codeloop-mcp-server 0.1.65 → 0.1.67

package/dist/index.js

@@ -4,6 +4,7 @@ import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js"
 import { z } from "zod";
 import { readFileSync, writeFileSync, existsSync, readdirSync, statSync } from "fs";
 import { slugForTargetChangeEntry } from "./tools/c7_slug.js";
+import { resolveAgentMode, MODE_PARAM_DESCRIPTION as AGENT_MODE_PARAM_DESC, } from "./evidence/agent_mode.js";
 function dirHasFile(dir, predicate) {
     try {
         if (!existsSync(dir))
@@ -605,10 +606,12 @@ Returns: structured report with pass/fail counts, artifact paths, and next-step
     project_dir: z.string().optional().describe("Absolute path to the project root. Defaults to CODELOOP_PROJECT_DIR env var or auto-discovered project directory. MUST be an actual project folder — passing the user's home directory is rejected. If your IDE launches the MCP server from the wrong cwd (common on Windows where Cursor uses C:\\Users\\<name> as cwd), set CODELOOP_PROJECT_DIR or pass this param explicitly."),
     workspace_root: z.string().optional().describe("[Alias for project_dir] Same semantics; accepted because many agents reach for this conventional name. Pass either `project_dir` OR `workspace_root` — they're equivalent."),
     tasks_completed: z.array(z.string()).optional().describe("0.1.52 C5 — free-text titles of the tasks the agent claims to have completed in this code change. Cross-checked against the change manifest produced by C1: every claim should map to >= 1 manifest entry and every manifest entry should map to >= 1 claim. Mismatches surface as warnings in the verify response and feed the change_coverage_evidence gate (C3)."),
+    mode: z.string().optional().describe(AGENT_MODE_PARAM_DESC),
 }, async (params) => {
     const cwd = resolveCwd(params);
     const explicitDir = params.project_dir || params.workspace_root;
     const cfg = explicitDir ? loadConfig(explicitDir) : config;
+    const auditMode = resolveAgentMode({ cwd, paramMode: params.mode, configMode: cfg.agent_mode }) === "audit";
     const result = await withAuth(async () => {
         const { runVerify } = await import("./tools/verify.js");
         const input = {
@@ -638,7 +641,12 @@ Returns: structured report with pass/fail counts, artifact paths, and next-step
     try {
         const { isUIProject } = await import("./tools/gate_check.js");
         const verifyResult = result;
-        if (verifyResult.run_id &&
+        if (auditMode) {
+            // Read-only: the user asked for a problem list, not an auto-fix loop.
+            const { buildAuditDirective } = await import("./evidence/agent_mode.js");
+            postscript = buildAuditDirective("verify");
+        }
+        else if (verifyResult.run_id &&
             (verifyResult.fail_count ?? 0) === 0 &&
             isUIProject(cwd)) {
             const { getArtifactsBaseDir, getRunDir, listRuns } = await import("./evidence/artifacts.js");
@@ -693,7 +701,10 @@ Returns: categorized issues with severity, evidence, root cause, and actionable
     focus_files: z.array(z.string()).optional(),
     project_dir: z.string().optional().describe("Absolute path to the project root. Defaults to CODELOOP_PROJECT_DIR env var or auto-discovered project directory. MUST be an actual project folder — passing the user's home directory is rejected. If your IDE launches the MCP server from the wrong cwd (common on Windows where Cursor uses C:\\Users\\<name> as cwd), set CODELOOP_PROJECT_DIR or pass this param explicitly."),
     workspace_root: z.string().optional().describe("[Alias for project_dir] Same semantics; accepted because many agents reach for this conventional name. Pass either `project_dir` OR `workspace_root` — they're equivalent."),
+    mode: z.string().optional().describe(AGENT_MODE_PARAM_DESC),
 }, async (params) => {
+    const diagnoseCwd = resolveCwd(params);
+    const auditMode = resolveAgentMode({ cwd: diagnoseCwd, paramMode: params.mode, configMode: config.agent_mode }) === "audit";
     const result = await withAuth(async () => {
         const { runDiagnose } = await import("./tools/diagnose.js");
         const input = {
@@ -711,17 +722,24 @@ Returns: categorized issues with severity, evidence, root cause, and actionable
     // result already enumerates the work; the agent must pick them up
     // and execute, not stop to ask "which would you like me to start
     // with?".
-    const fixDirective = [
-        "",
-        "",
-        "⚠️ EXECUTE REPAIR TASKS NOW — do not ask the user which one to start with ⚠️",
-        "The repair_tasks above are the agent's TODO list, in order. For EACH task:",
-        "  1. Modify ONLY the files listed in allowed_file_scope (do not refactor adjacent code).",
-        "  2. After all tasks for a single issue are applied, call codeloop_verify again.",
-        "  3. If new failures surface, call codeloop_diagnose again (max 8 attempts before escalating).",
-        "  4. When verify passes, continue the loop to codeloop_gate_check.",
-        "Do NOT report partial progress to the user between iterations. Do NOT pause to confirm the plan — the diagnose output IS the plan.",
-    ].join("\n");
+    let fixDirective;
+    if (auditMode) {
+        const { buildAuditDirective } = await import("./evidence/agent_mode.js");
+        fixDirective = buildAuditDirective("diagnose");
+    }
+    else {
+        fixDirective = [
+            "",
+            "",
+            "⚠️ EXECUTE REPAIR TASKS NOW — do not ask the user which one to start with ⚠️",
+            "The repair_tasks above are the agent's TODO list, in order. For EACH task:",
+            "  1. Modify ONLY the files listed in allowed_file_scope (do not refactor adjacent code).",
+            "  2. After all tasks for a single issue are applied, call codeloop_verify again.",
+            "  3. If new failures surface, call codeloop_diagnose again (max 8 attempts before escalating).",
+            "  4. When verify passes, continue the loop to codeloop_gate_check.",
+            "Do NOT report partial progress to the user between iterations. Do NOT pause to confirm the plan — the diagnose output IS the plan.",
+        ].join("\n");
+    }
     return {
         content: withInitHint([{ type: "text", text: JSON.stringify(result, null, 2) + fixDirective }]),
     };
@@ -761,7 +779,9 @@ Returns: pass/fail for each gate, overall confidence score, and recommendation.`
     project_dir: z.string().optional().describe("Absolute path to the project root. Defaults to CODELOOP_PROJECT_DIR env var or auto-discovered project directory. MUST be an actual project folder — passing the user's home directory is rejected. If your IDE launches the MCP server from the wrong cwd (common on Windows where Cursor uses C:\\Users\\<name> as cwd), set CODELOOP_PROJECT_DIR or pass this param explicitly."),
     workspace_root: z.string().optional().describe("[Alias for project_dir] Same semantics; accepted because many agents reach for this conventional name. Pass either `project_dir` OR `workspace_root` — they're equivalent."),
     recent_thinking: z.string().optional().describe("0.1.52 C6 — optional dump of the agent's recent thinking / rationale (last few turns of the loop). When present, the gate scans it for anti-rationalisation phrases ('comprehensive verification confirms', 'further interaction would be redundant', 'grid is empty so can't test', etc.) and surfaces specific matches in the continue_fixing postscript so the agent stops repeating the rationalisation and acts on the per-gate next steps instead. Safe to omit — the canonical FORBIDDEN list still ships in the directive without a hit."),
+    mode: z.string().optional().describe(AGENT_MODE_PARAM_DESC),
 }, async (params) => {
+    const gateAuditMode = resolveAgentMode({ cwd: resolveCwd(params), paramMode: params.mode, configMode: config.agent_mode }) === "audit";
     const result = await withAuth(async () => {
         const { runGateCheck } = await import("./tools/gate_check.js");
         const input = {
@@ -792,6 +812,14 @@ Returns: pass/fail for each gate, overall confidence score, and recommendation.`
     }, { tool: "codeloop_gate_check", cwd: resolveCwd(params), input: params });
     const resultJson = JSON.stringify(result, null, 2);
     const gateResult = result;
+    if (gateAuditMode) {
+        // Read-only: report gate results as an audit summary; do not push the
+        // auto-fix loop even when gates fail.
+        const { buildAuditDirective } = await import("./evidence/agent_mode.js");
+        return {
+            content: withInitHint([{ type: "text", text: resultJson + buildAuditDirective("gate_check") }], resolveCwd(params)),
+        };
+    }
     if (gateResult.recommendation === "continue_fixing") {
         // Per-gate next-step enumeration. The auto-fix loop's biggest
         // failure mode was the generic directive ("call verify, diagnose,
@@ -1421,6 +1449,7 @@ Returns: confirmation + the captured image as an MCP ImageContent block so you c
     project_dir: z.string().optional().describe("Absolute path to the project root. Defaults to CODELOOP_PROJECT_DIR env var or auto-discovered project directory. MUST be an actual project folder — passing the user's home directory is rejected. If your IDE launches the MCP server from the wrong cwd (common on Windows where Cursor uses C:\\Users\\<name> as cwd), set CODELOOP_PROJECT_DIR or pass this param explicitly."),
     workspace_root: z.string().optional().describe("[Alias for project_dir] Same semantics; accepted because many agents reach for this conventional name. Pass either `project_dir` OR `workspace_root` — they're equivalent."),
     target_change_entry: z.string().optional().describe("0.1.52 C7 — verbatim display name of the change-manifest entry this screenshot exercises (e.g. 'datagrid_column: \"Product Code\"' or 'PhotometricConfigurations.ProductCode'). When present, the screenshot file is auto-anchored: the filename is prefixed with a slugged form of the entry so the change_coverage_evidence (C3) gate's screenshot scan can credit this evidence to the correct manifest entry without fuzzy matching. The value is also persisted alongside the screenshot path in the response so downstream tools (interaction_replay, gate_check) can use it."),
+    target_type: z.string().optional().describe("Capture source. Omit for desktop/window capture (default). Set 'android_emulator' (or 'android') to capture the BOOTED Android emulator via adb, or 'ios_simulator' (or 'ios', macOS only) to capture the booted iOS simulator via simctl. Use this to screenshot a Flutter/native mobile app running on a simulator instead of the host desktop."),
 }, async (params) => {
     const authResult = await withAuth(async () => {
         const { captureScreenshot } = await import("./runners/screenshot.js");
@@ -1462,7 +1491,14 @@ Returns: confirmation + the captured image as an MCP ImageContent block so you c
         const finalScreenName = targetChangeEntry
             ? `${params.screen_name}--c7-${slugForTargetChangeEntry(targetChangeEntry)}`
             : params.screen_name;
-        const result = await captureScreenshot(screenshotsDir, finalScreenName, targetApp, undefined, { desktopAppMode: desktopApp });
+        // Mobile capture: when the agent asks for an emulator/simulator
+        // screenshot, route to the adb/simctl capture path instead of the
+        // desktop window grab (and skip the desktop-app honesty refusal, which
+        // only applies to host-window capture).
+        const ttRaw = normalizeTargetType(params.target_type);
+        const explicitTargetType = ttRaw === "android_emulator" || ttRaw === "ios_simulator" ? ttRaw : undefined;
+        const isMobileCapture = explicitTargetType !== undefined;
+        const result = await captureScreenshot(screenshotsDir, finalScreenName, isMobileCapture ? undefined : targetApp, explicitTargetType, { desktopAppMode: isMobileCapture ? false : desktopApp });
         // Photometry-DB E2E 8 follow-on: when we capture a desktop app
         // window, also resolve its on-screen bounds so the agent can
         // (a) compute window-relative coords from the returned image
@@ -1976,10 +2012,37 @@ App logs (stdout, logcat, simctl log) are automatically captured alongside the v
             }
             catch { /* best-effort */ }
         }
+        // Mobile device-readiness probe. adb/simctl screen capture attaches to a
+        // BOOTED emulator/simulator; if none is running the recording would fail
+        // deep inside the recorder with a cryptic error. Probe first and, when
+        // nothing is booted, push the agent to OPEN the right simulator(s) with
+        // concrete commands rather than letting it guess. CodeLoop does not boot
+        // devices itself (heavy + can hijack the desktop), it directs the agent.
+        let deviceReadinessDirective;
+        if (targetType === "android_emulator" || targetType === "ios_simulator") {
+            try {
+                const { probeBootedDevices, detectMobileTargets, buildOpenSimulatorsDirective } = await import("./runners/device_probe.js");
+                const booted = await probeBootedDevices();
+                const wantedBooted = targetType === "android_emulator"
+                    ? booted.android.length > 0
+                    : booted.ios.length > 0;
+                if (!wantedBooted) {
+                    deviceReadinessDirective = buildOpenSimulatorsDirective({
+                        booted,
+                        targets: detectMobileTargets(cwd),
+                        context: "recording",
+                    });
+                }
+            }
+            catch { /* best-effort */ }
+        }
         const result = await startBackgroundRecording(videosDir, appName ?? "", params.max_duration_seconds, targetType);
         if (autoLaunchSummary) {
             result.auto_launch = autoLaunchSummary;
         }
+        if (deviceReadinessDirective) {
+            result.device_readiness = deviceReadinessDirective;
+        }
         if (binaryFreshnessWarning) {
             result.binary_freshness_warning = binaryFreshnessWarning;
             result.binary_freshness = binaryFreshnessDetails;
@@ -2721,7 +2784,13 @@ pass app_name explicitly. This ensures interactions hit the app window, not the
 
 Core actions: click, double_click, right_click, hover, type, keystroke, hotkey, scroll,
   drag_drop, long_press, type_and_submit, type_and_tab, fill_form, select_option, toggle,
-  navigate_url, navigate_back, wait, sequence.
+  navigate_url, navigate_back, wait, sequence, get_text.
+get_text (browser + android_emulator): reads back rendered text. Browser = Playwright
+  textContent of a selector (or the whole page body); Android = uiautomator hierarchy text
+  (selector acts as a text/resource-id substring filter). Use it to VERIFY/ANALYZE dynamic
+  responses — e.g. after type_and_submit into a chatbox, get_text the assistant message to
+  read the AI's answer and judge it. Pass expect_contains to assert the response contains an
+  expected substring (success=false if not). iOS: use codeloop_capture_screenshot + vision.
 navigate_url works on ALL targets: browser (Playwright), desktop (opens Chrome/Safari via
   osascript/PowerShell/xdg-open), Android (adb deep link), iOS (simctl openurl).
 scroll works on desktop via CGEvent (macOS), user32 mouse_event (Windows), xdotool (Linux).
@@ -2735,7 +2804,8 @@ Windows: win_ui_inspect, win_ui_automate — PowerShell UI Automation for UWP/Wi
 MANDATORY for web apps: You MUST type into form fields, fill login/signup forms, test
 validation errors, and click submit buttons. Just navigating pages is NOT enough.
 Wait 1-2 seconds between interactions so video frames capture state changes.`, {
-    action: z.string().describe("Action to perform: click, double_click, right_click, hover, type, keystroke, hotkey, scroll, drag_drop, long_press, type_and_submit, type_and_tab, fill_form, select_option, toggle, upload_file, navigate_url, navigate_back, navigate_forward, wait, sequence, swipe, back_button, home_button, deep_link, grant_permission, rotate_device, biometric_auth, launch_app, clear_app_data, mock_location, simulate_network, maestro_flow, win_ui_inspect, win_ui_automate"),
+    action: z.string().describe("Action to perform: click, double_click, right_click, hover, type, keystroke, hotkey, scroll, drag_drop, long_press, type_and_submit, type_and_tab, fill_form, select_option, toggle, get_text, upload_file, navigate_url, navigate_back, navigate_forward, wait, sequence, swipe, back_button, home_button, deep_link, grant_permission, rotate_device, biometric_auth, launch_app, clear_app_data, mock_location, simulate_network, maestro_flow, win_ui_inspect, win_ui_automate"),
+    expect_contains: z.string().optional().describe("[get_text] Optional assertion: the read-back text must contain this substring (case-insensitive). When set, success is false if the substring is absent. Use to confirm an AI chatbot answer / dynamic response actually rendered the expected content."),
     target_type: targetTypeSchema.optional()
         .describe("Interaction target. Auto-detected if omitted. Accepts synonyms: `windows_desktop`/`mac_desktop`/`linux_desktop` → `desktop`; `web` → `browser`; `android` → `android_emulator`; `ios` → `ios_simulator`."),
     x: z.number().optional().describe("X coordinate for click/scroll/drag/swipe"),
@@ -2790,6 +2860,11 @@ Wait 1-2 seconds between interactions so video frames capture state changes.`, {
         const action = params.action;
         let success = false;
         let detail = "";
+        // Text read back from the UI (get_text action) so the agent can analyze
+        // dynamic responses — e.g. the AI chatbot's answer after submitting a
+        // prompt. Surfaced in the response + interaction log as evidence.
+        let extractedText = null;
+        let assertion;
         // Import runners lazily
         const wm = await import("./runners/window_manager.js");
         const bi = await import("./runners/browser_interaction.js");
@@ -3051,6 +3126,62 @@ Wait 1-2 seconds between interactions so video frames capture state changes.`, {
                 }
                 detail = `type "${(params.text || "").substring(0, 50)}"`;
                 break;
+            case "get_text": {
+                // Read back rendered text so the agent can verify/analyze a dynamic
+                // response (e.g. an AI chatbot answer). Browser: Playwright
+                // textContent of the selector (or <body> if omitted). Windows
+                // desktop: the UIA value/name readback already runs below via
+                // win_ui_automate; for a pure get_text on desktop, point the agent
+                // at win_ui_inspect. Mobile text scraping is not yet supported.
+                if (tt === "browser") {
+                    extractedText = await bi.browserGetText(params.selector);
+                    success = extractedText !== null;
+                    const preview = (extractedText ?? "").replace(/\s+/g, " ").trim().slice(0, 120);
+                    detail = success
+                        ? `get_text ${params.selector ?? "body"} → "${preview}${(extractedText ?? "").length > 120 ? "…" : ""}"`
+                        : `get_text failed (no page / selector "${params.selector ?? "body"}" not found)`;
+                }
+                else if (tt === "android_emulator") {
+                    // Read the rendered text off the booted Android emulator via the
+                    // uiautomator hierarchy — lets the agent analyze a Flutter/native
+                    // AI chatbot's answer without OCR. `selector` is treated as a
+                    // text/resource-id substring filter when provided.
+                    extractedText = await wm.adbGetText(params.selector);
+                    success = extractedText !== null && extractedText.length > 0;
+                    const preview = (extractedText ?? "").replace(/\s+/g, " ").trim().slice(0, 120);
+                    detail = success
+                        ? `get_text (android${params.selector ? ` ~"${params.selector}"` : ""}) → "${preview}${(extractedText ?? "").length > 120 ? "…" : ""}"`
+                        : "get_text (android) found no visible text — is the emulator booted and the app foregrounded?";
+                }
+                else if (tt === "ios_simulator") {
+                    // simctl has no UI-tree dump; the agent reads the answer from a
+                    // screenshot via its own vision. Point it there explicitly.
+                    success = false;
+                    detail =
+                        "get_text is not available for iOS simulators (simctl exposes no UI-tree dump). Call codeloop_capture_screenshot with target_type=\"ios_simulator\" and read the AI answer from the image with your vision.";
+                }
+                else {
+                    success = false;
+                    detail =
+                        "get_text is supported for browser and android_emulator targets. For Windows desktop, use win_ui_inspect to read element text; for iOS / other, capture a screenshot and read it visually.";
+                }
+                // Optional assertion: confirm the read text contains expected substring(s).
+                const expectContains = params.expect_contains;
+                if (typeof expectContains === "string" && expectContains.length > 0) {
+                    const hay = (extractedText ?? "").toLowerCase();
+                    const matched = hay.includes(expectContains.toLowerCase());
+                    assertion = {
+                        expected: expectContains,
+                        matched,
+                        actual_excerpt: (extractedText ?? "").replace(/\s+/g, " ").trim().slice(0, 200),
+                    };
+                    success = success && matched;
+                    detail += matched
+                        ? ` | assertion PASSED (contains "${expectContains}")`
+                        : ` | assertion FAILED (does NOT contain "${expectContains}")`;
+                }
+                break;
+            }
             case "keystroke":
                 if (params.key) {
                     if (tt === "browser") {
@@ -3924,6 +4055,10 @@ Wait 1-2 seconds between interactions so video frames capture state changes.`, {
             interactionEntry.console_errors = consoleErrors;
         if (clickEffectVerification)
             interactionEntry.verification = clickEffectVerification;
+        if (extractedText !== null)
+            interactionEntry.extracted_text = extractedText.slice(0, 1000);
+        if (assertion)
+            interactionEntry.assertion = assertion;
         try {
             const activeIds = vr.getActiveRecordingIds();
             if (activeIds.length > 0) {
@@ -3949,6 +4084,10 @@ Wait 1-2 seconds between interactions so video frames capture state changes.`, {
             action,
             detail,
         };
+        if (extractedText !== null)
+            ret.extracted_text = extractedText;
+        if (assertion)
+            ret.assertion = assertion;
         if (modalPersistenceDirective)
             ret._f4_directive = modalPersistenceDirective;
         return ret;