@a5c-ai/babysitter-sdk 0.0.183 → 0.0.184-staging.347227ea

package/dist/cli/commands/harnessCreateRun.js

@@ -68,7 +68,7 @@ async function handleHarnessCreateRun(parsed) {
             const workDir = workspace ?? process.cwd();
             processPath = await (0, harnessPhase1_1.runProcessDefinitionPhase)({
                 prompt: prompt,
-                outputPath: (0, harnessPhase1_1.getGeneratedProcessPath)(workDir),
+                outputDir: (0, harnessPhase1_1.getProcessOutputDir)(workDir),
                 workspace: workDir,
                 model,
                 interactive,

package/dist/cli/commands/harnessPhase1.d.ts

@@ -6,6 +6,8 @@
  */
 import * as readline from "node:readline";
 import { type ExternalWorkspaceAssessment, type CompressionConfig, type HarnessPromptContext as SessionCreatePromptContext } from "./harnessUtils";
+export declare function getProcessOutputDir(workDir: string): string;
+/** @deprecated Use getProcessOutputDir instead */
 export declare function getGeneratedProcessPath(workDir: string): string;
 export declare function execShellEffect(command: string, args: string[], cwd?: string): Promise<{
     stdout: string;
@@ -18,7 +20,7 @@ export declare function extractJsonObjectFromText(text: string): string | null;
 export declare function coerceAgentResultValue(taskDef: Record<string, unknown>, output: string): unknown;
 export declare function buildExternalProcessDefinitionPrompt(args: {
     prompt: string;
-    outputPath: string;
+    outputDir: string;
     workspace?: string;
     promptContext: SessionCreatePromptContext;
     workspaceAssessment: ExternalWorkspaceAssessment;
@@ -29,7 +31,7 @@ export declare function buildExternalProcessConformancePrompt(args: {
 }): string;
 export declare function runProcessDefinitionPhase(args: {
     prompt: string;
-    outputPath: string;
+    outputDir: string;
     workspace?: string;
     model?: string;
     interactive: boolean;

package/dist/cli/commands/harnessPhase1.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"harnessPhase1.d.ts","sourceRoot":"","sources":["../../../src/cli/commands/harnessPhase1.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAGH,OAAO,KAAK,QAAQ,MAAM,eAAe,CAAC;AAa1C,OAAO,EAIL,KAAK,2BAA2B,EAEhC,KAAK,iBAAiB,EAGtB,KAAK,oBAAoB,IAAI,0BAA0B,EAwBxD,MAAM,gBAAgB,CAAC;AAkBxB,wBAAgB,uBAAuB,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,CAE/D;AAyHD,wBAAgB,eAAe,CAC7B,OAAO,EAAE,MAAM,EACf,IAAI,EAAE,MAAM,EAAE,EACd,GAAG,CAAC,EAAE,MAAM,GACX,OAAO,CAAC;IAAE,MAAM,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,MAAM,CAAC;IAAC,QAAQ,EAAE,MAAM,CAAA;CAAE,CAAC,CAyB/D;AAg5BD,wBAAgB,gBAAgB,CAAC,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,MAAM,CAmDzE;AAED,wBAAgB,sCAAsC,CAAC,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,MAAM,EAAE,CAU/F;AAED,wBAAgB,yBAAyB,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,GAAG,IAAI,CAgBrE;AAED,wBAAgB,sBAAsB,CAAC,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,EAAE,MAAM,EAAE,MAAM,GAAG,OAAO,CAwBhG;AAwGD,wBAAgB,oCAAoC,CAAC,IAAI,EAAE;IACzD,MAAM,EAAE,MAAM,CAAC;IACf,UAAU,EAAE,MAAM,CAAC;IACnB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,aAAa,EAAE,0BAA0B,CAAC;IAC1C,mBAAmB,EAAE,2BAA2B,CAAC;CAClD,GAAG,MAAM,CA+ET;AAED,wBAAgB,qCAAqC,CAAC,IAAI,EAAE;IAC1D,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;CAChB,GAAG,MAAM,CA4BT;AAkND,wBAAsB,yBAAyB,CAAC,IAAI,EAAE;IACpD,MAAM,EAAE,MAAM,CAAC;IACf,UAAU,EAAE,MAAM,CAAC;IACnB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,OAAO,CAAC;IACrB,EAAE,EAAE,QAAQ,CAAC,SAAS,GAAG,IAAI,CAAC;IAC9B,IAAI,EAAE,OAAO,CAAC;IACd,OAAO,EAAE,OAAO,CAAC;IACjB,iBAAiB,EAAE,iBAAiB,GAAG,IAAI,CAAC;IAC5C,aAAa,EAAE,0BAA0B,CAAC;IAC1C,mBAAmB,EAAE,MAAM,CAAC;CAC7B,GAAG,OAAO,CAAC,MAAM,CAAC,CAwgBlB"}
+{"version":3,"file":"harnessPhase1.d.ts","sourceRoot":"","sources":["../../../src/cli/commands/harnessPhase1.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAGH,OAAO,KAAK,QAAQ,MAAM,eAAe,CAAC;AAa1C,OAAO,EAIL,KAAK,2BAA2B,EAEhC,KAAK,iBAAiB,EAGtB,KAAK,oBAAoB,IAAI,0BAA0B,EAyBxD,MAAM,gBAAgB,CAAC;AAkBxB,wBAAgB,mBAAmB,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,CAE3D;AAED,kDAAkD;AAClD,wBAAgB,uBAAuB,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,CAE/D;AA0HD,wBAAgB,eAAe,CAC7B,OAAO,EAAE,MAAM,EACf,IAAI,EAAE,MAAM,EAAE,EACd,GAAG,CAAC,EAAE,MAAM,GACX,OAAO,CAAC;IAAE,MAAM,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,MAAM,CAAC;IAAC,QAAQ,EAAE,MAAM,CAAA;CAAE,CAAC,CAyB/D;AA85BD,wBAAgB,gBAAgB,CAAC,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,MAAM,CAmDzE;AAED,wBAAgB,sCAAsC,CAAC,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,MAAM,EAAE,CAU/F;AAED,wBAAgB,yBAAyB,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,GAAG,IAAI,CAgBrE;AAED,wBAAgB,sBAAsB,CAAC,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,EAAE,MAAM,EAAE,MAAM,GAAG,OAAO,CAwBhG;AAiHD,wBAAgB,oCAAoC,CAAC,IAAI,EAAE;IACzD,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,aAAa,EAAE,0BAA0B,CAAC;IAC1C,mBAAmB,EAAE,2BAA2B,CAAC;CAClD,GAAG,MAAM,CA+ET;AAED,wBAAgB,qCAAqC,CAAC,IAAI,EAAE;IAC1D,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;CAChB,GAAG,MAAM,CA4BT;AAmND,wBAAsB,yBAAyB,CAAC,IAAI,EAAE;IACpD,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,OAAO,CAAC;IACrB,EAAE,EAAE,QAAQ,CAAC,SAAS,GAAG,IAAI,CAAC;IAC9B,IAAI,EAAE,OAAO,CAAC;IACd,OAAO,EAAE,OAAO,CAAC;IACjB,iBAAiB,EAAE,iBAAiB,GAAG,IAAI,CAAC;IAC5C,aAAa,EAAE,0BAA0B,CAAC;IAC1C,mBAAmB,EAAE,MAAM,CAAC;CAC7B,GAAG,OAAO,CAAC,MAAM,CAAC,CAmjBlB"}

package/dist/cli/commands/harnessPhase1.js

@@ -39,6 +39,7 @@ var __importStar = (this && this.__importStar) || (function () {
     };
 })();
 Object.defineProperty(exports, "__esModule", { value: true });
+exports.getProcessOutputDir = getProcessOutputDir;
 exports.getGeneratedProcessPath = getGeneratedProcessPath;
 exports.execShellEffect = execShellEffect;
 exports.buildAgentPrompt = buildAgentPrompt;
@@ -70,8 +71,12 @@ const dynamicImportModule = (() => {
     return new Function("specifier", "return import(specifier);");
 })();
 // ── Process File Utilities ───────────────────────────────────────────
+function getProcessOutputDir(workDir) {
+    return path.join(workDir, ".a5c", "processes");
+}
+/** @deprecated Use getProcessOutputDir instead */
 function getGeneratedProcessPath(workDir) {
-    return path.join(workDir, "generated-process.mjs");
+    return getProcessOutputDir(workDir);
 }
 function looksLikeProcessDefinitionSource(source) {
     const normalized = source.trim();
@@ -111,6 +116,7 @@ function extractMentionedProcessPaths(text, workspace) {
         /([A-Za-z]:[\\/][^\r\n"'`]+?\.m?js)\b/g,
         /((?:\.{0,2}[\\/]|\/)[^\r\n"'`]+?\.m?js)\b/g,
         /\b([A-Za-z0-9_.\-\\/]+generated-process\.m?js)\b/g,
+        /\b(\.a5c[\\/]processes[\\/][A-Za-z0-9_.-]+\.m?js)\b/g,
     ];
     const matches = new Set();
     for (const pattern of patterns) {
@@ -742,7 +748,7 @@ async function validateProcessExport(filePath) {
             category: harnessUtils_1.ErrorCategory.Validation,
             nextSteps: [
                 "Every TaskDef returned from defineTask(...) must include a top-level kind string",
-                "Use kind: \"agent\" with agent: { ... }, kind: \"shell\" with shell: { command: ... }, or kind: \"node\" with node: { entry, args? } as appropriate",
+                "Use kind: \"agent\" with agent: { ... }, kind: \"shell\" with shell: { command: ... }, or another supported effect kind that the agent will execute and post manually",
             ],
         });
     }
@@ -754,7 +760,7 @@ async function validateProcessExport(filePath) {
             category: harnessUtils_1.ErrorCategory.Validation,
             nextSteps: [
                 "Match each task's kind to its body shape",
-                "Agent tasks must use kind: \"agent\", shell tasks must use kind: \"shell\", and node tasks must use kind: \"node\"",
+                "Agent tasks must use kind: \"agent\" and shell tasks must use kind: \"shell\". Do not generate node task definitions in authored processes",
             ],
         });
     }
@@ -768,6 +774,16 @@ async function validateProcessExport(filePath) {
             ],
         });
     }
+    const nodeTaskIds = getDefineTaskIdsByKind(source, "node");
+    if (nodeTaskIds.length > 0) {
+        throw new harnessUtils_1.BabysitterRuntimeError("InvalidProcessSourceError", `Process file at ${filePath} defines forbidden node tasks: ${nodeTaskIds.join(", ")}`, {
+            category: harnessUtils_1.ErrorCategory.Validation,
+            nextSteps: [
+                "Replace each node task with an agent or skill task",
+                "If the work is a concrete existing CLI command, use a shell task and have the orchestrating agent execute it and post the result",
+            ],
+        });
+    }
     const invalidCtxTaskTargets = getInvalidCtxTaskTargets(source);
     if (invalidCtxTaskTargets.length > 0) {
         throw new harnessUtils_1.BabysitterRuntimeError("InvalidProcessSourceError", `Process file at ${filePath} calls ctx.task(...) with values that are not DefinedTask bindings created via defineTask(...): ${invalidCtxTaskTargets.join(", ")}`, {
@@ -1019,16 +1035,21 @@ function coerceAgentResultValue(taskDef, output) {
 }
 // ── Process Recovery ─────────────────────────────────────────────────
 async function recoverProcessDefinitionFromOutputs(args) {
-    const expectedPath = path.resolve(args.outputPath);
+    const resolvedDir = path.resolve(args.outputDir);
+    // Check if any .js/.mjs files already exist in the output directory
     try {
-        await waitForProcessFile(expectedPath, 1_000);
-        return {
-            processPath: expectedPath,
-            summary: "Recovered from missing process-definition tool report by using the expected output path.",
-        };
+        const entries = await node_fs_1.promises.readdir(resolvedDir);
+        const processFiles = entries.filter((e) => /\.m?js$/.test(e));
+        if (processFiles.length > 0) {
+            const candidatePath = path.join(resolvedDir, processFiles[0]);
+            return {
+                processPath: candidatePath,
+                summary: "Recovered from missing process-definition tool report by scanning the output directory.",
+            };
+        }
     }
     catch {
-        // continue to other recovery strategies
+        // directory may not exist yet
     }
     for (const output of args.outputs) {
         for (const candidatePath of extractMentionedProcessPaths(output, args.workspace)) {
@@ -1049,10 +1070,12 @@ async function recoverProcessDefinitionFromOutputs(args) {
         if (!extracted || !looksLikeProcessDefinitionSource(extracted)) {
             continue;
         }
-        await node_fs_1.promises.mkdir(path.dirname(expectedPath), { recursive: true });
-        await node_fs_1.promises.writeFile(expectedPath, extracted, "utf8");
+        const recoveredName = `recovered-process-${Date.now()}.mjs`;
+        const recoveredPath = path.join(resolvedDir, recoveredName);
+        await node_fs_1.promises.mkdir(resolvedDir, { recursive: true });
+        await node_fs_1.promises.writeFile(recoveredPath, extracted, "utf8");
         return {
-            processPath: expectedPath,
+            processPath: recoveredPath,
             summary: "Recovered process-definition output by writing a JavaScript code block returned by the agent.",
         };
     }
@@ -1060,10 +1083,12 @@ async function recoverProcessDefinitionFromOutputs(args) {
         if (!looksLikeProcessDefinitionSource(output)) {
             continue;
         }
-        await node_fs_1.promises.mkdir(path.dirname(expectedPath), { recursive: true });
-        await node_fs_1.promises.writeFile(expectedPath, output.trim(), "utf8");
+        const recoveredName = `recovered-process-${Date.now()}.mjs`;
+        const recoveredPath = path.join(resolvedDir, recoveredName);
+        await node_fs_1.promises.mkdir(resolvedDir, { recursive: true });
+        await node_fs_1.promises.writeFile(recoveredPath, output.trim(), "utf8");
         return {
-            processPath: expectedPath,
+            processPath: recoveredPath,
             summary: "Recovered process-definition output by writing the agent's direct JavaScript response.",
         };
     }
@@ -1082,7 +1107,7 @@ async function recoverReportedProcessDefinition(args) {
         return args.state.report;
     }
     const recovered = await recoverProcessDefinitionFromOutputs({
-        outputPath: args.outputPath,
+        outputDir: args.outputDir,
         workspace: args.workspace,
         outputs: args.outputs,
     });
@@ -1125,7 +1150,7 @@ function buildExternalProcessDefinitionPrompt(args) {
         "Task:",
         `- User request: ${args.prompt}`,
         `- Workspace: ${workspace}`,
-        `- Output path: ${path.resolve(args.outputPath)}`,
+        `- Process output directory: ${path.resolve(args.outputDir)}`,
         `- Workspace assessment: ${args.workspaceAssessment.kind} (${workspaceSummary})`,
         "",
         "Requirements:",
@@ -1161,7 +1186,7 @@ function buildExternalProcessDefinitionPrompt(args) {
         "- External harnesses do not provide PI sandbox guardrails for their own tool execution. Keep security-sensitive shell work on the internal PI worker by using shell effects without routing them to an external harness.",
         "",
         "Output rules:",
-        "- Write the file now to the exact output path.",
+        `- Choose a descriptive kebab-case filename (e.g. "user-auth-tdd.mjs", "data-pipeline-setup.js") and write the file to the process output directory.`,
         "- Return a short summary that confirms what you wrote and the final path.",
         "- Do not rely on AskUserQuestion or babysitter_report_process_definition. Those tools are not available here.",
         "- Do not return pseudocode, placeholders, or a plan without writing the file.",
@@ -1195,7 +1220,7 @@ function buildExternalProcessConformancePrompt(args) {
         "- Put instructions inside `agent.prompt.task`, `agent.prompt.instructions`, and related prompt fields rather than top-level `instructions` fields.",
         "- Agent tasks must use `kind: \"agent\"` with `agent: { name, prompt, outputSchema }`.",
         "- Shell tasks must use `kind: \"shell\"` with `shell: { command: \"...\" }`.",
-        "- Node tasks must use `kind: \"node\"` with `node: { entry, args? }`.",
+        "- Do not introduce `kind: \"node\"` task definitions in generated or repaired processes. If logic would have been a node task, convert it to an `agent` or `skill` task instead.",
         "- Any task passed to `ctx.task(...)` must be a DefinedTask created via `defineTask(...)`; do not pass plain object task definitions or ad-hoc task objects.",
         "- The exported `process(inputs, ctx)` function must run tasks with `await ctx.task(definedTask, args)`; do not invent alternate task runners.",
         "- Inside the named `process(inputs, ctx)` export, never reference Node's global process object as `process.*`; use `globalThis.process` or an imported alias like `nodeProcess` instead.",
@@ -1229,9 +1254,9 @@ function buildInternalProcessConformancePrompt(args) {
         "- If the process needs the workspace root, do not assume `ctx.workspaceDir` or `ctx.cwd` exists. Resolve it from the module location using `import.meta.url`, for example with `path.dirname(fileURLToPath(import.meta.url))`.",
         "- Agent tasks must use `kind: \"agent\"` with `agent: { name, prompt, outputSchema }`.",
         "- Shell tasks must use `kind: \"shell\"` with `shell: { command: \"...\" }`.",
-        "- Node tasks must use `kind: \"node\"` with `node: { entry, args? }`.",
+        "- Do not introduce `kind: \"node\"` task definitions in generated or repaired processes. If logic would have been a node task, convert it to an `agent` or `skill` task instead.",
         "- The exported `process(inputs, ctx)` function must call tasks with `await ctx.task(definedTask, args)`.",
-        "- Use `babysitter_write_process_definition` to rewrite the full file to the exact target path.",
+        "- Use `babysitter_write_process_definition` with the `filename` parameter to rewrite the process file in the output directory.",
         "- After rewriting the file, call `babysitter_report_process_definition` exactly once with the same path.",
         "- Do not answer with plain text only.",
     ].join("\n");
@@ -1265,7 +1290,7 @@ async function _runExternalProcessDefinitionPhase(args) {
         (0, harnessUtils_1.writeVerboseBlock)(args.verbose, args.json, label, value, maxChars);
     };
     (0, harnessUtils_1.emitProgress)({ phase: "1", status: "started", harness: `${args.selectedHarnessName} (headless)` }, args.json, args.verbose);
-    writeVerbose(`[phase1 setup] workspace=${path.resolve(args.workspace ?? process.cwd())} model=${args.model ?? "(default)"} outputPath=${path.resolve(args.outputPath)} harness=${args.selectedHarnessName}`);
+    writeVerbose(`[phase1 setup] workspace=${path.resolve(args.workspace ?? process.cwd())} model=${args.model ?? "(default)"} outputDir=${path.resolve(args.outputDir)} harness=${args.selectedHarnessName}`);
     const workspaceAssessment = await assessWorkspaceForExternalAuthoring(args.workspace);
     writeVerboseData("phase1 workspace assessment", workspaceAssessment);
     const invokeProcessAuthor = async (label, prompt, timeout) => {
@@ -1285,13 +1310,13 @@ async function _runExternalProcessDefinitionPhase(args) {
     };
     await invokeProcessAuthor("phase1 initial", buildExternalProcessDefinitionPrompt({
         prompt: args.prompt,
-        outputPath: args.outputPath,
+        outputDir: args.outputDir,
         workspace: args.workspace,
         promptContext: args.promptContext,
         workspaceAssessment,
     }), 900_000);
     let report = await recoverProcessDefinitionFromOutputs({
-        outputPath: args.outputPath,
+        outputDir: args.outputDir,
         workspace: args.workspace,
         outputs: phaseOutputs,
     });
@@ -1300,12 +1325,13 @@ async function _runExternalProcessDefinitionPhase(args) {
     }
     if (!report) {
         await invokeProcessAuthor("phase1 recovery", [
-            `Write the full process file now to ${args.outputPath}.`,
+            `Write the full process file now to the output directory ${args.outputDir}.`,
+            "Choose a descriptive kebab-case filename (e.g. recovered-process.mjs).",
             "Do not describe the plan only; materialize the file in the workspace.",
             "After writing it, return either a concise summary or the full file in a ```javascript fenced block.",
         ].join("\n"), 300_000);
         report = await recoverProcessDefinitionFromOutputs({
-            outputPath: args.outputPath,
+            outputDir: args.outputDir,
             workspace: args.workspace,
             outputs: phaseOutputs,
         });
@@ -1315,12 +1341,12 @@ async function _runExternalProcessDefinitionPhase(args) {
     }
     if (!report) {
         await invokeProcessAuthor("phase1 final recovery", [
-            `Final recovery step: write the complete JavaScript process file to ${args.outputPath}.`,
+            `Final recovery step: write the complete JavaScript process file to the output directory ${args.outputDir}.`,
             "Return the full file in a ```javascript fenced block after it exists on disk.",
             "Do not omit the file write.",
         ].join("\n"), 300_000);
         report = await recoverProcessDefinitionFromOutputs({
-            outputPath: args.outputPath,
+            outputDir: args.outputDir,
             workspace: args.workspace,
             outputs: phaseOutputs,
         });
@@ -1366,27 +1392,29 @@ async function runProcessDefinitionPhase(args) {
         {
             name: "babysitter_write_process_definition",
             label: "Write Process Definition",
-            description: "Write the final JavaScript process file to the exact required output path.",
+            description: "Write the final JavaScript process file to the process output directory with a descriptive filename.",
             parameters: typebox_1.Type.Object({
                 source: typebox_1.Type.String(),
-                processPath: typebox_1.Type.Optional(typebox_1.Type.String()),
+                filename: typebox_1.Type.String(),
             }),
             execute: async (_toolCallId, params) => {
-                const resolvedOutputPath = path.resolve(args.outputPath);
-                const requestedPath = params.processPath
-                    ? path.resolve(params.processPath)
-                    : resolvedOutputPath;
-                if (requestedPath !== resolvedOutputPath) {
-                    throw new harnessUtils_1.BabysitterRuntimeError("ProcessDefinitionPathMismatch", `Process definitions must be written to ${resolvedOutputPath}.`, { category: harnessUtils_1.ErrorCategory.Runtime });
+                const filename = params.filename;
+                if (!/\.m?js$/.test(filename)) {
+                    throw new harnessUtils_1.BabysitterRuntimeError("ProcessDefinitionInvalidFilename", `Filename must end in .js or .mjs, got: ${filename}`, { category: harnessUtils_1.ErrorCategory.Validation });
+                }
+                if (/[/\\]/.test(filename) || filename.includes("..")) {
+                    throw new harnessUtils_1.BabysitterRuntimeError("ProcessDefinitionInvalidFilename", `Filename must be a simple name without path separators or '..', got: ${filename}`, { category: harnessUtils_1.ErrorCategory.Validation });
                 }
+                const resolvedDir = path.resolve(args.outputDir);
+                const resolvedPath = path.join(resolvedDir, filename);
                 const normalizedSource = normalizeProcessDefinitionSource(params.source);
                 writeVerboseData("phase1 tool babysitter_write_process_definition", {
-                    processPath: requestedPath,
+                    processPath: resolvedPath,
                     sourcePreview: (0, harnessUtils_1.truncateForVerboseLog)(normalizedSource, 600),
                 });
-                await node_fs_1.promises.mkdir(path.dirname(resolvedOutputPath), { recursive: true });
-                await node_fs_1.promises.writeFile(resolvedOutputPath, normalizedSource, "utf8");
-                return (0, harnessUtils_1.formatToolResult)({ processPath: resolvedOutputPath }, "Process definition written.");
+                await node_fs_1.promises.mkdir(resolvedDir, { recursive: true });
+                await node_fs_1.promises.writeFile(resolvedPath, normalizedSource, "utf8");
+                return (0, harnessUtils_1.formatToolResult)({ processPath: resolvedPath }, "Process definition written.");
             },
         },
         {
@@ -1495,15 +1523,15 @@ async function runProcessDefinitionPhase(args) {
     });
     const mergedCustomTools = [...customTools, ...agenticTools];
     (0, harnessUtils_1.emitProgress)({ phase: "1", status: "started", harness: "pi (agentic)" }, args.json, args.verbose);
-    writeVerbose(`[phase1 setup] workspace=${path.resolve(args.workspace ?? process.cwd())} model=${args.model ?? "(default)"} outputPath=${path.resolve(args.outputPath)}`);
+    writeVerbose(`[phase1 setup] workspace=${path.resolve(args.workspace ?? process.cwd())} model=${args.model ?? "(default)"} outputDir=${path.resolve(args.outputDir)}`);
     writeVerboseData("phase1 tools", mergedCustomTools.map((tool) => ({
         name: tool.name,
         label: tool.label,
     })));
     const workspaceAssessment = await assessWorkspaceForExternalAuthoring(args.workspace);
     writeVerboseData("phase1 workspace assessment", workspaceAssessment);
-    const processDefinitionSystemPrompt = (0, harnessPrompts_1.buildProcessDefinitionSystemPrompt)(args.outputPath, args.promptContext);
-    const initialMetaPrompt = (0, harnessPrompts_1.buildProcessDefinitionUserPrompt)(args.prompt, args.outputPath, {
+    const processDefinitionSystemPrompt = await (0, harnessPrompts_1.buildProcessDefinitionSystemPrompt)(args.outputDir, args.promptContext, args.interactive);
+    const initialMetaPrompt = (0, harnessPrompts_1.buildProcessDefinitionUserPrompt)(args.prompt, args.outputDir, {
         interactive: args.interactive,
         workspaceAssessment: workspaceAssessment.kind,
         workspaceEntries: workspaceAssessment.entries,
@@ -1540,10 +1568,18 @@ async function runProcessDefinitionPhase(args) {
         const result = await (0, harnessUtils_1.promptPiWithRetry)({
             session,
             message: initialMetaPrompt,
-            timeout: 300_000,
+            timeout: harnessUtils_1.PI_DEFAULT_PROMPT_TIMEOUT_MS,
             label: "phase1 initial",
             writeVerbose,
             writeVerboseData,
+        }).catch((err) => {
+            const isTimeout = err instanceof harnessUtils_1.BabysitterRuntimeError &&
+                (err.name === "PiTimeoutError" || (err.message ?? "").includes("timed out"));
+            if (isTimeout) {
+                writeVerbose("[phase1] Pi prompt timed out, converting to failure result");
+                return { success: false, output: `Pi prompt timed out: ${err instanceof Error ? err.message : String(err)}` };
+            }
+            throw err;
         });
         phaseOutputs.push(result.output);
         if (unsubscribe)
@@ -1554,7 +1590,7 @@ async function runProcessDefinitionPhase(args) {
             writeVerboseData("phase1 agent failure output", result.output);
             const recovered = await recoverReportedProcessDefinition({
                 state,
-                outputPath: args.outputPath,
+                outputDir: args.outputDir,
                 workspace: args.workspace,
                 outputs: phaseOutputs,
                 verbose: args.verbose,
@@ -1572,7 +1608,7 @@ async function runProcessDefinitionPhase(args) {
             writeVerboseProcessDefinitionRecovery(args.json);
             const recoveryPrompt = [
                 "Recovery step:",
-                `- Write the full process file now to ${args.outputPath}`,
+                `- Write the process file now to the output directory ${args.outputDir} using babysitter_write_process_definition with a descriptive filename.`,
                 "- Then call babysitter_report_process_definition exactly once.",
                 "- Do not just describe the process in plain text.",
             ].join("\n");
@@ -1584,13 +1620,21 @@ async function runProcessDefinitionPhase(args) {
                 label: "phase1 recovery",
                 writeVerbose,
                 writeVerboseData,
+            }).catch((err) => {
+                const isTimeout = err instanceof harnessUtils_1.BabysitterRuntimeError &&
+                    (err.name === "PiTimeoutError" || (err.message ?? "").includes("timed out"));
+                if (isTimeout) {
+                    writeVerbose("[phase1 recovery] Pi prompt timed out, converting to failure result");
+                    return { success: false, output: `Pi prompt timed out: ${err instanceof Error ? err.message : String(err)}` };
+                }
+                throw err;
             });
             phaseOutputs.push(recovery.output);
             if (!recovery.success) {
                 writeVerboseData("phase1 recovery failure output", recovery.output);
                 const recovered = await recoverReportedProcessDefinition({
                     state,
-                    outputPath: args.outputPath,
+                    outputDir: args.outputDir,
                     workspace: args.workspace,
                     outputs: phaseOutputs,
                     verbose: args.verbose,
@@ -1609,7 +1653,7 @@ async function runProcessDefinitionPhase(args) {
             writeVerbose("[phase1 recovery] attempting host-side recovery from agent outputs");
             await recoverReportedProcessDefinition({
                 state,
-                outputPath: args.outputPath,
+                outputDir: args.outputDir,
                 workspace: args.workspace,
                 outputs: phaseOutputs,
                 verbose: args.verbose,
@@ -1619,7 +1663,7 @@ async function runProcessDefinitionPhase(args) {
         if (!state.report?.processPath) {
             const finalRecoveryPrompt = [
                 "Final recovery step:",
-                `- Write the complete JavaScript process file to ${args.outputPath}.`,
+                `- Write the process file to the output directory ${args.outputDir} using babysitter_write_process_definition with a descriptive filename.`,
                 "- If you already wrote it, do not rewrite unnecessarily.",
                 "- Call babysitter_report_process_definition exactly once after the file exists.",
                 "- Do not answer with plain text only.",
@@ -1633,13 +1677,21 @@ async function runProcessDefinitionPhase(args) {
                 label: "phase1 final recovery",
                 writeVerbose,
                 writeVerboseData,
+            }).catch((err) => {
+                const isTimeout = err instanceof harnessUtils_1.BabysitterRuntimeError &&
+                    (err.name === "PiTimeoutError" || (err.message ?? "").includes("timed out"));
+                if (isTimeout) {
+                    writeVerbose("[phase1 final recovery] Pi prompt timed out, converting to failure result");
+                    return { success: false, output: `Pi prompt timed out: ${err instanceof Error ? err.message : String(err)}` };
+                }
+                throw err;
             });
             phaseOutputs.push(finalRecovery.output);
             if (!finalRecovery.success) {
                 writeVerboseData("phase1 final recovery failure output", finalRecovery.output);
                 const recovered = await recoverReportedProcessDefinition({
                     state,
-                    outputPath: args.outputPath,
+                    outputDir: args.outputDir,
                     workspace: args.workspace,
                     outputs: phaseOutputs,
                     verbose: args.verbose,
@@ -1655,7 +1707,7 @@ async function runProcessDefinitionPhase(args) {
             }
             await recoverReportedProcessDefinition({
                 state,
-                outputPath: args.outputPath,
+                outputDir: args.outputDir,
                 workspace: args.workspace,
                 outputs: phaseOutputs,
                 verbose: args.verbose,
@@ -1702,6 +1754,14 @@ async function runProcessDefinitionPhase(args) {
                     label: "phase1 conformance repair",
                     writeVerbose,
                     writeVerboseData,
+                }).catch((err) => {
+                    const isTimeout = err instanceof harnessUtils_1.BabysitterRuntimeError &&
+                        (err.name === "PiTimeoutError" || (err.message ?? "").includes("timed out"));
+                    if (isTimeout) {
+                        writeVerbose("[phase1 conformance repair] Pi prompt timed out, converting to failure result");
+                        return { success: false, output: `Pi prompt timed out: ${err instanceof Error ? err.message : String(err)}` };
+                    }
+                    throw err;
                 });
                 phaseOutputs.push(repair.output);
                 if (!repair.success) {

package/dist/cli/commands/harnessPhase2.d.ts

@@ -7,6 +7,12 @@
  */
 import * as readline from "node:readline";
 import { type ResolveEffectResult, type CompressionConfig, type HarnessDiscoveryResult, type HarnessPromptContext as SessionCreatePromptContext, type EffectAction, PiSessionHandle } from "./harnessUtils";
+/**
+ * Maximum number of consecutive Pi prompt timeouts before the orchestration
+ * run is failed.  A single timeout is recoverable — the loop simply continues
+ * — but repeated timeouts indicate a persistent infrastructure issue.
+ */
+export declare const MAX_CONSECUTIVE_TIMEOUTS = 3;
 export declare function resolveEffect(action: EffectAction, harnessName: string, options: {
     workspace?: string;
     model?: string;

package/dist/cli/commands/harnessPhase2.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"harnessPhase2.d.ts","sourceRoot":"","sources":["../../../src/cli/commands/harnessPhase2.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAGH,OAAO,KAAK,QAAQ,MAAM,eAAe,CAAC;AAkB1C,OAAO,EAOL,KAAK,mBAAmB,EACxB,KAAK,iBAAiB,EACtB,KAAK,sBAAsB,EAI3B,KAAK,oBAAoB,IAAI,0BAA0B,EACvD,KAAK,YAAY,EA8BjB,eAAe,EAIhB,MAAM,gBAAgB,CAAC;AAaxB,wBAAsB,aAAa,CACjC,MAAM,EAAE,YAAY,EACpB,WAAW,EAAE,MAAM,EACnB,OAAO,EAAE;IACP,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,WAAW,CAAC,EAAE,OAAO,CAAC;IACtB,iBAAiB,CAAC,EAAE,iBAAiB,GAAG,IAAI,CAAC;CAC9C,EACD,SAAS,CAAC,EAAE,eAAe,GAAG,IAAI,EAClC,UAAU,CAAC,EAAE,sBAAsB,EAAE,EACrC,EAAE,CAAC,EAAE,QAAQ,CAAC,SAAS,GAAG,IAAI,EAC9B,IAAI,CAAC,EAAE,OAAO,EACd,sBAAsB,CAAC,EAAE,CAAC,CAAC,MAAM,EAAE,OAAO,KAAK,OAAO,CAAC,OAAO,CAAC,CAAC,GAAG,IAAI,GACtE,OAAO,CAAC,mBAAmB,CAAC,CAwO9B;AA2ND,wBAAsB,qBAAqB,CAAC,IAAI,EAAE;IAChD,WAAW,EAAE,MAAM,CAAC;IACpB,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,OAAO,EAAE,MAAM,CAAC;IAChB,aAAa,EAAE,MAAM,CAAC;IACtB,IAAI,EAAE,OAAO,CAAC;IACd,OAAO,EAAE,OAAO,CAAC;IACjB,WAAW,EAAE,OAAO,CAAC;IACrB,EAAE,EAAE,QAAQ,CAAC,SAAS,GAAG,IAAI,CAAC;IAC9B,mBAAmB,EAAE,MAAM,CAAC;IAC5B,UAAU,EAAE,sBAAsB,EAAE,CAAC;IACrC,iBAAiB,EAAE,iBAAiB,GAAG,IAAI,CAAC;IAC5C,aAAa,EAAE,0BAA0B,CAAC;IAC1C,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,cAAc,CAAC,EAAE,MAAM,CAAC;CACzB,GAAG,OAAO,CAAC,MAAM,CAAC,CAw1ClB"}
+{"version":3,"file":"harnessPhase2.d.ts","sourceRoot":"","sources":["../../../src/cli/commands/harnessPhase2.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAGH,OAAO,KAAK,QAAQ,MAAM,eAAe,CAAC;AAkB1C,OAAO,EAOL,KAAK,mBAAmB,EACxB,KAAK,iBAAiB,EACtB,KAAK,sBAAsB,EAI3B,KAAK,oBAAoB,IAAI,0BAA0B,EACvD,KAAK,YAAY,EA+BjB,eAAe,EAIhB,MAAM,gBAAgB,CAAC;AAQxB;;;;GAIG;AACH,eAAO,MAAM,wBAAwB,IAAI,CAAC;AAO1C,wBAAsB,aAAa,CACjC,MAAM,EAAE,YAAY,EACpB,WAAW,EAAE,MAAM,EACnB,OAAO,EAAE;IACP,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,WAAW,CAAC,EAAE,OAAO,CAAC;IACtB,iBAAiB,CAAC,EAAE,iBAAiB,GAAG,IAAI,CAAC;CAC9C,EACD,SAAS,CAAC,EAAE,eAAe,GAAG,IAAI,EAClC,UAAU,CAAC,EAAE,sBAAsB,EAAE,EACrC,EAAE,CAAC,EAAE,QAAQ,CAAC,SAAS,GAAG,IAAI,EAC9B,IAAI,CAAC,EAAE,OAAO,EACd,sBAAsB,CAAC,EAAE,CAAC,CAAC,MAAM,EAAE,OAAO,KAAK,OAAO,CAAC,OAAO,CAAC,CAAC,GAAG,IAAI,GACtE,OAAO,CAAC,mBAAmB,CAAC,CAwO9B;AAkOD,wBAAsB,qBAAqB,CAAC,IAAI,EAAE;IAChD,WAAW,EAAE,MAAM,CAAC;IACpB,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,OAAO,EAAE,MAAM,CAAC;IAChB,aAAa,EAAE,MAAM,CAAC;IACtB,IAAI,EAAE,OAAO,CAAC;IACd,OAAO,EAAE,OAAO,CAAC;IACjB,WAAW,EAAE,OAAO,CAAC;IACrB,EAAE,EAAE,QAAQ,CAAC,SAAS,GAAG,IAAI,CAAC;IAC9B,mBAAmB,EAAE,MAAM,CAAC;IAC5B,UAAU,EAAE,sBAAsB,EAAE,CAAC;IACrC,iBAAiB,EAAE,iBAAiB,GAAG,IAAI,CAAC;IAC5C,aAAa,EAAE,0BAA0B,CAAC;IAC1C,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,cAAc,CAAC,EAAE,MAAM,CAAC;CACzB,GAAG,OAAO,CAAC,MAAM,CAAC,CA05ClB"}