return { error: `${projectConfig.relativePath} pins the managed web port to ${pinnedPort}, but that port is already in use`, details: { remediation: "Stop the process holding the port, change web.port, or provide an already-running owned --url." } };
+  }
+  const port = pinnedPort || declaredPort || (frameworkPort && await localPortAvailable(frameworkPort) ? frameworkPort : await openLocalPort());
+  const portBasis = pinnedPort ? "project-pinned" : declaredPort ? "repository-declared" : port === frameworkPort ? "framework-default" : "available-ephemeral";
   let command = "npm";
   let startArgs;
   let startDir = projectDir;
@@ -2846,7 +2869,8 @@ server.setRequestHandler(ListToolsRequestSchema, async () => ({
         "Opt-in AI assertion: {assert_ai: '<claim about the current screen>'} (judged host-side; needs a key; " +
         "skipped otherwise). Pass a flow inline via `flow`, or a repo-relative `flowPath` to a .yml/.json. " +
         "A failed assertion fails the flow and is reported like a QA finding. $TEST_EMAIL/$TEST_PASSWORD and any " +
-        "flow `vars` are substituted; pass testEmail/testPassword for real credential values.",
+        "flow `vars` are substituted; pass testEmail/testPassword for real credential values, or `actor` to " +
+        "resolve them from a configured actor's environment-variable bindings.",
       inputSchema: {
         type: "object",
         properties: {
@@ -2865,6 +2889,7 @@ server.setRequestHandler(ListToolsRequestSchema, async () => ({
           androidSerial: { type: "string", description: "Android adb device serial." },
           testEmail: { type: "string", description: "Value for $TEST_EMAIL" },
           testPassword: { type: "string", description: "Value for $TEST_PASSWORD" },
+          actor: { type: "string", description: "Configured actor name; resolves $TEST_EMAIL/$TEST_PASSWORD from the actor's env-var bindings (explicit testEmail/testPassword win)" },
         },
       },
     },
@@ -3040,7 +3065,10 @@ server.setRequestHandler(ListToolsRequestSchema, async () => ({
         "launch per action. Returns the initial screen {screenTitle, elements[]}. Drive it with " +
         "tapp_focus for any named destination (source + shortest observed route), then tapp_session_act only for remaining actions; finish with tapp_session_end. " +
         "For a focused user request, ALWAYS pass it in `focus` so the session reaches that surface before returning. Only one session at a time. Starts from a " +
-        "fresh launch. Use appLaunchArgs/appLaunchEnv for apps that need a backend override or login bypass. " +
+        "fresh COLD launch (terminate + relaunch, for a deterministic starting screen): persisted data such as Keychain credentials survives, but the app opens on its " +
+        "launch screen, NOT a resumed foreground state — an app that gates each cold start behind sign-in will show its login wall, so plan a `login` act (or a login bypass " +
+        "launch argument) as the first step. CLI `tapp tree`/`tapp screenshot` warm-resume the currently foregrounded app instead, which is why they can look signed-in when a session does not. " +
+        "Use appLaunchArgs/appLaunchEnv for apps that need a backend override or login bypass. " +
         "When you reach a screen with input fields and don't have values for them, ASK THE USER what to type " +
         "(offer defaults/skip) before typing — the session does not prompt on its own.",
       inputSchema: {
@@ -3091,8 +3119,10 @@ server.setRequestHandler(ListToolsRequestSchema, async () => ({
         "on refocus, so step-by-step login flows lose the password), 'tap' (by `id` = accessibility identifier " +
         "or visible/partial label or placeholder, or by `x`/`y` coordinates), 'type' (`text`, optional `id` to " +
         "target a field — always REPLACES the field's content), 'swipe' (`direction`), 'back', 'wait' (block " +
-        "until an element with `id`/`text` appears, up to `timeoutMs`), 'tree' (re-inspect without acting), " +
-        "'screenshot'. Returns {status, screenTitle, elements[]}; status 'not_found'/'timeout'/'still_on_login' " +
+        "until an element with `id`/`text` appears, up to `timeoutMs`), 'tree' (re-inspect without acting; " +
+        "pass `full: true` for every raw element with frames — the default projection dedupes and caps at " +
+        "160 elements, so grep-style checks against it can miss text that IS on screen), " +
+        "'screenshot'. Returns {status, screenTitle, elements[], durationMs}; status 'not_found'/'timeout'/'still_on_login' " +
         "etc. with a `detail` explaining login failures.",
       inputSchema: {
         type: "object",
@@ -3107,6 +3137,7 @@ server.setRequestHandler(ListToolsRequestSchema, async () => ({
           text: { type: "string", description: "Text to type, or the label/text to wait for" },
           direction: { type: "string", enum: ["up", "down", "left", "right"], description: "Swipe direction" },
           timeoutMs: { type: "integer", minimum: 500, maximum: 60000, default: 5000, description: "For 'wait': how long to poll for the element" },
+          full: { type: "boolean", default: false, description: "For 'tree': return the complete raw element list (frames included) instead of the deduplicated agent-facing projection capped at 160 elements" },
           label: { type: "string", description: "Optional screenshot label" },
         },
         required: ["action"],
@@ -3927,6 +3958,22 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
     const runEnv = { ...process.env, FLOW_LOG: flowLog, TAPP_FLOW_EVIDENCE_DIR: evidenceDir };
     if (isNonEmptyString(args.testEmail)) runEnv.OCQA_TEST_EMAIL = args.testEmail.trim();
     if (isNonEmptyString(args.testPassword)) runEnv.OCQA_TEST_PASSWORD = args.testPassword.trim();
+    if (isNonEmptyString(args.actor)) {
+      const { readProjectConfig } = await import("./project-config.js");
+      const loaded = readProjectConfig(workspaceRoot);
+      if (loaded.errors.length) return errorResult(`Invalid ${loaded.relativePath}: ${loaded.errors.join("; ")}`);
+      const actor = loaded.config.actors?.[args.actor.trim()];
+      if (!actor) return errorResult(`Actor '${args.actor.trim()}' is not configured in ${loaded.relativePath}`, { hint: "Configure it with tapp_actor_set / tapp actor set first" });
+      // Actors store env-var NAMES only; resolve values here. Explicit testEmail/testPassword win.
+      for (const [credential, argName, envKey] of [["email", "testEmail", "OCQA_TEST_EMAIL"], ["password", "testPassword", "OCQA_TEST_PASSWORD"]]) {
+        if (isNonEmptyString(args[argName])) continue;
+        const binding = actor.credentials?.[credential];
+        if (!binding) continue;
+        const value = process.env[binding.env];
+        if (!value) return errorResult(`Actor '${args.actor.trim()}' binds ${credential} to $${binding.env}, but that environment variable is not set in the MCP server's environment.`);
+        runEnv[envKey] = value;
+      }
+    }
     let run = { stdout: "", stderr: "", code: 0 };
     if (platform === "web") {
       try {
@@ -4376,7 +4423,7 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
     const detailNote = !ok && r.detail ? ` — ${r.detail}` : "";
     const head = `${did} — ${ok ? "ok" : `⚠️ ${r.status}${detailNote}`} → now on **${r.screenTitle || "Unknown"}**`;
     const rec = typeof r.recordedSteps === "number" ? `\n\n🔴 Recording — ${r.recordedSteps} step(s). \`tapp_flow_save\` to keep it as a test.` : "";
-    const screen = agentScreenProjection(r);
+    const screen = agentScreenProjection(r, { full: action === "tree" && args.full === true });
     const result = richResult(head + "\n\n" + formatScreen(screen.screenTitle, screen.elements) + rec, { ...r, ...screen });
     if (!ok) result.isError = true;
     return result;

package/mcp-server/src/project-config.js

@@ -22,7 +22,7 @@ function cleanActor(actor) {
 export function validateProjectConfig(config) {
   const errors = [];
   if (!config || typeof config !== "object" || Array.isArray(config)) return ["Project configuration must be an object"];
-  for (const key of Object.keys(config)) if (!["kind", "schemaVersion", "actors", "lifecycle", "provenance"].includes(key)) errors.push(`unsupported project configuration field '${key}'`);
+  for (const key of Object.keys(config)) if (!["kind", "schemaVersion", "actors", "lifecycle", "provenance", "web"].includes(key)) errors.push(`unsupported project configuration field '${key}'`);
   if (config.kind !== "tapp-project-config") errors.push("kind must be 'tapp-project-config'");
   if (config.schemaVersion !== 1) errors.push("schemaVersion must be 1");
   if (config.actors !== undefined && (!config.actors || typeof config.actors !== "object" || Array.isArray(config.actors))) errors.push("actors must be an object");
@@ -43,6 +43,13 @@ export function validateProjectConfig(config) {
       for (const key of Object.keys(binding || {})) if (key !== "env") errors.push(`actor '${name}' credential '${credential}' may contain only env; credential values are forbidden`);
     }
   }
+  if (config.web !== undefined) {
+    if (!config.web || typeof config.web !== "object" || Array.isArray(config.web)) errors.push("web must be an object");
+    else {
+      for (const key of Object.keys(config.web)) if (key !== "port") errors.push(`web has unsupported field '${key}'; the managed web host is always 127.0.0.1`);
+      if (config.web.port !== undefined && (!Number.isInteger(config.web.port) || config.web.port < 1 || config.web.port > 65535)) errors.push("web.port must be an integer between 1 and 65535");
+    }
+  }
   if (config.lifecycle !== undefined && (!config.lifecycle || typeof config.lifecycle !== "object" || Array.isArray(config.lifecycle))) errors.push("lifecycle must be an object");
   for (const key of Object.keys(config.lifecycle || {})) if (!["setup", "teardown"].includes(key)) errors.push(`lifecycle has unsupported phase '${key}'`);
   if (config.provenance !== undefined && (!config.provenance || typeof config.provenance !== "object" || Array.isArray(config.provenance))) errors.push("provenance must be an object");

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aarwitz/tapp",
-  "version": "0.17.3",
+  "version": "0.17.4",
   "mcpName": "io.github.aarwitz/tapp",
   "description": "Let coding agents verify UI changes on real iOS, Android, and web surfaces, then enforce reviewed proof in deterministic CI.",
   "license": "MIT",

package/skills/tapp/references/commands.md

@@ -48,6 +48,11 @@ npx -y @aarwitz/tapp@latest open https://example.com --tap "Not now" --wait-for
 - `tapp_focus`: source-locate a named screen/control and execute the shortest observed route in the active session.
 - `tapp_ui_tree` / `tapp_screenshot`: inspect the current real surface.
 - `tapp_session_start` → `tapp_session_act` → `tapp_session_end`: drive one persistent journey.
+  Sessions begin from a **cold** launch: persisted data survives, but an app that gates each cold
+  start behind sign-in shows its login wall first — make `{action:"login"}` the first act (plain
+  `tapp_ui_tree`/`tapp_screenshot` warm-resume the foregrounded app, so they can look signed-in
+  when a fresh session does not). A `tree` act accepts `full:true` for the complete raw element
+  list when the default ≤160-element projection might omit the text you are checking for.
 - `tapp_explore`: autonomous iOS, Android, or web exploration; observation only.
 - `tapp_flow_save` / `tapp_flow_run`: save a driven journey and replay it deterministically.
 - `tapp_release_contract`: validate, compile, or run a reviewed business guarantee.
@@ -82,6 +87,10 @@ For autonomous exploration, pass test-only values when authorized:
 - CLI: `--email`, `--password`, repeated `--launch-arg`, and JSON `--launch-env`.
 - MCP: `testEmail`, `testPassword`, `inputOverrides`, `appLaunchArgs`, `appLaunchEnv`, or explicit
   `loginSteps`.
+- Reusable: `tapp actor set NAME --email-env ENV --password-env ENV` stores env-var **names**
+  (never values); Flow replay resolves them via `tapp flow run FILE --actor NAME` or
+  `tapp_flow_run {actor:"NAME"}`. Re-running `actor set` on an existing actor requires
+  `--replace` — idempotent setup scripts must include it.
 
 Do not persist secrets in `.tapp/`. If the result reports input fields and no values were supplied,
 ask the user rather than pretending the explored surface was complete.