verikun 0.16.0 → 0.18.0

package/README.md

@@ -104,7 +104,7 @@ vk screenshot                   # -> ./.verikun/screen.png
 | Command | Description |
 |---|---|
 | `ai <file> [--model m] [--max-cost-usd n] [--timeout dur] [--cost-override in/out] [--effort e] [--package pkg] [--app-build id] [--server url] [--show-plan] [--recompile] [--json]` | Run a plain-English test: compile it to a deterministic plan once, replay it model-free, and self-heal failures via the model. Needs `ANTHROPIC_API_KEY` or `OPENAI_API_KEY` (per model), or no key with `--model codex-cli` / `cursor-cli` (a logged-in `codex` / `cursor-agent` CLI). See [AI](#ai--natural-language-tests). |
-| `suite <dir> [--app <id>] [--name n] [--server url] [--json]` (+ all `ai` flags) | Run every `*.md` in `<dir>` as one sequential suite with an overview report and a non-zero exit on failure — the CI gate. See [Suites](#suites--run-a-directory-of-tests). |
+| `suite <dir> [--app <id>] [--name n] [--retries n] [--server url] [--json]` (+ all `ai` flags) | Run every `*.md` in `<dir>` as one sequential suite with an overview report and a non-zero exit on failure — the CI gate. See [Suites](#suites--run-a-directory-of-tests). |
 
 ### Remote
 | Command | Description |
@@ -122,7 +122,7 @@ vk screenshot                   # -> ./.verikun/screen.png
 |---|---|
 | `run start [name] [--force]` | Begin a named run. One auto-starts on the first action if you don't. |
 | `run status` | Show the active run and its recorded steps. |
-| `run archive [name]` | Write JUnit + HTML report to `./.verikun/runs/<id>/`; exits non-zero if any step failed. |
+| `run archive [name] [--no-logs]` | Write JUnit + HTML report to `./.verikun/runs/<id>/`; exits non-zero if any step failed. Captures `artifacts/logcat.txt` by default. |
 | `run clear` | Discard the active run without a report. |
 
 ## Test runs & reports
@@ -130,10 +130,11 @@ vk screenshot                   # -> ./.verikun/screen.png
 Actions are recorded into a **test run** — one auto-starts on the first action
 (set `VERIKUN_NO_RUN=1` to disable). Every command becomes a step with its
 timing, the selector + identifier it resolved through, and pass/fail; a failing
-step also captures a screenshot **and** the UI hierarchy of the page. When a
-step fails you can additionally run `vk log <package>` to pull the device logs —
-that step records the logs **into the same run**, so the crash trace shows up in
-the report alongside the failure.
+step also captures a screenshot **and** the UI hierarchy of the page. Device
+logs are captured automatically at archive time into `artifacts/logcat.txt`
+(session-scoped when possible). You can still run `vk log <package>` mid-run to
+pull a snapshot into a step for the report; `--no-logs` / `VERIKUN_NO_LOGS`
+skips the archive dump on green runs (failures still capture).
 
 `vk run archive` finalizes the run into `./.verikun/runs/<id>/`:
 
@@ -141,8 +142,11 @@ the report alongside the failure.
   for failed assertions, `<error>` for environment errors, and the resolved
   identifier in `<system-out>`. Drops straight into CI.
 - **`report.html`** — a self-contained report: every step, the identifiers used,
-  any screenshots taken, the screenshot + hierarchy of any failed page, and any
-  device logs captured via `vk log`.
+  any screenshots taken, the screenshot + hierarchy of any failed page, a link to
+  the full device log, an app-scoped log accordion when the run launched an app,
+  and any per-step logs from `vk log`.
+- **`artifacts/logcat.txt`** — full device log for the run window (default).
+- **`artifacts/logcat-app.txt`** — app-scoped log (when a package/bundle was launched).
 - **`run.json`** — the raw recording.
 
 `vk run archive` exits non-zero when the run contained failures, so the same
@@ -287,12 +291,29 @@ vk suite tests/ --app com.example.app --server "$VERIKUN_SERVER"   # remote devi
 - **Each test is a full `vk ai` run** — plan cache, self-healing, cost budget, and
   its own archived JUnit + HTML report under `./.verikun/runs/<id>/`. A test that
   fails (or errors) doesn't stop the suite; the rest still run.
-- **But a broken *environment* does stop it.** If a test dies from an environment
-  error (exit 3 — tool gone, device unplugged, server unreachable), the toolchain is
-  re-probed; only if it is *still* broken does the suite abort. That re-probe matters:
-  a transient `uiautomator` dump failure also exits 3, and shouldn't vaporize a
-  20-test run. Continuing on a genuinely dead box just produces one identical red row
-  per remaining test — noise that reads exactly like a mass regression.
+- **`--retries N` recovers from flakes.** A failed test is re-run up to N times
+  (default `0` — opt-in, so CI cost/time stay predictable). If a later attempt
+  passes, the suite exits `0` and the flake is a **warning**, not a hard failure.
+  Failed attempt archives stay linked from the suite overview (`attempts` on the
+  test row + a `warnings` list on the manifest), so flakiness remains visible.
+  Cost and duration sum across attempts.
+- **What earns a retry:** anything that might come out differently — a flaky
+  selector, a wedged app, and **a broken environment**, including a `vk server`
+  connection dropping mid-suite. The bias is intentional: an attempt costs one test,
+  giving up costs the whole suite plus a human rerunning it. Environment retries wait
+  a little longer each time (an outage that survives the health probe usually needs
+  seconds, not milliseconds) and each one lands in `warnings`, so riding out a wobble
+  is never silent. Exactly two failures are never retried, because a rerun cannot
+  change them: a **budget abort** (each attempt gets its own ceiling, so it would just
+  re-abort having spent twice) and a **usage error** (exit `2` — an unreadable test
+  file, a payload the server refuses).
+- **But a broken *environment* does stop it, once the attempts are gone.** If a test
+  dies from an environment error (exit 3 — tool gone, device unplugged, server
+  unreachable), the toolchain is re-probed; only if it is *still* broken **and** no
+  retries remain does the suite abort. That re-probe matters: a transient
+  `uiautomator` dump failure also exits 3, and shouldn't vaporize a 20-test run.
+  Continuing on a genuinely dead box just produces one identical red row per remaining
+  test — noise that reads exactly like a mass regression.
 - **The suite writes an overview** to `./.verikun/suites/<id>/`:
   - **`index.json`** — a stable, `schemaVersion`ed manifest: per-test pass/fail,
     steps, model repairs, cost, duration, and the run id, plus suite totals. This
@@ -300,17 +321,21 @@ vk suite tests/ --app com.example.app --server "$VERIKUN_SERVER"   # remote devi
     it (see the [CI recipe](#ci-recipe)) instead of verikun growing upload plugins.
     On an abort it also carries `aborted: {reason, notRun}`; the not-run tests get
     **no rows and no place in `totals`**, so `passed + failed === tests` still holds
-    and nothing downstream mistakes a skipped test for a regression.
+    and nothing downstream mistakes a skipped test for a regression. Retried flakes
+    add `flaky` / `attempts` on the test row and suite-level `warnings` (additive;
+    `schemaVersion` stays `1`).
   - **`index.html`** — a summary page linking every test's `report.html`, with a
-    banner naming the not-run tests when the suite aborted.
-- **Exit code is the CI gate:** `0` all green · `1` a test failed · `2` bad/empty
-  directory · `3` environment (the provider or the device toolchain is unavailable,
-  or the box broke mid-run). The `1`-vs-`3` split is the point: `1` is a regression
-  to investigate, `3` is a machine to fix. All `ai` flags (`--model`,
-  `--max-cost-usd`, `--timeout`, …) apply to every test; both the provider
-  (`ANTHROPIC_API_KEY` / `OPENAI_API_KEY`, or the `codex` / `cursor-agent` CLI for
-  `--model codex-cli` / `cursor-cli`) **and** the device toolchain (`adb` / `idb` +
-  a resolvable device) are checked up front, before anything is compiled.
+    banner naming the not-run tests when the suite aborted, and a warnings banner
+    when a flake recovered on retry (prior failed attempts stay linked).
+- **Exit code is the CI gate:** `0` all green (including flakes that recovered with
+  `--retries`) · `1` a test failed · `2` bad/empty directory · `3` environment (the
+  provider or the device toolchain is unavailable, or the box broke mid-run). The
+  `1`-vs-`3` split is the point: `1` is a regression to investigate, `3` is a
+  machine to fix. All `ai` flags (`--model`, `--max-cost-usd`, `--timeout`, …)
+  apply to every test; both the provider (`ANTHROPIC_API_KEY` / `OPENAI_API_KEY`,
+  or the `codex` / `cursor-agent` CLI for `--model codex-cli` / `cursor-cli`)
+  **and** the device toolchain (`adb` / `idb` + a resolvable device) are checked up
+  front, before anything is compiled.
 
 ## Remote devices — `vk server`
 

package/dist/agent/remote.js

@@ -112,7 +112,7 @@ function createRemoteBackend(opts, health) {
     const execRaw = async (req, record) => {
         const res = await t.postJson('/v1/exec', req, EXEC_TIMEOUT_MS);
         if (record && res.step)
-            opts.onStep?.(res.step, decodeArtifacts(res.artifacts));
+            opts.onStep?.(res.step, decodeArtifacts(res.artifacts), res.logStart);
         return { code: res.code, error: res.error ? (0, rpc_1.rebuildError)(res.error) : undefined };
     };
     return {
@@ -121,6 +121,10 @@ function createRemoteBackend(opts, health) {
             const res = await t.postJson('/v1/elements', {}, ELEMENTS_TIMEOUT_MS);
             return res.elements;
         },
+        async getLogs(logOpts = {}) {
+            const res = await t.postJson('/v1/logs', logOpts, ELEMENTS_TIMEOUT_MS);
+            return res.logs ?? '';
+        },
         async install(appPath) {
             // v1 remote installs are single-file uploads; the extension is the only thing
             // the client tells the server about the artifact (never a path).

package/dist/args.js

@@ -42,6 +42,7 @@ const BOOLEAN = new Set([
     'recompile',
     'no-cache',
     'no-restart',
+    'no-logs',
     'allow-install',
     'allow-unsafe-anonymous',
     // Selector state modifiers (STATE_ATTRS in ui/selector.ts) and their negations.

package/dist/cli.js

@@ -975,16 +975,47 @@ function cmdRun(positionals, flags, platform, device) {
         case 'archive':
         case 'finish':
         case 'save': {
-            const { dir, xmlPath, htmlPath, state } = run_1.Recorder.archive(positionals[1]);
+            const noLogs = (0, args_1.flagBool)(flags, 'no-logs');
+            // Best-effort: archive-time log capture needs a device. Prefer the run's
+            // bound serial/platform so a multi-device host hits the right one. A
+            // missing/broken toolchain must not prevent sealing the report.
+            let fetchLogs;
+            const active = run_1.Recorder.status();
+            const hasFailures = !!active?.steps.some((s) => s.status !== 'passed');
+            if ((0, run_1.wantsArchiveLogs)(hasFailures, noLogs)) {
+                try {
+                    const plat = active?.platform === 'ios' || active?.platform === 'android'
+                        ? active.platform
+                        : platform;
+                    const driver = (0, drivers_1.getDriver)(plat, active?.device || device);
+                    fetchLogs = (opts) => driver.getLogs(opts);
+                }
+                catch (e) {
+                    (0, output_1.err)(`[verikun] archive log capture unavailable (${e.message})`);
+                }
+            }
+            const { dir, xmlPath, htmlPath, state } = run_1.Recorder.archive(positionals[1], { noLogs, fetchLogs });
             const { passed, failed } = tally(state.steps);
             if (asJson) {
-                (0, output_1.json)({ archived: dir, report: htmlPath, junit: xmlPath, steps: state.steps.length, passed, failed });
+                (0, output_1.json)({
+                    archived: dir,
+                    report: htmlPath,
+                    junit: xmlPath,
+                    steps: state.steps.length,
+                    passed,
+                    failed,
+                    ...(state.logFile ? { logFile: state.logFile } : {}),
+                });
             }
             else {
                 (0, output_1.out)(dir); // primary result: the archived run directory
                 (0, output_1.err)(`archived '${state.name}': ${state.steps.length} step(s), ${passed} passed, ${failed} failed/error`);
                 (0, output_1.err)(`  JUnit: ${xmlPath}`);
                 (0, output_1.err)(`  HTML:  ${htmlPath}`);
+                if (state.logFile)
+                    (0, output_1.err)(`  Logs:  ${(0, node_path_1.join)(dir, state.logFile)}`);
+                if (state.appLogFile)
+                    (0, output_1.err)(`  App:   ${(0, node_path_1.join)(dir, state.appLogFile)}`);
             }
             // Exit non-zero when the run contained failures, so CI can gate on it.
             return failed > 0 ? 1 : 0;
@@ -1280,6 +1311,7 @@ async function resolveBackend(platform, device, flags) {
             backend: {
                 exec: (command, positionals, f) => executeOutcome(command, positionals, f, driver),
                 getElements: () => driver.getElements(),
+                getLogs: (opts) => driver.getLogs(opts),
                 install: (appPath) => driver.install(appPath),
                 reset: (appId) => {
                     assertSafeAppId(appId);
@@ -1300,8 +1332,9 @@ async function resolveBackend(platform, device, flags) {
         url: server,
         authKey: (0, args_1.flagStr)(flags, 'auth-key') || process.env.VERIKUN_SERVER_AUTH_KEY || undefined,
         // Each remote step is spliced into the local active run so the archived report
-        // is identical to a local run's.
-        onStep: (step, artifacts) => run_1.Recorder.appendForeignStep(step, artifacts, runCtx),
+        // is identical to a local run's. logStart travels from the server's device clock
+        // so archive-time / vk log scoping works without a local driver.
+        onStep: (step, artifacts, logStart) => run_1.Recorder.appendForeignStep(step, artifacts, { ...runCtx, logStart }),
     };
     const health = await (0, remote_1.pingServer)(opts); // fails fast (exit 3) on a bad URL or key
     runCtx = { platform: health.platform, device: health.serial };
@@ -1325,6 +1358,45 @@ async function resolveBackend(platform, device, flags) {
         remote: { url: server, version: health.version },
     };
 }
+/**
+ * Best-effort archive-time device-log capture via an ExecBackend (local driver or
+ * remote `/v1/logs`). Writes `artifacts/logcat.txt` onto the active run so a later
+ * `Recorder.archive()` finds `logFile` already set. Never throws — a gone device
+ * must not prevent sealing the report.
+ */
+async function prefetchArchiveLogs(backend, noLogs = false) {
+    const state = run_1.Recorder.status();
+    if (!state)
+        return;
+    const hasFailures = state.steps.some((s) => s.status !== 'passed');
+    if (!(0, run_1.wantsArchiveLogs)(hasFailures, noLogs))
+        return;
+    if (!backend.getLogs)
+        return;
+    // Skip only when both artifacts are already attached (e.g. a prior prefetch).
+    if (state.logFile && state.appLogFile)
+        return;
+    const window = (0, run_1.archiveLogWindow)(state);
+    try {
+        const full = state.logFile ? undefined : await backend.getLogs(window);
+        const appId = (0, run_1.inferRunAppId)(state);
+        let app;
+        if (appId && !state.appLogFile) {
+            try {
+                app = await backend.getLogs({ ...window, appId, scopedOnly: true });
+            }
+            catch (e) {
+                (0, output_1.err)(`[verikun] could not capture archive app logs (${e.message})`);
+            }
+        }
+        if (full !== undefined || (app !== undefined && app !== '')) {
+            run_1.Recorder.attachArchiveLogs(full, app);
+        }
+    }
+    catch (e) {
+        (0, output_1.err)(`[verikun] could not capture archive device logs (${e.message})`);
+    }
+}
 /**
  * Run one natural-language test through a backend and return DATA — no stdout
  * writes (stdout stays the caller's one result; progress streams to stderr).
@@ -1365,10 +1437,14 @@ async function runAiTest(file, opts, backend, platform, device) {
     if (existing && existing.steps.length > 0) {
         // Seal the pre-existing run into the archive instead of letting start(force=true)
         // discard it — a manual in-progress run should never be silently lost.
+        await prefetchArchiveLogs(backend);
         const sealed = run_1.Recorder.archive();
         (0, output_1.err)(`[ai] archived the active run ('${existing.name}', ${existing.steps.length} step(s)) → ${sealed.dir}`);
     }
     const started = run_1.Recorder.start(`ai: ${(0, node_path_1.basename)(file)}`, platform, device, true);
+    // Prefer --package when the prose never launches by id (rare but possible).
+    if (opts.pkg)
+        run_1.Recorder.annotateRun({ appId: opts.pkg });
     // Suppress per-step `out()` so stdout stays the one final result; progress -> stderr.
     const prevQuiet = (0, output_1.setOutputQuiet)(true);
     let result;
@@ -1395,6 +1471,7 @@ async function runAiTest(file, opts, backend, platform, device) {
         // to roll over. Then let the error map to an exit code as usual.
         run_1.Recorder.annotateRun({ ai: { ok: false, cost: cost.summaryLine(), modelRepairs: 0, improvements: [] } });
         try {
+            await prefetchArchiveLogs(backend);
             run_1.Recorder.archive();
         }
         catch (sealErr) {
@@ -1422,6 +1499,7 @@ async function runAiTest(file, opts, backend, platform, device) {
     run_1.Recorder.annotateRun({
         ai: { ok: result.ok, cost: costLine, modelRepairs: result.modelRepairs, improvements: result.improvements },
     });
+    await prefetchArchiveLogs(backend);
     const { dir, xmlPath, htmlPath, state } = run_1.Recorder.archive();
     const status = result.ok
         ? 'PASS'
@@ -1537,7 +1615,7 @@ async function cmdInstall(positionals, flags) {
 async function cmdSuiteEntry(positionals, flags) {
     const dirArg = positionals[0];
     if (!dirArg)
-        throw new errors_1.CliError('Usage: verikun suite <dir> [--app <id>] [--server url] [--name n] [--json]', 2);
+        throw new errors_1.CliError('Usage: verikun suite <dir> [--app <id>] [--server url] [--name n] [--retries n] [--json]', 2);
     const opts = parseAiOptions(flags);
     // Pre-flight the provider BEFORE touching any device/server: every test needs it
     // to compile (on a cache miss) or to repair at runtime.
@@ -1708,11 +1786,23 @@ async function executeForServer(command, positionals, flags, driver, platform) {
     catch {
         /* surfaced by the command handler below */
     }
+    // Sample the device clock up front so the caller's run can set logStart (the
+    // ephemeral recorder never persists RunState). Best-effort — empty/unavailable
+    // just means archive / vk log fall back to last-N.
+    let logStart;
+    try {
+        const t = driver.deviceTime();
+        if (t)
+            logStart = t;
+    }
+    catch {
+        /* device clock unavailable */
+    }
     const recorder = run_1.Recorder.beginEphemeralStep(command, positionals, flags, platform, serial);
     const ctx = { driver, platform, device: serial, positionals, flags, record: recorder };
     const outcome = await runRecorded(command, ctx, recorder, driver);
     const { step, artifacts } = recorder.takeEphemeral();
-    return { ...outcome, step, artifacts };
+    return { ...outcome, step, artifacts, ...(logStart ? { logStart } : {}) };
 }
 /**
  * Run one already-parsed command for the CLI / `batch`: dispatch meta-commands,
@@ -1849,14 +1939,19 @@ AI (run a natural-language test — compile once, replay model-free, self-heal)
                                       codex-cli | cursor-cli.
 
 SUITE (run a directory of natural-language tests as one gated suite)
-  suite <dir> [--app <id>] [--name n] [--json]  (+ all \`ai\` flags, incl. --server)
+  suite <dir> [--app <id>] [--name n] [--retries n] [--json]
+                                      (+ all \`ai\` flags, incl. --server)
                                       Run every *.md in <dir> (lexicographic order —
                                       prefix 01-, 02- to sequence; README.md skipped)
                                       through \`vk ai\`. With --app, app data is reset
-                                      between tests (iOS: force-stop). Writes a suite
-                                      overview to ./.verikun/suites/<id>/{index.json,
-                                      index.html} linking each test's report. Exits 1
-                                      if any test failed — the CI gate.
+                                      between tests (iOS: force-stop). --retries N
+                                      re-runs a failed test up to N times; a later
+                                      pass recovers the suite (exit 0) and surfaces a
+                                      warning, keeping failed-attempt evidence in the
+                                      report. Writes a suite overview to
+                                      ./.verikun/suites/<id>/{index.json, index.html}
+                                      linking each test's report. Exits 1 if any test
+                                      failed — the CI gate.
 
 SERVER (expose a locally-connected device to remote verikun clients)
   server [--bind addr] [--port n] [--auth-key k] [--allow-install]
@@ -1879,7 +1974,10 @@ ENVIRONMENT
 TEST RUNS (actions are recorded; a run auto-starts on first action)
   run start [name] [--force]          Begin a named run (else one starts implicitly)
   run status                          Show the active run, its device/session, and steps
-  run archive [name]                  Write JUnit + HTML report, move to ./.verikun/runs/<id>/
+  run archive [name] [--no-logs]      Write JUnit + HTML report, move to ./.verikun/runs/<id>/
+                                      Captures artifacts/logcat.txt by default (session-scoped);
+                                      --no-logs / VERIKUN_NO_LOGS skips on green runs (failures
+                                      still capture). Capture is best-effort and never blocks archive.
   run clear                           Discard the active run with no report
   An implicit run auto-closes (archives) and rolls over on a device change, a
   VERIKUN_SESSION change, or VERIKUN_RUN_IDLE_MIN minutes idle (default 30; 0 off).

package/dist/drivers/adb.js

@@ -356,15 +356,36 @@ class AdbDriver {
             args.push('-t', String(n));
         }
         if (opts.appId) {
-            // Scope to the app's process when it's alive. If pidof is empty the app
-            // isn't running (likely crashed) — fall through to a system-wide dump so
-            // the FATAL EXCEPTION, still in the crash buffer, is not missed.
-            const pid = this.shell(['pidof', opts.appId]).trim().split(/\s+/)[0];
-            if (pid)
-                args.push(`--pid=${pid}`);
+            // Prefer --uid: it survives process death/restart (crash traces stay under the
+            // package's uid). Fall back to --pid for a live process when uid isn't known.
+            // When neither works: vk log falls through to system-wide so a FATAL EXCEPTION
+            // isn't missed; archive accordion passes scopedOnly to keep the dump empty.
+            const uid = this.packageUid(opts.appId);
+            if (uid) {
+                args.push(`--uid=${uid}`);
+            }
+            else {
+                const pid = this.shell(['pidof', opts.appId]).trim().split(/\s+/)[0];
+                if (pid)
+                    args.push(`--pid=${pid}`);
+                else if (opts.scopedOnly)
+                    return '';
+            }
         }
         return this.shell(args, 15000);
     }
+    /** Android userId for an installed package, or '' if unknown. Used to scope
+     *  logcat across process restarts (unlike pidof, which only sees a live process). */
+    packageUid(appId) {
+        try {
+            const out = this.shell(['dumpsys', 'package', appId], 10000);
+            const m = /\buserId=(\d+)\b/.exec(out);
+            return m?.[1] ?? '';
+        }
+        catch {
+            return '';
+        }
+    }
     deviceTime() {
         // logcat's default timestamp is MM-DD HH:MM:SS.mmm in the device's LOCAL time.
         // Sample it from the device clock with a space-free format (so no device-shell

package/dist/drivers/ios.js

@@ -433,6 +433,11 @@ class IdbDriver {
             const proc = opts.appId.split('.').pop() || opts.appId;
             args.push('--predicate', `process CONTAINS "${proc}"`);
         }
+        else if (opts.scopedOnly) {
+            // scopedOnly without an appId can't mean anything on iOS — empty rather than
+            // a full dump (mirrors Android's "couldn't scope" behaviour).
+            return '';
+        }
         const out = (0, exec_1.runText)(XCRUN, args, { timeout: 20000 }).stdout;
         if (opts.since)
             return out; // the whole session window

package/dist/report.js

@@ -95,14 +95,20 @@ function toJUnitXml(run) {
         .join('\n');
     const suiteAttrs = `name="${xmlAttr(run.name)}" tests="${c.tests}" failures="${c.failures}" ` +
         `errors="${c.errors}" time="${suiteTime}" timestamp="${xmlAttr(run.startedAt)}"`;
+    const suiteExtras = [];
+    if (run.logFile)
+        suiteExtras.push(`device log: ${run.logFile}`);
+    if (run.ai) {
+        suiteExtras.push('vk ai: ' +
+            run.ai.cost +
+            (run.ai.improvements.length ? '\nSuggested improvements:\n' + run.ai.improvements.join('\n') : ''));
+    }
     return (`<?xml version="1.0" encoding="UTF-8"?>\n` +
         `<testsuites name="verikun" tests="${c.tests}" failures="${c.failures}" errors="${c.errors}" time="${suiteTime}">\n` +
         `<testsuite ${suiteAttrs}>\n` +
         `${cases}\n` +
-        (run.ai
-            ? `  <system-out>${xmlText('vk ai: ' +
-                run.ai.cost +
-                (run.ai.improvements.length ? '\nSuggested improvements:\n' + run.ai.improvements.join('\n') : ''))}</system-out>\n`
+        (suiteExtras.length
+            ? `  <system-out>${xmlText(suiteExtras.join('\n'))}</system-out>\n`
             : '') +
         `</testsuite>\n</testsuites>\n`);
 }
@@ -118,6 +124,7 @@ const STYLE = `
   .summary { display:flex; gap:8px; flex-wrap:wrap; align-items:center; margin-bottom: 20px; }
   .chip { font-weight:600; font-size:13px; padding:4px 10px; border-radius:999px; color:#fff; }
   .chip.pass{background:var(--pass)} .chip.fail{background:var(--fail)} .chip.err{background:var(--err)}
+  .chip.warn{background:var(--err)}
   .chip.muted{ background:#eaeef2; color:var(--muted); }
   ol.steps { list-style:none; margin:0; padding:0; }
   li.step { background:#fff; border:1px solid var(--line); border-left-width:4px; border-radius:8px; margin-bottom:10px; padding:12px 14px; }
@@ -135,6 +142,10 @@ const STYLE = `
   .msg.fail { color:var(--fail); }
   img.shot { display:block; margin-top:10px; max-width:300px; max-height:520px; border:1px solid var(--line); border-radius:6px; }
   details { margin-top:8px; }
+  details.run-log { margin-top:20px; background:#fff; border:1px solid var(--line); border-radius:8px; padding:10px 14px; }
+  details.run-log > summary { font-weight:600; color:#1f2328; }
+  details.run-log > summary a { font-weight:400; }
+  details.run-log pre { max-height:480px; }
   summary { cursor:pointer; color:var(--muted); font-size:13px; }
   pre { background:#0d1117; color:#e6edf3; padding:12px; border-radius:6px; overflow:auto; font-size:12px; line-height:1.45; max-height:360px; }
   .aibox { background:#fff; border:1px solid var(--line); border-radius:8px; padding:12px 14px; margin-bottom:20px; font-size:13px; }
@@ -206,19 +217,40 @@ const SUITE_STYLE = `
   table.tests td.num { text-align:right; font-variant-numeric:tabular-nums; white-space:nowrap; }
   table.tests a { color:inherit; }
   .fail-reason { color:var(--fail); font-size:12px; margin-top:2px; }
+  .flake-note { color:var(--err); font-size:12px; margin-top:2px; }
+  .attempts { margin-top:4px; font-size:12px; color:var(--muted); }
+  .attempts a { color:var(--fail); }
   .aborted { background:#fff4e5; border:1px solid #f0b429; border-radius:8px; padding:12px 14px; margin:0 0 14px; font-size:13px; }
   .aborted strong { color:#8a5300; }
   .aborted ul { margin:6px 0 0; padding-left:20px; color:var(--muted); }
+  .warnings { background:#fff8c5; border:1px solid #d4a72c; border-radius:8px; padding:12px 14px; margin:0 0 14px; font-size:13px; }
+  .warnings strong { color:#7d4e00; }
+  .warnings ul { margin:6px 0 0; padding-left:20px; color:var(--muted); }
 `;
+function suiteAttemptLinks(attempts, linkBase) {
+    const links = attempts
+        .map((a, i) => {
+        const label = `attempt ${i + 1}`;
+        if (!a.id)
+            return htmlEsc(label);
+        return `<a href="${htmlEsc(`${linkBase}runs/${encodeURIComponent(a.id)}/report.html`)}">${htmlEsc(label)}</a>`;
+    })
+        .join(', ');
+    return `<div class="attempts">prior failed: ${links}</div>`;
+}
 function suiteTestRow(t, linkBase) {
     // A test that errored before its run started (id '') has no report to link.
     const label = t.id
         ? `<a href="${htmlEsc(`${linkBase}runs/${encodeURIComponent(t.id)}/report.html`)}">${htmlEsc(t.name)}</a>`
         : htmlEsc(t.name);
     const failure = t.failure ? `<div class="fail-reason">${htmlEsc(t.failure)}</div>` : '';
+    const flake = t.flaky ? `<div class="flake-note">passed on retry (flake)</div>` : '';
+    const prior = t.attempts?.length ? suiteAttemptLinks(t.attempts, linkBase) : '';
+    const status = t.flaky ? 'FLAKY' : t.ok ? 'PASS' : 'FAIL';
+    const statusClass = t.ok ? 'passed' : 'failed';
     return `  <tr>
-    <td><span class="st ${t.ok ? 'passed' : 'failed'}">${t.ok ? 'PASS' : 'FAIL'}</span></td>
-    <td>${label}${failure}</td>
+    <td><span class="st ${statusClass}">${status}</span></td>
+    <td>${label}${flake}${failure}${prior}</td>
     <td class="num">${t.passedSteps}/${t.steps}${t.failedSteps ? ` (${t.failedSteps} failed)` : ''}</td>
     <td class="num">${t.modelRepairs || ''}</td>
     <td class="num">$${t.costUsd.toFixed(4)}</td>
@@ -233,9 +265,11 @@ function suiteTestRow(t, linkBase) {
 function toSuiteHtml(suite, opts = {}) {
     const linkBase = opts.linkBase ?? '../../';
     const t = suite.totals;
+    const flaky = suite.tests.filter((x) => x.flaky).length;
     const chips = [
         `<span class="chip pass">${t.passed} passed</span>`,
         t.failed ? `<span class="chip fail">${t.failed} failed</span>` : '',
+        flaky ? `<span class="chip warn">${flaky} flaky</span>` : '',
         suite.aborted ? `<span class="chip fail">ABORTED</span>` : '',
         `<span class="chip muted">${t.tests} tests &middot; ${t.steps} steps &middot; ${fmtDuration(t.durationMs)} &middot; $${t.costUsd.toFixed(4)}</span>`,
     ]
@@ -250,6 +284,13 @@ function toSuiteHtml(suite, opts = {}) {
 ${suite.aborted.notRun.length
             ? `    <ul>${suite.aborted.notRun.map((f) => `<li>${htmlEsc(f)} — not run</li>`).join('')}</ul>\n`
             : ''}  </div>
+`
+        : '';
+    const warningsBanner = suite.warnings?.length
+        ? `  <div class="warnings">
+    <strong>Warnings</strong>
+    <ul>${suite.warnings.map((w) => `<li>${htmlEsc(w)}</li>`).join('')}</ul>
+  </div>
 `
         : '';
     const metaBits = [
@@ -274,7 +315,7 @@ ${suite.aborted.notRun.length
   <div class="summary">
       ${chips}
   </div>
-${abortedBanner}  <table class="tests">
+${abortedBanner}${warningsBanner}  <table class="tests">
     <thead><tr><th></th><th>Test</th><th>Steps</th><th>Repairs</th><th>Cost</th><th>Duration</th></tr></thead>
     <tbody>
 ${suite.tests.map((x) => suiteTestRow(x, linkBase)).join('\n')}
@@ -285,7 +326,12 @@ ${suite.tests.map((x) => suiteTestRow(x, linkBase)).join('\n')}
 </html>
 `;
 }
-function toHtml(run) {
+/**
+ * @param opts.appLog  app-scoped logcat body for the bottom accordion (kept out
+ *                     of RunState / run.json — pass the file contents when writing
+ *                     report.html). The full device dump stays a meta-row file link.
+ */
+function toHtml(run, opts = {}) {
     const c = counts(run);
     const chips = [
         `<span class="chip pass">${c.passed} passed</span>`,
@@ -301,7 +347,20 @@ function toHtml(run) {
         `started ${htmlEsc(run.startedAt)}`,
         run.finishedAt ? `finished ${htmlEsc(run.finishedAt)}` : '',
         run.implicit ? 'implicit run' : '',
+        run.logFile ? `<a href="${htmlEsc(run.logFile)}">device log</a>` : '',
     ].filter(Boolean);
+    // Bottom accordion: app-scoped dump only. The noisy full device log stays a
+    // download via the meta link — same shape on every framework (package/uid scope).
+    const appLabel = run.appId ? ` for ${htmlEsc(run.appId)}` : '';
+    const appFileLink = run.appLogFile
+        ? ` (<a href="${htmlEsc(run.appLogFile)}">${htmlEsc(run.appLogFile)}</a>)`
+        : '';
+    const appLogPanel = opts.appLog !== undefined && opts.appLog !== ''
+        ? `\n  <details class="run-log">
+    <summary>App log${appLabel}${appFileLink}</summary>
+    <pre>${htmlEsc(opts.appLog)}</pre>
+  </details>`
+        : '';
     return `<!doctype html>
 <html lang="en">
 <head>
@@ -320,7 +379,7 @@ function toHtml(run) {
   ${run.ai ? aiPanelHtml(run.ai) : ''}
   <ol class="steps">
     ${run.steps.map(stepHtml).join('\n    ')}
-  </ol>
+  </ol>${appLogPanel}
 </div>
 </body>
 </html>

package/dist/run.js

@@ -1,10 +1,14 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.Recorder = void 0;
+exports.Recorder = exports.RUN_APP_LOG_ARTIFACT = exports.RUN_LOG_ARTIFACT = void 0;
 exports.isRecordable = isRecordable;
+exports.wantsArchiveLogs = wantsArchiveLogs;
+exports.archiveLogWindow = archiveLogWindow;
+exports.inferRunAppId = inferRunAppId;
 exports.runId = runId;
 exports.uniqueDir = uniqueDir;
 exports.rolloverReason = rolloverReason;
+exports.rolloverLogsSameDevice = rolloverLogsSameDevice;
 exports.stepName = stepName;
 const node_fs_1 = require("node:fs");
 const node_path_1 = require("node:path");
@@ -32,6 +36,58 @@ function isRecordable(command) {
 }
 const HIERARCHY_CAP = 24000; // chars of failure hierarchy kept inline in run.json
 const LOG_CAP = 50000; // chars of device logs kept inline in run.json (tail-kept)
+/** Max chars written to the archive-time `artifacts/logcat.txt` (tail-kept). Larger
+ *  than the per-step LOG_CAP because the file is not inlined into run.json. */
+const LOG_FILE_CAP = 512_000;
+/** When no session `logStart` marker exists, archive captures this many trailing lines. */
+const ARCHIVE_LOG_LINES = 5000;
+exports.RUN_LOG_ARTIFACT = 'artifacts/logcat.txt';
+exports.RUN_APP_LOG_ARTIFACT = 'artifacts/logcat-app.txt';
+const APP_LIFECYCLE = new Set(['launch', 'open', 'stop', 'clear']);
+/** Whether archive-time device-log capture should run. Default on; `--no-logs` /
+ *  `VERIKUN_NO_LOGS` opts out of *green* runs only — a failed run always captures
+ *  (best-effort), because that is when the log is load-bearing. Exported for tests. */
+function wantsArchiveLogs(hasFailures, noLogsFlag = false) {
+    if (hasFailures)
+        return true;
+    if (noLogsFlag)
+        return false;
+    if (process.env.VERIKUN_NO_LOGS)
+        return false;
+    return true;
+}
+/** Window passed to `getLogs` at archive time: session-scoped when `logStart` is
+ *  known, otherwise a bounded trailing dump. Exported for tests. */
+function archiveLogWindow(state) {
+    if (state.logStart)
+        return { since: state.logStart };
+    return { lines: ARCHIVE_LOG_LINES };
+}
+/** Package / bundle id the run exercised, if we can tell. Prefer an explicit
+ *  `state.appId` (set on launch/open/stop/clear); otherwise recover from those
+ *  steps' names. Exported for tests. */
+function inferRunAppId(state) {
+    if (state.appId && /^[A-Za-z0-9._-]+$/.test(state.appId))
+        return state.appId;
+    for (let i = state.steps.length - 1; i >= 0; i--) {
+        const s = state.steps[i];
+        if (!APP_LIFECYCLE.has(s.command))
+            continue;
+        // stepName is `launch com.foo` / `clear com.foo` (no flags in the name).
+        const m = /^(?:launch|open|stop|clear)\s+([A-Za-z0-9._-]+)\s*$/.exec(s.name);
+        if (m)
+            return m[1];
+    }
+    return undefined;
+}
+function truncateLogFile(text) {
+    if (text.length <= LOG_FILE_CAP)
+        return text;
+    return '…(truncated)\n' + text.slice(-LOG_FILE_CAP);
+}
+function runHasFailures(state) {
+    return state.steps.some((s) => s.status !== 'passed');
+}
 // --- paths & persistence --------------------------------------------------
 const activeDir = () => (0, node_path_1.join)((0, output_1.artifactDir)(), 'run');
 const archiveBase = () => (0, node_path_1.join)((0, output_1.artifactDir)(), 'runs');
@@ -123,6 +179,14 @@ function rolloverReason(state, serial, session) {
         return `idle for ${fmtAge(ageMs(state))} (>${idle}m)`;
     return null;
 }
+/** Whether the incoming step's driver is safe to use for archive-time logs of the
+ *  run being sealed. On a device-change rollover the driver is already pointed at
+ *  the *new* serial — pulling logcat from it would attribute the wrong device's
+ *  output to the old run. Idle/session rollover on the same device is fine.
+ *  Exported for tests. */
+function rolloverLogsSameDevice(runDevice, newSerial) {
+    return !runDevice || !newSerial || runDevice === newSerial;
+}
 function stepName(command, positionals, flags) {
     const p = positionals;
     const at = typeof flags['at'] === 'string' ? flags['at'] : undefined;
@@ -187,7 +251,13 @@ class Recorder {
             const reason = rolloverReason(state, serial, session);
             if (reason) {
                 try {
-                    const dest = Recorder.seal(state, dir);
+                    // Only reuse this step's driver when it still targets the run's device.
+                    // A device-change rollover must not write the new device's logcat into
+                    // the old run's archive (vk run archive binds a driver to state.device).
+                    const fetchLogs = driver && rolloverLogsSameDevice(state.device, serial)
+                        ? (opts) => driver.getLogs(opts)
+                        : undefined;
+                    const dest = Recorder.seal(state, dir, { fetchLogs });
                     (0, output_1.err)(`[verikun] previous run '${state.name}' (${state.steps.length} step(s)) auto-closed → ${dest} (${reason}); starting a fresh run`);
                     state = null;
                     rolledOver = true;
@@ -220,6 +290,11 @@ class Recorder {
             if (!state.session && session)
                 state.session = session;
         }
+        // Remember the app under test from lifecycle commands so archive can scope
+        // the accordion log without requiring a separate flag.
+        if (APP_LIFECYCLE.has(command) && positionals[0] && /^[A-Za-z0-9._-]+$/.test(positionals[0])) {
+            state.appId = positionals[0];
+        }
         // Anchor the log window at the session's first step (covers both implicit
         // creation and an explicit `vk run start`, which records no marker itself).
         // Best-effort — a missing marker just means `vk log` falls back to last-N.
@@ -386,18 +461,87 @@ class Recorder {
         this.state.updatedAt = nowIso();
         saveState(this.dir, this.state);
     }
-    /** Finalize a run: write reports next to it, then move it into ./.verikun/runs/<id>/. */
-    static seal(state, dir) {
+    /** Finalize a run: optionally capture device logs, write reports next to it, then
+     *  move it into ./.verikun/runs/<id>/. Log capture is best-effort — a device that
+     *  is gone (often why the run failed) must never prevent sealing the report. */
+    static seal(state, dir, opts = {}) {
+        Recorder.maybeCaptureArchiveLogs(state, dir, opts);
         state.finishedAt = nowIso();
         state.updatedAt = nowIso();
         saveState(dir, state);
         (0, node_fs_1.writeFileSync)((0, node_path_1.join)(dir, 'report.xml'), (0, report_1.toJUnitXml)(state));
-        (0, node_fs_1.writeFileSync)((0, node_path_1.join)(dir, 'report.html'), (0, report_1.toHtml)(state));
+        // Accordion embeds the *app*-scoped dump (readable); the full device dump stays
+        // a file link in the meta row. Missing file is fine — capture is best-effort.
+        let appLog;
+        if (state.appLogFile) {
+            try {
+                appLog = (0, node_fs_1.readFileSync)((0, node_path_1.join)(dir, state.appLogFile), 'utf8');
+            }
+            catch {
+                /* report still writes */
+            }
+        }
+        (0, node_fs_1.writeFileSync)((0, node_path_1.join)(dir, 'report.html'), (0, report_1.toHtml)(state, { appLog }));
         (0, node_fs_1.mkdirSync)(archiveBase(), { recursive: true });
         const dest = uniqueDir((0, node_path_1.join)(archiveBase(), state.id));
         (0, node_fs_1.renameSync)(dir, dest);
         return dest;
     }
+    /** Write archive-time log artifacts on the ACTIVE run without sealing. Used by
+     *  remote (`--server`) callers that fetch asynchronously before `archive()`.
+     *  Pass `full` / `app` only for the pieces you have — omitted sides are left alone. */
+    static attachArchiveLogs(full, app) {
+        const dir = activeDir();
+        const state = loadState(dir);
+        if (!state)
+            return;
+        if (full !== undefined) {
+            Recorder.writeArchiveLog(state, dir, full, exports.RUN_LOG_ARTIFACT, 'logFile');
+        }
+        if (app !== undefined && app !== '') {
+            if (!state.appId)
+                state.appId = inferRunAppId(state);
+            Recorder.writeArchiveLog(state, dir, app, exports.RUN_APP_LOG_ARTIFACT, 'appLogFile');
+        }
+        saveState(dir, state);
+    }
+    static writeArchiveLog(state, dir, text, rel, field) {
+        const body = truncateLogFile(text);
+        (0, node_fs_1.mkdirSync)((0, node_path_1.join)(dir, 'artifacts'), { recursive: true });
+        (0, node_fs_1.writeFileSync)((0, node_path_1.join)(dir, rel), body);
+        state[field] = rel;
+    }
+    static maybeCaptureArchiveLogs(state, dir, opts) {
+        if (!wantsArchiveLogs(runHasFailures(state), opts.noLogs))
+            return;
+        if (!opts.fetchLogs)
+            return;
+        const window = archiveLogWindow(state);
+        if (!state.logFile) {
+            try {
+                const text = opts.fetchLogs(window);
+                if (text !== undefined && text !== null) {
+                    Recorder.writeArchiveLog(state, dir, text, exports.RUN_LOG_ARTIFACT, 'logFile');
+                }
+            }
+            catch (e) {
+                (0, output_1.err)(`[verikun] could not capture archive device logs (${e.message})`);
+            }
+        }
+        const appId = inferRunAppId(state);
+        if (appId)
+            state.appId = appId;
+        if (appId && !state.appLogFile) {
+            try {
+                const text = opts.fetchLogs({ ...window, appId, scopedOnly: true });
+                if (text)
+                    Recorder.writeArchiveLog(state, dir, text, exports.RUN_APP_LOG_ARTIFACT, 'appLogFile');
+            }
+            catch (e) {
+                (0, output_1.err)(`[verikun] could not capture archive app logs (${e.message})`);
+            }
+        }
+    }
     /** One-line context summary for `vk run status`. */
     static contextLine(state) {
         const bits = [];
@@ -428,6 +572,8 @@ class Recorder {
      * a local one. Re-indexes the step, writes its artifact buffers under the new
      * index, and rewrites the step's artifact references to match. Creates an
      * implicit run first if none is active (parity with beginStep's auto-start).
+     * `ctx.logStart` (device-clock marker from the server) is recorded on the first
+     * splice so archive-time / `vk log` scoping works the same as a local run.
      */
     static appendForeignStep(step, artifacts = {}, ctx = {}) {
         if (process.env.VERIKUN_NO_RUN)
@@ -448,6 +594,17 @@ class Recorder {
             };
             (0, output_1.err)('[verikun] recording test run (implicit) — archive: `vk run archive` · discard: `vk run clear`');
         }
+        // Anchor the log window from the server's device clock (beginEphemeralStep
+        // never persists state, so the marker has to travel on the wire).
+        if (!state.logStart && ctx.logStart)
+            state.logStart = ctx.logStart;
+        // Mirror local beginStep: remember the app from lifecycle steps so archive
+        // can scope the accordion without a local launch having set state.appId.
+        if (APP_LIFECYCLE.has(step.command)) {
+            const m = /^(?:launch|open|stop|clear)\s+([A-Za-z0-9._-]+)\s*$/.exec(step.name);
+            if (m)
+                state.appId = m[1];
+        }
         const index = state.steps.length;
         const spliced = { ...step, index };
         for (const [rel, buf] of Object.entries(artifacts)) {
@@ -521,8 +678,10 @@ class Recorder {
             (0, node_fs_1.rmSync)(dir, { recursive: true, force: true });
         return existing;
     }
-    /** Write JUnit + HTML reports and move the run into ./.verikun/runs/<id>/. */
-    static archive(name) {
+    /** Write JUnit + HTML reports and move the run into ./.verikun/runs/<id>/.
+     *  By default captures a bounded device-log tail into `artifacts/logcat.txt`
+     *  (see wantsArchiveLogs / ArchiveLogOpts). */
+    static archive(name, opts = {}) {
         const dir = activeDir();
         if (!(0, node_fs_1.existsSync)(statePath(dir))) {
             throw new errors_1.CliError('No active test run to archive. Run an action first, or `vk run start`.', 1);
@@ -532,7 +691,7 @@ class Recorder {
             throw new errors_1.CliError('Active run state is unreadable (.verikun/run/run.json is corrupt).', 3);
         if (name)
             state.name = name;
-        const dest = Recorder.seal(state, dir);
+        const dest = Recorder.seal(state, dir, opts);
         return { dir: dest, xmlPath: (0, node_path_1.join)(dest, 'report.xml'), htmlPath: (0, node_path_1.join)(dest, 'report.html'), state };
     }
 }

package/dist/server.js

@@ -11,6 +11,7 @@
 //    through the SAME validateNode gate that guards `vk ai` model repairs, so only
 //    KNOWN_COMMANDS action verbs execute — never `ui`/`log`, never a shell. The
 //    server's driver is fixed at startup; client flags can never repoint the device.
+//    Archive-time log capture uses the dedicated /v1/logs endpoint instead.
 //  - /v1/install is a privileged management verb: auth PLUS --allow-install, body
 //    streamed to a server-generated temp path (the client supplies only an
 //    allowlisted extension — never a path), optional sha256 verification.
@@ -154,13 +155,14 @@ function buildServer(config) {
         if (node.type !== 'command')
             throw new HttpError(400, 'rejected: not a command leaf');
         const t0 = Date.now();
-        const { code, error, step, artifacts } = await (0, cli_1.executeForServer)(node.command, node.positionals, (0, ir_1.leafToFlags)(node), config.driver, config.platform);
+        const { code, error, step, artifacts, logStart } = await (0, cli_1.executeForServer)(node.command, node.positionals, (0, ir_1.leafToFlags)(node), config.driver, config.platform);
         (0, output_1.err)(`[server] exec ${node.command} ${node.positionals.join(' ')} → exit ${code} (${Date.now() - t0}ms)`);
         const payload = {
             code,
             ...(error ? { error: (0, rpc_1.describeError)(error) } : {}),
             ...(step ? { step } : {}),
             ...(artifacts && Object.keys(artifacts).length ? { artifacts: encodeArtifacts(artifacts) } : {}),
+            ...(logStart ? { logStart } : {}),
         };
         sendJson(res, 200, payload);
     }
@@ -169,6 +171,45 @@ function buildServer(config) {
         const elements = config.driver.getElements(); // CliError(3) on dump failure → 500 below
         sendJson(res, 200, { elements });
     }
+    async function handleLogs(req, res) {
+        const body = await readBody(req, EXEC_BODY_CAP);
+        let parsed = {};
+        if (body.length) {
+            try {
+                parsed = JSON.parse(body.toString('utf8'));
+            }
+            catch {
+                throw new HttpError(400, 'invalid JSON body');
+            }
+        }
+        // Mirror the driver's --since charset gate so a remote caller cannot inject
+        // into the device shell via a crafted marker (see AdbDriver.getLogs).
+        if (parsed.since !== undefined && parsed.since !== null) {
+            if (typeof parsed.since !== 'string' || !/^[0-9 :.\-]+$/.test(parsed.since)) {
+                throw new HttpError(400, `invalid since: only a logcat timestamp (digits, space, '-', ':', '.') is allowed`);
+            }
+        }
+        const lines = parsed.lines === undefined || parsed.lines === null
+            ? undefined
+            : typeof parsed.lines === 'number' && Number.isFinite(parsed.lines) && parsed.lines > 0
+                ? Math.floor(parsed.lines)
+                : undefined;
+        const appId = parsed.appId === undefined || parsed.appId === null
+            ? undefined
+            : typeof parsed.appId === 'string' && /^[A-Za-z0-9_.-]+$/.test(parsed.appId)
+                ? parsed.appId
+                : (() => {
+                    throw new HttpError(400, `invalid appId '${String(parsed.appId)}'`);
+                })();
+        const logs = config.driver.getLogs({
+            ...(lines !== undefined ? { lines } : {}),
+            ...(parsed.since ? { since: parsed.since } : {}),
+            ...(appId ? { appId } : {}),
+            ...(parsed.scopedOnly ? { scopedOnly: true } : {}),
+        });
+        const payload = { logs };
+        sendJson(res, 200, payload);
+    }
     async function handleInstall(req, res) {
         const ext = String(req.headers['x-verikun-ext'] ?? '').toLowerCase();
         if (ext !== 'apk' && ext !== 'ipa') {
@@ -256,6 +297,8 @@ function buildServer(config) {
             return deviceEndpoint(() => handleExec(req, res));
         if (req.method === 'POST' && path === '/v1/elements')
             return deviceEndpoint(() => handleElements(req, res));
+        if (req.method === 'POST' && path === '/v1/logs')
+            return deviceEndpoint(() => handleLogs(req, res));
         if (req.method === 'POST' && path === '/v1/install') {
             if (!config.allowInstall) {
                 throw new HttpError(403, 'install is disabled on this server (start it with --allow-install)', 3);

package/dist/suite.js

@@ -13,6 +13,8 @@ Object.defineProperty(exports, "__esModule", { value: true });
 exports.sortTestFiles = sortTestFiles;
 exports.listTestFiles = listTestFiles;
 exports.toSuiteResult = toSuiteResult;
+exports.toSuiteAttempt = toSuiteAttempt;
+exports.mergeSuiteAttempts = mergeSuiteAttempts;
 exports.cmdSuite = cmdSuite;
 const node_fs_1 = require("node:fs");
 const node_path_1 = require("node:path");
@@ -100,6 +102,64 @@ function toSuiteResult(file, r, durationMs) {
         ...(r.ok ? {} : { failure: failure ?? 'failed' }),
     };
 }
+/** Compact one attempt for the `attempts` evidence array (pure). */
+function toSuiteAttempt(r) {
+    return {
+        id: r.id,
+        ok: r.ok,
+        durationMs: r.durationMs,
+        costUsd: r.costUsd,
+        ...(r.failure ? { failure: r.failure } : {}),
+    };
+}
+/**
+ * Merge a sequence of attempt rows into the final suite row: primary `id` is the last
+ * attempt (winning green, or last red), cost/duration/repairs sum across attempts, and
+ * prior attempts are retained as flake evidence.
+ */
+function mergeSuiteAttempts(attempts) {
+    if (attempts.length === 0)
+        throw new Error('mergeSuiteAttempts: empty');
+    const last = attempts[attempts.length - 1];
+    if (attempts.length === 1)
+        return last;
+    const round = (n) => Number(n.toFixed(4));
+    const prior = attempts.slice(0, -1).map(toSuiteAttempt);
+    const flaky = last.ok && prior.some((a) => !a.ok);
+    return {
+        ...last,
+        durationMs: attempts.reduce((a, t) => a + t.durationMs, 0),
+        costUsd: round(attempts.reduce((a, t) => a + t.costUsd, 0)),
+        modelRepairs: attempts.reduce((a, t) => a + t.modelRepairs, 0),
+        attempts: prior,
+        ...(flaky ? { flaky: true } : {}),
+    };
+}
+// What --retries will and won't spend an attempt on. The bias is deliberate and
+// asymmetric: a retry costs one test, while giving up costs the whole suite plus a
+// human rerunning it. So the rule is *retry unless a rerun provably cannot change the
+// outcome* — the two predicates below are the only "provably" cases, everything else
+// (flaky selector, wedged app, a wobbling network to `vk server`) earns another go.
+/** Budget aborts won't heal on retry: each attempt gets its own cost ceiling, so a
+ *  rerun just re-aborts at the same place having spent the money twice. */
+function isRetryable(r) {
+    return !r.ok && !r.abortedForBudget;
+}
+/** A thrown USAGE error (exit 2) is the one throw a rerun cannot change — an unreadable
+ *  test file, a payload the server refuses, a flag it doesn't understand. Everything
+ *  else, including every environment error, is retried while attempts remain. */
+function isRetryableThrow(e) {
+    return !(e instanceof errors_1.CliError && e.exitCode === 2);
+}
+function parseRetries(flags) {
+    const n = (0, args_1.flagNum)(flags, 'retries');
+    if (n === undefined)
+        return 0;
+    if (!Number.isInteger(n) || n < 0) {
+        throw new errors_1.CliError(`--retries must be a non-negative integer, got '${n}'`, 2);
+    }
+    return n;
+}
 async function cmdSuite(dirArg, flags, deps) {
     const dir = (0, node_path_1.resolve)(process.cwd(), dirArg);
     if (!(0, node_fs_1.existsSync)(dir) || !(0, node_fs_1.statSync)(dir).isDirectory()) {
@@ -109,69 +169,133 @@ async function cmdSuite(dirArg, flags, deps) {
     if (files.length === 0) {
         throw new errors_1.CliError(`suite: no test files (*.md) in '${dirArg}'`, 2);
     }
+    const retries = parseRetries(flags);
     const suiteId = (0, run_1.runId)();
     const name = (0, args_1.flagStr)(flags, 'name') || (0, node_path_1.basename)(dir);
     const startedAt = new Date().toISOString();
-    (0, output_1.err)(`[suite] '${name}': ${files.length} test(s) from ${dirArg} (${deps.platform}${deps.device ? ` · ${deps.device}` : ''})`);
+    (0, output_1.err)(`[suite] '${name}': ${files.length} test(s) from ${dirArg} (${deps.platform}${deps.device ? ` · ${deps.device}` : ''})${retries > 0 ? ` · up to ${retries} retry(ies) on failure` : ''}`);
     const results = [];
+    const warnings = [];
     let aborted;
+    async function resetApp(label) {
+        // Returns the abort reason when the suite should stop (confirmed env break during reset).
+        if (!deps.reset)
+            return undefined;
+        try {
+            await deps.reset();
+            (0, output_1.err)(`[suite] app state reset${label}`);
+            return undefined;
+        }
+        catch (e) {
+            // A reset that failed because the BOX is broken means nothing after it is
+            // trustworthy — but only if a re-probe agrees. Otherwise surface and continue:
+            // a flaky reset should not zero out the whole suite, and the test itself will
+            // fail loudly if the stale state actually matters.
+            const broken = (0, errors_1.isEnvError)(e) ? await stillBroken(deps) : undefined;
+            if (broken)
+                return broken;
+            (0, output_1.err)(`[suite] reset failed (${e.message}) — continuing`);
+            return undefined;
+        }
+    }
+    /** A confirmed env break with attempts left: say so, pause, and let the loop retry.
+     *  The pause matters — the failures this rides out (a server restart, a wifi drop, a
+     *  USB re-enumeration) clear in seconds, and retrying into the same dead socket
+     *  immediately would burn every attempt inside the outage. */
+    async function noteEnvRetry(file, attempt, reason) {
+        const warn = `${file}: environment error on attempt ${attempt + 1} (${reason}) — retried`;
+        warnings.push(warn);
+        (0, output_1.err)(`[suite] WARN ${warn}`);
+        await sleep((deps.probeRetryMs ?? PROBE_RETRY_MS) * (attempt + 1));
+    }
     for (let i = 0; i < files.length && !aborted; i++) {
         const file = files[i];
         (0, output_1.err)(`[suite] ── (${i + 1}/${files.length}) ${file} ──`);
-        if (deps.reset) {
+        const attemptRows = [];
+        for (let attempt = 0; attempt <= retries; attempt++) {
+            // The last attempt is where a retryable failure becomes the verdict: a confirmed
+            // env break aborts the suite, anything else stands as this test's failed row.
+            const lastAttempt = attempt === retries;
+            if (attempt > 0)
+                (0, output_1.err)(`[suite] retry ${attempt}/${retries} for ${file}`);
+            // Re-isolate before EVERY attempt — between tests and between retries alike.
+            const resetBreak = await resetApp(attempt > 0 ? ' (retry)' : '');
+            if (resetBreak) {
+                if (!lastAttempt) {
+                    await noteEnvRetry(file, attempt, `reset failed: ${resetBreak}`);
+                    continue;
+                }
+                // With no attempt row this test never ran, so notRun starts at the CURRENT file.
+                aborted = {
+                    reason: `reset failed: ${resetBreak}`,
+                    notRun: files.slice(attemptRows.length ? i + 1 : i),
+                };
+                break;
+            }
+            const t0 = Date.now();
             try {
-                await deps.reset();
-                (0, output_1.err)('[suite] app state reset');
+                const r = await deps.runTest((0, node_path_1.join)(dir, file));
+                attemptRows.push(toSuiteResult(file, r, Date.now() - t0));
+                if (r.abortedForEnv) {
+                    const broken = await stillBroken(deps);
+                    if (broken) {
+                        if (!lastAttempt) {
+                            // Even a CONFIRMED break is worth an attempt: the probe window is a couple
+                            // of seconds, which a server restart outlives — and aborting costs the run.
+                            await noteEnvRetry(file, attempt, broken);
+                            continue;
+                        }
+                        aborted = { reason: broken, notRun: files.slice(i + 1) };
+                        break;
+                    }
+                    // Transient env blip: retryable like any other failure.
+                }
+                if (r.ok || !isRetryable(r) || lastAttempt)
+                    break;
             }
             catch (e) {
-                // A reset that failed because the BOX is broken means nothing after it is
-                // trustworthy — but only if a re-probe agrees. Otherwise surface and continue:
-                // a flaky reset should not zero out the whole suite, and the test itself will
-                // fail loudly if the stale state actually matters.
+                // A test that THREW (device gone, server unreachable, bad file) still becomes a
+                // failed row — one broken test must not vaporize the suite report for the tests
+                // that already ran. Out of attempts, a confirmed env break stops the suite.
+                const msg = e instanceof Error ? e.message : String(e);
+                (0, output_1.err)(`[suite] ${file} errored: ${msg}`);
+                attemptRows.push({
+                    id: '',
+                    file,
+                    name: (0, node_path_1.basename)(file, (0, node_path_1.extname)(file)),
+                    ok: false,
+                    durationMs: Date.now() - t0,
+                    costUsd: 0,
+                    steps: 0,
+                    passedSteps: 0,
+                    failedSteps: 0,
+                    modelRepairs: 0,
+                    failure: msg.split('\n')[0],
+                });
                 const broken = (0, errors_1.isEnvError)(e) ? await stillBroken(deps) : undefined;
-                if (broken) {
-                    // This test never ran, so it gets no row — notRun starts at the CURRENT file.
-                    aborted = { reason: `reset failed: ${broken}`, notRun: files.slice(i) };
+                if (lastAttempt) {
+                    if (broken)
+                        aborted = { reason: broken, notRun: files.slice(i + 1) };
                     break;
                 }
-                (0, output_1.err)(`[suite] reset failed (${e.message}) — continuing`);
-            }
-        }
-        const t0 = Date.now();
-        try {
-            const r = await deps.runTest((0, node_path_1.join)(dir, file));
-            results.push(toSuiteResult(file, r, Date.now() - t0));
-            // The test itself reported an environment abort (exit 3 mid-plan). Same rule:
-            // fatal only if the box is still broken. This test HAS a row and a real report,
-            // so notRun starts after it.
-            if (r.abortedForEnv) {
-                const broken = await stillBroken(deps);
+                if (!isRetryableThrow(e))
+                    break;
                 if (broken)
-                    aborted = { reason: broken, notRun: files.slice(i + 1) };
+                    await noteEnvRetry(file, attempt, broken);
             }
         }
-        catch (e) {
-            // A test that THREW (device gone, server unreachable, bad file) still becomes a
-            // failed row — one broken test must not vaporize the suite report for the tests
-            // that already ran. But if it threw because the environment is gone, stop.
-            const msg = e instanceof Error ? e.message : String(e);
-            (0, output_1.err)(`[suite] ${file} errored: ${msg}`);
-            results.push({
-                id: '',
-                file,
-                name: (0, node_path_1.basename)(file, (0, node_path_1.extname)(file)),
-                ok: false,
-                durationMs: Date.now() - t0,
-                costUsd: 0,
-                steps: 0,
-                passedSteps: 0,
-                failedSteps: 0,
-                modelRepairs: 0,
-                failure: msg.split('\n')[0],
-            });
-            const broken = (0, errors_1.isEnvError)(e) ? await stillBroken(deps) : undefined;
-            if (broken)
-                aborted = { reason: broken, notRun: files.slice(i + 1) };
+        if (attemptRows.length === 0) {
+            // Every attempt was blocked by a failing reset, so the test never ran and gets no
+            // row — `aborted.notRun` (set above) already names it. Nothing to merge.
+            break;
+        }
+        const merged = mergeSuiteAttempts(attemptRows);
+        results.push(merged);
+        if (merged.flaky) {
+            const n = merged.attempts?.length ?? 0;
+            const warn = `${file} passed on retry after ${n} failed attempt${n === 1 ? '' : 's'}`;
+            warnings.push(warn);
+            (0, output_1.err)(`[suite] WARN ${warn}`);
         }
     }
     if (aborted) {
@@ -189,6 +313,7 @@ async function cmdSuite(dirArg, flags, deps) {
         totals: (0, report_1.suiteTotals)(results),
         tests: results,
         ...(aborted ? { aborted } : {}),
+        ...(warnings.length ? { warnings } : {}),
     };
     // .verikun/suites/<id>/ sits beside .verikun/runs/<id>/, so index.html reaches a
     // test report at ../../runs/<id>/report.html — the linkBase below.
@@ -198,8 +323,12 @@ async function cmdSuite(dirArg, flags, deps) {
     (0, node_fs_1.writeFileSync)((0, node_path_1.join)(outDir, 'index.html'), (0, report_1.toSuiteHtml)(suite, { linkBase: '../../' }));
     const t = suite.totals;
     (0, output_1.err)(`[suite] ${t.passed}/${t.tests} passed · ${t.steps} steps · $${t.costUsd.toFixed(4)} · ${(t.durationMs / 1000).toFixed(1)}s`);
-    for (const r of results)
-        (0, output_1.err)(`  ${r.ok ? 'PASS' : 'FAIL'} ${r.file}${r.failure ? ` — ${r.failure}` : ''}`);
+    for (const r of results) {
+        const tag = r.flaky ? 'FLAKY' : r.ok ? 'PASS' : 'FAIL';
+        (0, output_1.err)(`  ${tag} ${r.file}${r.failure ? ` — ${r.failure}` : r.flaky ? ' — passed on retry' : ''}`);
+    }
+    if (warnings.length)
+        (0, output_1.err)(`[suite] ${warnings.length} warning(s)`);
     (0, output_1.err)(`[suite] overview: ${(0, node_path_1.join)(outDir, 'index.html')}`);
     if ((0, args_1.flagBool)(flags, 'json'))
         (0, output_1.json)(suite);
@@ -207,6 +336,7 @@ async function cmdSuite(dirArg, flags, deps) {
         (0, output_1.out)(outDir); // primary machine result: the suite directory
     // The CI gate: any failed test fails the invocation (mirrors `vk run archive`). An
     // environment abort exits 3 instead, so CI can tell "the runner is broken" from "the
-    // app regressed" — the whole point of stopping early.
+    // app regressed" — the whole point of stopping early. A flake that recovered is ok
+    // (exit 0) with a warning — that is the whole point of --retries.
     return aborted ? 3 : t.failed > 0 ? 1 : 0;
 }

package/dist/version.js

@@ -3,4 +3,4 @@ Object.defineProperty(exports, "__esModule", { value: true });
 exports.VERSION = void 0;
 // GENERATED by scripts/gen-version.mjs from package.json's "version" at build time
 // (the `prebuild` script). Do NOT edit by hand; bump package.json instead.
-exports.VERSION = '0.16.0';
+exports.VERSION = '0.18.0';

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "verikun",
-  "version": "0.16.0",
+  "version": "0.18.0",
   "description": "Drive Android emulators/devices and iOS simulators for AI agents: tap, type, swipe, screenshot, and inspect the UI hierarchy by semantic identifiers — like Puppeteer for native apps.",
   "keywords": [
     "android",