@sagentlab/navarch-runtime 0.1.138 → 0.1.140

package/dist/browser-evidence.cjs

@@ -22,18 +22,20 @@ const node_path_1 = __importDefault(require("node:path"));
 const BROWSER_USE_RUNTIMES = ["codex", "claude-code", "antigravity", "opencode", "acp", "muse"];
 /**
  * Which navarch_browser tools a host session gets. `browserUse` adds capture
- * and local_browser_* for browser-use tasks. `playwrightSuites` adds
- * start_playwright_browser to every macOS Codex build/verify/qa session,
- * because Codex's seatbelt cannot launch the project's own Playwright suite's
- * Chromium whether or not the task asked for browser-use.
+ * and local_browser_* for browser-use build/verify/qa tasks. `playwrightSuites`
+ * adds start_playwright_browser to every macOS Codex build/verify/qa/ops
+ * session, because Codex's seatbelt cannot launch the project's own Playwright
+ * suite's Chromium whether or not the task asked for browser-use. Ops tasks
+ * diagnose suite failures too (ShiftSwap issue #154).
  */
 function navarchBrowserTools(input) {
-    if (input.mcpConfigPath || input.sandboxMode !== "host" || !["build", "verify", "qa"].includes(input.taskType)) {
+    if (input.mcpConfigPath || input.sandboxMode !== "host")
         return { browserUse: false, playwrightSuites: false };
-    }
     return {
-        browserUse: input.requiredCapabilities?.includes("browser-use") === true && BROWSER_USE_RUNTIMES.includes(input.runtime),
-        playwrightSuites: input.runtime === "codex" && input.platform === "darwin",
+        browserUse: ["build", "verify", "qa"].includes(input.taskType)
+            && input.requiredCapabilities?.includes("browser-use") === true && BROWSER_USE_RUNTIMES.includes(input.runtime),
+        playwrightSuites: ["build", "verify", "qa", "ops"].includes(input.taskType)
+            && input.runtime === "codex" && input.platform === "darwin",
     };
 }
 async function installEvidenceBrowser() {
@@ -214,23 +216,25 @@ async function startBrowserEvidence(options) {
         }
         const server = new mcp_js_1.McpServer({ name: "navarch-browser", version: "1.0.0" });
         const transport = new streamableHttp_js_1.StreamableHTTPServerTransport({ sessionIdGenerator: undefined, enableJsonResponse: true });
-        server.registerTool("capture_visual_evidence", {
-            description: "Capture and upload desktop and mobile screenshots using Navarch's managed local browser. Works outside the coding-agent sandbox. Returns durable artifact URLs. Capture does not certify acceptance or that the URL serves the supplied commit. Evidence is automatically attached to the delivery PR at session completion.",
-            inputSchema: captureShape,
-        }, async (args) => {
-            try {
-                if (busy)
-                    throw new Error("A capture is already running. Wait for it to finish.");
-                const pending = capture(args);
-                activeCapture = pending;
-                const { captured, images } = await pending;
-                return { content: [{ type: "text", text: JSON.stringify(captured) }, ...images] };
-            }
-            catch (error) {
-                return { isError: true, content: [{ type: "text", text: String(error) }] };
-            }
-        });
-        (0, local_browser_cjs_1.registerLocalBrowserTools)(server, localBrowser);
+        if (options.browserUse !== false) {
+            server.registerTool("capture_visual_evidence", {
+                description: "Capture and upload desktop and mobile screenshots using Navarch's managed local browser. Works outside the coding-agent sandbox. Returns durable artifact URLs. Capture does not certify acceptance or that the URL serves the supplied commit. Evidence is automatically attached to the delivery PR at session completion.",
+                inputSchema: captureShape,
+            }, async (args) => {
+                try {
+                    if (busy)
+                        throw new Error("A capture is already running. Wait for it to finish.");
+                    const pending = capture(args);
+                    activeCapture = pending;
+                    const { captured, images } = await pending;
+                    return { content: [{ type: "text", text: JSON.stringify(captured) }, ...images] };
+                }
+                catch (error) {
+                    return { isError: true, content: [{ type: "text", text: String(error) }] };
+                }
+            });
+            (0, local_browser_cjs_1.registerLocalBrowserTools)(server, localBrowser);
+        }
         if (options.playwrightServer)
             (0, playwright_server_cjs_1.registerPlaywrightServerTool)(server, options.playwrightServer);
         res.on("close", () => { void transport.close(); void server.close(); });

package/dist/delivery.cjs

@@ -176,10 +176,11 @@ function deliveryIncompleteSection(args) {
         "so your earlier edits, commits, and branch are all still here — start with `git status` and",
         "`git log --oneline` and continue from there rather than redoing the work.",
         "",
-        "Inspect tracked and untracked changes before deciding what is undelivered work. Preserve required",
-        "evidence in task artifacts, remove only temporary files you created (such as `.review-*/` scratch",
-        "directories), and restore only tracked edits you made solely for local checks. Do not delete",
-        "unfamiliar or pre-existing changes, hide leftovers with ignore rules, or commit scratch files.",
+        "Inspect tracked and untracked changes before deciding what is undelivered work. Required evidence",
+        "belongs in your final report or in task or issue comments; files in the worktree are not evidence.",
+        "Remove only temporary files you created (such as `.review-*/` scratch directories or evidence files",
+        "whose content you already reported), and restore only tracked edits you made solely for local checks.",
+        "Do not delete unfamiliar or pre-existing changes, hide leftovers with ignore rules, or commit scratch files.",
         ...(args.taskType === "review"
             ? [
                 "This is a read-only review task. The generic delivery failure does not authorize commits,",

package/dist/playwright-server.cjs

@@ -259,7 +259,11 @@ function createPlaywrightServer(options) {
             canonicalCore,
             token,
         ];
-        const childEnv = { TMPDIR: params.PRIVATE };
+        // Chromium's macOS temp dir ignores TMPDIR and falls back to the user's
+        // /var/folders/.../T, which the profile does not let it write. Every
+        // download creates a temp file there first, so without this each one
+        // ended `canceled` (ShiftSwap issue #154).
+        const childEnv = { TMPDIR: params.PRIVATE, MAC_CHROMIUM_TMPDIR: params.PRIVATE };
         for (const name of ["HOME", "PATH", "LANG", "PLAYWRIGHT_BROWSERS_PATH"]) {
             if (env[name] !== undefined)
                 childEnv[name] = env[name];

package/dist/prompt.cjs

@@ -12,7 +12,7 @@ exports.renderCompletionRemediationPrompt = renderCompletionRemediationPrompt;
 exports.renderResumedCompletionRemediationPrompt = renderResumedCompletionRemediationPrompt;
 exports.renderHostedCheckpointNote = renderHostedCheckpointNote;
 const CODE_GRAPH_GUIDANCE = "For JS/TS dependency inspection, the runtime provides `navarch-runtime code-graph --query <symbol-or-file> --mode search|callers|callees|impact --commit HEAD`. If the runtime is not on PATH, use its installed entry point; do not install another copy mid-task. Results describe committed source and include commit/freshness warnings; inspect current files for uncommitted changes and unresolved dependencies.";
-const PROCESS_GUIDANCE = "Process management: this host may run other agent sessions. Keep the PID or tool handle of each preview server you start and stop only that owned process. Never use broad process-name or command-line matching (pkill, killall) to clean up servers: the agent launcher carries the task prompt in its arguments and can match those patterns, terminating your own session. Use a different port when a port is occupied by a process you do not own. Keep scratch files in the current worktree; paths in previous transcripts belong to earlier sessions. Before your final report, inspect `git status --short`, including untracked files. Preserve required evidence in task artifacts, then remove only temporary files you created and restore only tracked edits you made solely for local checks. Do not delete unfamiliar or pre-existing changes, hide leftovers with ignore rules, or commit scratch files merely to make the worktree clean. Review tasks deliver a native review and attestation, never commits, pushes, or a new PR.";
+const PROCESS_GUIDANCE = "Process management: this host may run other agent sessions. Keep the PID or tool handle of each preview server you start and stop only that owned process. Never use broad process-name or command-line matching (pkill, killall) to clean up servers: the agent launcher carries the task prompt in its arguments and can match those patterns, terminating your own session. Use a different port when a port is occupied by a process you do not own. Keep scratch files in the current worktree; paths in previous transcripts belong to earlier sessions. Before your final report, inspect `git status --short`, including untracked files. Put required evidence in your final report or in task or issue comments, not in files: a file left in the worktree is not delivered evidence, and an untracked one fails delivery as uncommitted work. Then remove only temporary files you created, including evidence files whose content you already reported, and restore only tracked edits you made solely for local checks. Do not delete unfamiliar or pre-existing changes, hide leftovers with ignore rules, or commit scratch files merely to make the worktree clean. Review tasks deliver a native review and attestation, never commits, pushes, or a new PR.";
 /** Overrides the playbooks' "default branch" PR base when the project sets a working branch. */
 function workingBranchNote(workingBranch) {
     if (!workingBranch)

package/dist/session.cjs

@@ -731,7 +731,7 @@ async function runClaimedSession(deps, claimed, sessionId, lifecycle) {
                 registry.register(playwrightServer.token);
             browserEvidence = await (0, browser_evidence_cjs_1.startBrowserEvidence)({
                 apiBase: api.getBaseUrl(), machineToken: api.getToken() ?? "", leaseId,
-                isActive, playwrightServer,
+                isActive, playwrightServer, browserUse,
             });
             registry.register(browserEvidence.token);
             const mcp = JSON.parse(await node_fs_1.promises.readFile(mcpConfigPath, "utf8"));

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sagentlab/navarch-runtime",
-  "version": "0.1.138",
+  "version": "0.1.140",
   "description": "Navarch machine-side session manager: claims delivery tasks and runs them through local coding agents.",
   "type": "commonjs",
   "license": "MIT",