@a5c-ai/babysitter-sdk 0.0.183-staging.dd5dc0db → 0.0.183-staging.e7446f7c

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

@@ -0,0 +1,8 @@
+import type { HarnessInstallOptions } from "../../harness";
+export interface HarnessInstallCommandArgs extends HarnessInstallOptions {
+    harnessName?: string;
+}
+export declare function handleHarnessInstall(args: HarnessInstallCommandArgs): Promise<number>;
+export declare function handleHarnessInstallPlugin(args: HarnessInstallCommandArgs): Promise<number>;
+export declare function formatHarnessInstallError(error: unknown, json: boolean): number;
+//# sourceMappingURL=harnessInstall.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"harnessInstall.d.ts","sourceRoot":"","sources":["../../../src/cli/commands/harnessInstall.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAEV,qBAAqB,EAEtB,MAAM,eAAe,CAAC;AAQvB,MAAM,WAAW,yBAA0B,SAAQ,qBAAqB;IACtE,WAAW,CAAC,EAAE,MAAM,CAAC;CACtB;AA0FD,wBAAsB,oBAAoB,CAAC,IAAI,EAAE,yBAAyB,GAAG,OAAO,CAAC,MAAM,CAAC,CAM3F;AAED,wBAAsB,0BAA0B,CAAC,IAAI,EAAE,yBAAyB,GAAG,OAAO,CAAC,MAAM,CAAC,CAMjG;AAED,wBAAgB,yBAAyB,CAAC,KAAK,EAAE,OAAO,EAAE,IAAI,EAAE,OAAO,GAAG,MAAM,CAW/E"}

package/dist/cli/commands/harnessInstall.js

@@ -0,0 +1,94 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.handleHarnessInstall = handleHarnessInstall;
+exports.handleHarnessInstallPlugin = handleHarnessInstallPlugin;
+exports.formatHarnessInstallError = formatHarnessInstallError;
+const harness_1 = require("../../harness");
+const exceptions_1 = require("../../runtime/exceptions");
+function supportsColors() {
+    if (process.env.NO_COLOR !== undefined)
+        return false;
+    if (process.env.FORCE_COLOR !== undefined)
+        return true;
+    return Boolean(process.stderr?.isTTY);
+}
+function formatInstallResult(payload, json) {
+    if (json) {
+        console.log(JSON.stringify(payload, null, 2));
+        return;
+    }
+    if (payload.warning) {
+        console.log(`Warning: ${payload.warning}`);
+    }
+    if (payload.summary) {
+        console.log(payload.summary);
+    }
+    if (payload.command) {
+        console.log(`Command: ${payload.command}`);
+    }
+    if (payload.output) {
+        console.log(payload.output);
+    }
+    if (payload.location) {
+        console.log(`Location: ${payload.location}`);
+    }
+}
+function requireHarnessName(harnessName, commandName) {
+    if (!harnessName) {
+        throw new exceptions_1.BabysitterRuntimeError("MissingArgument", `${commandName} requires a harness name as the first argument`, {
+            category: exceptions_1.ErrorCategory.Validation,
+            suggestions: [`babysitter ${commandName} codex`],
+        });
+    }
+    return harnessName;
+}
+function requireInstallMethod(harnessName, commandName, kind) {
+    const adapter = (0, harness_1.getAdapterByName)(harnessName);
+    if (!adapter) {
+        throw new exceptions_1.BabysitterRuntimeError("UnsupportedHarnessInstall", `${commandName} does not support "${harnessName}" yet.`, {
+            category: exceptions_1.ErrorCategory.Validation,
+            suggestions: [
+                "Supported harnesses: claude-code, codex, gemini-cli, pi, oh-my-pi",
+            ],
+        });
+    }
+    const run = adapter[kind];
+    if (!run) {
+        throw new exceptions_1.BabysitterRuntimeError("UnsupportedHarnessInstall", `${commandName} does not support "${harnessName}" yet.`, {
+            category: exceptions_1.ErrorCategory.Validation,
+            suggestions: [
+                "Supported harnesses: claude-code, codex, gemini-cli, pi, oh-my-pi",
+            ],
+        });
+    }
+    return {
+        adapter,
+        run: run.bind(adapter),
+    };
+}
+async function handleHarnessInstall(args) {
+    const harnessName = requireHarnessName(args.harnessName, "harness:install");
+    const { run } = requireInstallMethod(harnessName, "harness:install", "installHarness");
+    const result = await run(args);
+    formatInstallResult(result, args.json);
+    return 0;
+}
+async function handleHarnessInstallPlugin(args) {
+    const harnessName = requireHarnessName(args.harnessName, "harness:install-plugin");
+    const { run } = requireInstallMethod(harnessName, "harness:install-plugin", "installPlugin");
+    const result = await run(args);
+    formatInstallResult(result, args.json);
+    return 0;
+}
+function formatHarnessInstallError(error, json) {
+    const err = error instanceof Error
+        ? error
+        : new Error(String(error));
+    if (json) {
+        console.error(JSON.stringify((0, exceptions_1.toStructuredError)(err), null, 2));
+    }
+    else {
+        console.error((0, exceptions_1.formatErrorWithContext)(err, { colors: supportsColors() }));
+    }
+    return 1;
+}

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

@@ -1 +1 @@
-{"version":3,"file":"sessionCreate.d.ts","sourceRoot":"","sources":["../../../src/cli/commands/sessionCreate.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAWH,OAAO,KAAK,EACV,sBAAsB,EAGvB,MAAM,qBAAqB,CAAC;AAwB7B,MAAM,WAAW,iBAAiB;IAChC,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,OAAO,EAAE,MAAM,CAAC;IAChB,IAAI,EAAE,OAAO,CAAC;IACd,OAAO,EAAE,OAAO,CAAC;IACjB,WAAW,CAAC,EAAE,OAAO,CAAC;CACvB;AAqGD,wBAAgB,aAAa,CAC3B,UAAU,EAAE,sBAAsB,EAAE,EACpC,SAAS,CAAC,EAAE,MAAM,GACjB,sBAAsB,GAAG,SAAS,CAYpC;AA0rCD,wBAAsB,mBAAmB,CACvC,MAAM,EAAE,iBAAiB,GACxB,OAAO,CAAC,MAAM,CAAC,CAsEjB"}
+{"version":3,"file":"sessionCreate.d.ts","sourceRoot":"","sources":["../../../src/cli/commands/sessionCreate.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAWH,OAAO,KAAK,EACV,sBAAsB,EAIvB,MAAM,qBAAqB,CAAC;AA0B7B,MAAM,WAAW,iBAAiB;IAChC,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,OAAO,EAAE,MAAM,CAAC;IAChB,IAAI,EAAE,OAAO,CAAC;IACd,OAAO,EAAE,OAAO,CAAC;IACjB,WAAW,CAAC,EAAE,OAAO,CAAC;CACvB;AA+ID,wBAAgB,aAAa,CAC3B,UAAU,EAAE,sBAAsB,EAAE,EACpC,SAAS,CAAC,EAAE,MAAM,GACjB,sBAAsB,GAAG,SAAS,CAYpC;AA2hDD,wBAAsB,mBAAmB,CACvC,MAAM,EAAE,iBAAiB,GACxB,OAAO,CAAC,MAAM,CAAC,CAiFjB"}

package/dist/cli/commands/sessionCreate.js

@@ -91,6 +91,38 @@ function compressInternalHarnessPrompt(text, compressionConfig, taskKind = "agen
     const targetReduction = layer.perTaskKind?.[taskKind] ?? layer.targetReduction;
     return (0, density_filter_1.densityFilterText)(text, targetReduction);
 }
+function buildPromptContext(args) {
+    const envNames = [
+        "CI",
+        "BABYSITTER_COMPRESSION_ENABLED",
+        "BABYSITTER_PI_SANDBOX_IMAGE",
+        "AZURE_OPENAI_API_KEY",
+        "AZURE_OPENAI_PROJECT_NAME",
+        "AZURE_OPENAI_DEPLOYMENT",
+        "OPENAI_API_KEY",
+        "ANTHROPIC_API_KEY",
+    ];
+    return {
+        platform: process.platform,
+        arch: process.arch,
+        nodeVersion: process.version,
+        cwd: process.cwd(),
+        workspace: path.resolve(args.workspace ?? process.cwd()),
+        selectedHarnessName: args.selectedHarnessName,
+        discoveredHarnesses: args.discovered,
+        compressionEnabled: Boolean(args.compressionConfig?.enabled &&
+            args.compressionConfig.layers.sdkContextHook.enabled),
+        secureSandboxImage: process.env.BABYSITTER_PI_SANDBOX_IMAGE || "node:22-bookworm",
+        piDefaultBashSandbox: "local",
+        piIsolationDefault: false,
+        envFlags: envNames.map((name) => ({
+            name,
+            value: process.env[name]
+                ? (name.endsWith("_API_KEY") ? "set" : process.env[name])
+                : "unset",
+        })),
+    };
+}
 const HARNESS_PRIORITY = [
     "pi",
     "oh-my-pi",
@@ -134,6 +166,66 @@ function buildMetaPrompt(userPrompt, workDir) {
     template = template.replace("{{INTERVIEW_CONTEXT}}", "");
     return template;
 }
+function buildFallbackProcessDefinitionSource(userPrompt) {
+    const serializedPrompt = JSON.stringify(userPrompt);
+    return `import { defineTask } from "@a5c-ai/babysitter-sdk";
+
+const implementRequest = defineTask("implement-request", (args) => ({
+  kind: "agent",
+  title: "Implement the requested work",
+  agent: {
+    prompt: {
+      role: "delivery engineer",
+      task: "Implement the user's request in the current workspace.",
+      context: {
+        request: args.request,
+      },
+      instructions: [
+        "Create or update the necessary project files in the current workspace.",
+        "Keep the implementation as small and direct as possible while still satisfying the request.",
+        "Run lightweight verification where practical and include what you verified in the result.",
+        "Return JSON with summary, changedFiles, and verificationNotes."
+      ],
+      outputFormat: "JSON with summary, changedFiles, verificationNotes"
+    }
+  }
+}));
+
+const verifyRequest = defineTask("verify-request", (args) => ({
+  kind: "agent",
+  title: "Verify the requested work",
+  agent: {
+    prompt: {
+      role: "verification engineer",
+      task: "Verify that the user's request was completed and identify any remaining gaps.",
+      context: {
+        request: args.request,
+        implementation: args.implementation
+      },
+      instructions: [
+        "Inspect the workspace state produced by the implementation step.",
+        "Confirm whether the requested work exists and is usable.",
+        "If there are remaining issues, describe them precisely.",
+        "Return JSON with success, findings, and recommendedNextSteps."
+      ],
+      outputFormat: "JSON with success, findings, recommendedNextSteps"
+    }
+  }
+}));
+
+export async function process(inputs, ctx) {
+  const request = inputs?.prompt || ${serializedPrompt};
+  const implementation = await ctx.task(implementRequest, { request });
+  const verification = await ctx.task(verifyRequest, { request, implementation });
+  return {
+    request,
+    implementation,
+    verification,
+    success: verification?.success !== false,
+  };
+}
+`;
+}
 const POLL_INTERVAL_MS = 1_000;
 const POLL_TIMEOUT_MS = 60_000;
 async function waitForProcessFile(filePath, timeoutMs = POLL_TIMEOUT_MS) {
@@ -258,6 +350,43 @@ function resolveTaskHarness(action, defaultHarness, discovered) {
     }
     return defaultHarness;
 }
+function isPiHarness(harnessName) {
+    return harnessName === "pi" || harnessName === "oh-my-pi";
+}
+function readBooleanMetadata(metadata, key) {
+    const value = metadata?.[key];
+    if (typeof value === "boolean") {
+        return value;
+    }
+    if (value === "true") {
+        return true;
+    }
+    if (value === "false") {
+        return false;
+    }
+    return undefined;
+}
+function readBashSandboxMetadata(metadata) {
+    const value = metadata?.bashSandbox;
+    return value === "auto" || value === "secure" || value === "local"
+        ? value
+        : undefined;
+}
+function buildPiWorkerSessionOptions(args) {
+    const metadata = args.action.taskDef?.metadata;
+    const isolated = readBooleanMetadata(metadata, "isolated");
+    const enableCompaction = readBooleanMetadata(metadata, "enableCompaction");
+    const bashSandbox = readBashSandboxMetadata(metadata);
+    return {
+        workspace: args.workspace,
+        model: args.model,
+        toolsMode: "coding",
+        ephemeral: true,
+        ...(isolated !== undefined ? { isolated } : {}),
+        ...(enableCompaction !== undefined ? { enableCompaction } : {}),
+        ...(bashSandbox ? { bashSandbox } : {}),
+    };
+}
 async function resolveEffect(action, harnessName, options, piSession, discovered, rl, json) {
     const kind = action.kind;
     if (kind === "node" || kind === "orchestrator_task") {
@@ -526,6 +655,11 @@ function emitProgress(payload, json, verbose) {
             break;
     }
 }
+function writeVerboseProcessDefinitionRecovery(json) {
+    if (!json) {
+        process.stderr.write(`${DIM}Phase 1 recovery: the agent did not report the process file, retrying with an explicit write-and-report instruction...${RESET}\n`);
+    }
+}
 async function runProcessDefinitionPhase(args) {
     const state = {};
     const customTools = [
@@ -568,11 +702,8 @@ async function runProcessDefinitionPhase(args) {
         model: args.model,
         toolsMode: "coding",
         customTools,
-        appendSystemPrompt: [(0, sessionCreatePrompts_1.buildProcessDefinitionSystemPrompt)(args.outputPath)],
-        isolated: true,
+        appendSystemPrompt: [(0, sessionCreatePrompts_1.buildProcessDefinitionSystemPrompt)(args.outputPath, args.promptContext)],
         ephemeral: true,
-        bashSandbox: "secure",
-        enableCompaction: true,
     });
     try {
         await session.initialize();
@@ -596,7 +727,38 @@ async function runProcessDefinitionPhase(args) {
             throw new exceptions_1.BabysitterRuntimeError("ProcessDefinitionFailed", result.output, { category: exceptions_1.ErrorCategory.External });
         }
         if (!state.report?.processPath) {
-            throw new exceptions_1.BabysitterRuntimeError("ProcessDefinitionReportMissing", "The process-definition agent finished without calling babysitter_report_process_definition.", { category: exceptions_1.ErrorCategory.Runtime });
+            writeVerboseProcessDefinitionRecovery(args.json);
+            const recovery = await session.prompt(compressInternalHarnessPrompt([
+                "Recovery step:",
+                `- Write the full process file now to ${args.outputPath}`,
+                "- Then call babysitter_report_process_definition exactly once.",
+                "- Do not just describe the process in plain text.",
+            ].join("\n"), args.compressionConfig, "agent"), 180_000);
+            if (!recovery.success) {
+                throw new exceptions_1.BabysitterRuntimeError("ProcessDefinitionFailed", recovery.output, { category: exceptions_1.ErrorCategory.External });
+            }
+        }
+        if (!state.report?.processPath) {
+            try {
+                await waitForProcessFile(args.outputPath, 5_000);
+                state.report = {
+                    processPath: args.outputPath,
+                    summary: "Recovered from missing process-definition tool report by using the expected output path.",
+                };
+            }
+            catch {
+                if (!args.interactive) {
+                    const fallbackSource = buildFallbackProcessDefinitionSource(args.prompt);
+                    await node_fs_1.promises.writeFile(args.outputPath, fallbackSource, "utf8");
+                    state.report = {
+                        processPath: args.outputPath,
+                        summary: "Recovered from phase-1 tool failure by writing a generic multi-step fallback process definition.",
+                    };
+                }
+                else {
+                    throw new exceptions_1.BabysitterRuntimeError("ProcessDefinitionReportMissing", "The process-definition agent finished without calling babysitter_report_process_definition.", { category: exceptions_1.ErrorCategory.Runtime });
+                }
+            }
         }
         await waitForProcessFile(state.report.processPath);
         await validateProcessExport(state.report.processPath);
@@ -629,6 +791,33 @@ async function runOrchestrationPhase(args) {
         pendingEffectResults: new Map(),
     };
     let orchestrationSession = null;
+    const writeVerbose = (message) => {
+        if (!args.json && args.verbose) {
+            process.stderr.write(`${DIM}${message}${RESET}\n`);
+        }
+    };
+    const summarizeAgentText = (text) => {
+        const normalized = text.replace(/\s+/g, " ").trim();
+        if (!normalized) {
+            return "(no summary emitted)";
+        }
+        return normalized.length > 240 ? `${normalized.slice(0, 237)}...` : normalized;
+    };
+    const describePendingActions = () => Array.from(state.pendingActions.values()).map((action) => ({
+        effectId: action.effectId,
+        kind: action.kind,
+        title: action.taskDef?.title,
+        harness: resolveTaskHarness(action, args.selectedHarnessName, args.discovered),
+    }));
+    const ensureTerminalResult = () => {
+        if (state.lastIterationResult?.status === "completed") {
+            return 0;
+        }
+        if (state.lastIterationResult?.status === "failed") {
+            return 1;
+        }
+        return null;
+    };
     const customTools = [
         {
             name: "AskUserQuestion",
@@ -818,16 +1007,12 @@ async function runOrchestrationPhase(args) {
                 }
                 const taskHarness = resolveTaskHarness(action, args.selectedHarnessName, args.discovered);
                 let workerSession = null;
-                if (action.kind === "shell" || taskHarness === "pi" || taskHarness === "oh-my-pi") {
-                    workerSession = (0, piWrapper_1.createPiSession)({
+                if (action.kind === "shell" || isPiHarness(taskHarness)) {
+                    workerSession = (0, piWrapper_1.createPiSession)(buildPiWorkerSessionOptions({
+                        action,
                         workspace: args.workspace,
                         model: args.model,
-                        toolsMode: "coding",
-                        isolated: true,
-                        ephemeral: true,
-                        bashSandbox: "secure",
-                        enableCompaction: true,
-                    });
+                    }));
                 }
                 try {
                     const effectResult = await resolveEffect(action, args.selectedHarnessName, {
@@ -935,22 +1120,69 @@ async function runOrchestrationPhase(args) {
             },
         },
     ];
+    const tools = customTools;
+    const getTool = (name) => tools.find((tool) => tool.name === name);
+    const runCreateTool = getTool("babysitter_run_create");
+    const bindSessionTool = getTool("babysitter_bind_session");
+    const runIterateTool = getTool("babysitter_run_iterate");
+    const executeEffectTool = getTool("babysitter_execute_effect");
+    const taskPostTool = getTool("babysitter_task_post_result");
+    const finishTool = getTool("babysitter_finish_orchestration");
+    const invokeTool = async (tool, name, params = {}) => {
+        if (!tool?.execute) {
+            throw new exceptions_1.BabysitterRuntimeError("MissingSessionCreateTool", `Required orchestration tool is unavailable: ${name}`, { category: exceptions_1.ErrorCategory.Internal });
+        }
+        const result = tool.execute(`host-${name}`, params);
+        return await Promise.resolve(result);
+    };
+    const runDeterministicPendingEffects = async (reason) => {
+        const pending = describePendingActions();
+        if (pending.length === 0) {
+            return;
+        }
+        writeVerbose(`[phase2 fallback] ${reason}`);
+        for (const effect of pending) {
+            writeVerbose(`[phase2 fallback] resolving ${effect.effectId} (${effect.kind}) via ${effect.harness ?? args.selectedHarnessName}`);
+            const action = state.pendingActions.get(effect.effectId);
+            if (!action) {
+                continue;
+            }
+            if (action.kind !== "breakpoint") {
+                await invokeTool(executeEffectTool, "babysitter_execute_effect", { effectId: effect.effectId });
+            }
+            await invokeTool(taskPostTool, "babysitter_task_post_result", { effectId: effect.effectId });
+        }
+    };
+    const promptOrchestrationAgent = async (message, options) => {
+        if (!orchestrationSession) {
+            throw new exceptions_1.BabysitterRuntimeError("OrchestrationSessionMissing", "The orchestration PI session has not been created.", { category: exceptions_1.ErrorCategory.Runtime });
+        }
+        if (!args.json && args.verbose) {
+            const label = options?.label ?? "phase2";
+            process.stderr.write(`\n${DIM}[${label}] agent turn${RESET}\n`);
+        }
+        const result = await orchestrationSession.prompt(compressInternalHarnessPrompt(message, args.compressionConfig, "agent"), 900_000);
+        if (!result.success) {
+            throw new exceptions_1.BabysitterRuntimeError("OrchestrationAgentFailed", result.output, { category: exceptions_1.ErrorCategory.External });
+        }
+        writeVerbose(`[phase2 agent] ${summarizeAgentText(result.output)}`);
+        if (options?.fallbackReason && state.pendingActions.size > 0) {
+            await runDeterministicPendingEffects(options.fallbackReason);
+        }
+    };
     orchestrationSession = (0, piWrapper_1.createPiSession)({
         workspace: args.workspace,
         model: args.model,
         toolsMode: "readonly",
         customTools,
-        appendSystemPrompt: [(0, sessionCreatePrompts_1.buildOrchestrationSystemPrompt)(args.selectedHarnessName)],
-        isolated: true,
+        appendSystemPrompt: [(0, sessionCreatePrompts_1.buildOrchestrationSystemPrompt)(args.selectedHarnessName, args.promptContext)],
         ephemeral: true,
-        bashSandbox: "secure",
-        enableCompaction: true,
     });
     emitProgress({ phase: "2", status: "started", harness: args.selectedHarnessName }, args.json, args.verbose);
+    let unsubscribe = null;
     try {
         await orchestrationSession.initialize();
-        let unsubscribe = null;
-        if (!args.json) {
+        if (!args.json && args.verbose) {
             unsubscribe = orchestrationSession.subscribe((event) => {
                 if (event.type === "text_delta") {
                     const text = event.text;
@@ -959,27 +1191,79 @@ async function runOrchestrationPhase(args) {
                 }
             });
         }
-        const result = await orchestrationSession.prompt(compressInternalHarnessPrompt((0, sessionCreatePrompts_1.buildOrchestrationUserPrompt)(path.resolve(args.processPath), args.prompt, args.maxIterations), args.compressionConfig, "agent"), 900_000);
-        if (unsubscribe)
-            unsubscribe();
-        if (!args.json)
-            process.stderr.write("\n");
-        if (!result.success) {
-            throw new exceptions_1.BabysitterRuntimeError("OrchestrationAgentFailed", result.output, { category: exceptions_1.ErrorCategory.External });
+        await promptOrchestrationAgent((0, sessionCreatePrompts_1.buildOrchestrationBootstrapPrompt)(path.resolve(args.processPath), args.prompt, args.maxIterations), { label: "phase2 bootstrap" });
+        if (!state.runId || !state.runDir) {
+            writeVerbose("[phase2 fallback] agent did not create a run; creating it in the host");
+            await invokeTool(runCreateTool, "babysitter_run_create", args.prompt ? { prompt: args.prompt } : {});
+        }
+        if (!state.sessionBound) {
+            writeVerbose("[phase2 fallback] agent did not bind the session; binding in the host");
+            await invokeTool(bindSessionTool, "babysitter_bind_session");
         }
         if (!state.runId || !state.runDir) {
-            throw new exceptions_1.BabysitterRuntimeError("RunNotCreated", "The orchestration agent exited without creating a run.", { category: exceptions_1.ErrorCategory.Runtime });
+            throw new exceptions_1.BabysitterRuntimeError("RunNotCreated", "The orchestration session could not establish a run after bootstrap.", { category: exceptions_1.ErrorCategory.Runtime });
         }
-        if (!state.finished) {
-            throw new exceptions_1.BabysitterRuntimeError("OrchestrationFinishMissing", "The orchestration agent exited without calling babysitter_finish_orchestration.", { category: exceptions_1.ErrorCategory.Runtime });
+        while (state.iteration < args.maxIterations) {
+            const terminal = ensureTerminalResult();
+            if (terminal !== null) {
+                break;
+            }
+            const iterationBeforeTurn = state.iteration;
+            await promptOrchestrationAgent((0, sessionCreatePrompts_1.buildOrchestrationTurnPrompt)({
+                processPath: path.resolve(args.processPath),
+                userPrompt: args.prompt,
+                maxIterations: args.maxIterations,
+                currentIteration: state.iteration,
+                runId: state.runId,
+                runDir: state.runDir,
+                lastStatus: state.lastIterationResult?.status,
+                pendingEffects: describePendingActions(),
+            }), {
+                label: `phase2 iteration ${state.iteration + 1}`,
+                fallbackReason: "Agent left pending effects unresolved; the host resolved and posted them to keep the orchestration loop moving.",
+            });
+            if (ensureTerminalResult() !== null) {
+                break;
+            }
+            if (state.iteration === iterationBeforeTurn && state.pendingActions.size === 0) {
+                writeVerbose("[phase2 fallback] agent did not advance the run; executing one iteration in the host");
+                await invokeTool(runIterateTool, "babysitter_run_iterate");
+            }
+            if (state.lastIterationResult?.status === "waiting" && state.pendingActions.size > 0) {
+                await runDeterministicPendingEffects("Agent left pending effects unresolved after the turn; the host completed them deterministically.");
+            }
+            const postTurnTerminal = ensureTerminalResult();
+            if (postTurnTerminal !== null) {
+                break;
+            }
         }
-        if (state.lastIterationResult?.status === "completed") {
-            return 0;
+        if (state.lastIterationResult?.status !== "completed" && state.lastIterationResult?.status !== "failed") {
+            state.lastIterationResult = {
+                status: "failed",
+                error: { message: `Max iterations (${args.maxIterations}) reached without completion` },
+            };
+            emitProgress({
+                phase: "2",
+                status: "failed",
+                runId: state.runId,
+                runDir: state.runDir,
+                iteration: state.iteration,
+                runStatus: "failed",
+                error: `Max iterations (${args.maxIterations}) reached without completion`,
+            }, args.json, args.verbose);
         }
-        if (state.lastIterationResult?.status === "failed") {
-            return 1;
+        if (state.lastIterationResult?.status === "completed" || state.lastIterationResult?.status === "failed") {
+            await invokeTool(finishTool, "babysitter_finish_orchestration", {
+                summary: state.lastIterationResult.status === "completed"
+                    ? `Run ${state.runId} completed after ${state.iteration} iterations.`
+                    : `Run ${state.runId} failed after ${state.iteration} iterations.`,
+            });
         }
-        throw new exceptions_1.BabysitterRuntimeError("OrchestrationIncomplete", "The orchestration agent exited before the run reached a terminal state.", { category: exceptions_1.ErrorCategory.Runtime });
+        const exitCode = ensureTerminalResult();
+        if (exitCode !== null) {
+            return exitCode;
+        }
+        throw new exceptions_1.BabysitterRuntimeError("OrchestrationIncomplete", "The orchestration phase ended without a terminal run state.", { category: exceptions_1.ErrorCategory.Runtime });
     }
     catch (error) {
         emitProgress({
@@ -993,6 +1277,10 @@ async function runOrchestrationPhase(args) {
         return 1;
     }
     finally {
+        if (unsubscribe)
+            unsubscribe();
+        if (!args.json && args.verbose)
+            process.stderr.write("\n");
         orchestrationSession?.dispose();
     }
 }
@@ -1013,8 +1301,16 @@ async function handleSessionCreate(parsed) {
         }
         const discovered = await (0, discovery_1.discoverHarnesses)();
         const selected = selectHarness(discovered, preferredHarness);
-        const selectedHarnessName = selected?.name ?? "pi";
+        const selectedHarnessName = preferredHarness === "pi" || preferredHarness === "oh-my-pi"
+            ? preferredHarness
+            : (selected?.name ?? "pi");
         const compressionConfig = loadSessionCompressionConfig(workspace);
+        const promptContext = buildPromptContext({
+            workspace,
+            selectedHarnessName,
+            discovered,
+            compressionConfig,
+        });
         let processPath = providedProcessPath;
         if (processPath) {
             emitProgress({ phase: "1", status: "skipped", processPath }, json, verbose);
@@ -1031,6 +1327,7 @@ async function handleSessionCreate(parsed) {
                 json,
                 verbose,
                 compressionConfig,
+                promptContext,
             });
         }
         return await runOrchestrationPhase({
@@ -1047,6 +1344,7 @@ async function handleSessionCreate(parsed) {
             selectedHarnessName,
             discovered,
             compressionConfig,
+            promptContext,
         });
     }
     finally {

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

@@ -1,5 +1,39 @@
-export declare function buildProcessDefinitionSystemPrompt(outputPath: string): string;
+import type { HarnessDiscoveryResult } from "../../harness/types";
+export interface SessionCreatePromptContext {
+    platform: string;
+    arch: string;
+    nodeVersion: string;
+    cwd: string;
+    workspace: string;
+    selectedHarnessName?: string;
+    discoveredHarnesses: HarnessDiscoveryResult[];
+    compressionEnabled: boolean;
+    secureSandboxImage: string;
+    piDefaultBashSandbox: "local" | "auto" | "secure";
+    piIsolationDefault: boolean;
+    envFlags: Array<{
+        name: string;
+        value: string;
+    }>;
+}
+export declare function buildProcessDefinitionSystemPrompt(outputPath: string, context: SessionCreatePromptContext): string;
 export declare function buildProcessDefinitionUserPrompt(userPrompt: string, outputPath: string): string;
-export declare function buildOrchestrationSystemPrompt(selectedHarnessName: string): string;
-export declare function buildOrchestrationUserPrompt(processPath: string, userPrompt: string | undefined, maxIterations: number): string;
+export declare function buildOrchestrationSystemPrompt(selectedHarnessName: string, context: SessionCreatePromptContext): string;
+export declare function buildOrchestrationBootstrapPrompt(processPath: string, userPrompt: string | undefined, maxIterations: number): string;
+export declare function buildOrchestrationTurnPrompt(args: {
+    processPath: string;
+    userPrompt?: string;
+    maxIterations: number;
+    currentIteration: number;
+    runId?: string;
+    runDir?: string;
+    lastStatus?: string;
+    pendingEffects?: Array<{
+        effectId: string;
+        kind: string;
+        title?: string;
+        harness?: string;
+    }>;
+    fallbackReason?: string;
+}): string;
 //# sourceMappingURL=sessionCreatePrompts.d.ts.map

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

@@ -1 +1 @@
-{"version":3,"file":"sessionCreatePrompts.d.ts","sourceRoot":"","sources":["../../../src/cli/commands/sessionCreatePrompts.ts"],"names":[],"mappings":"AAAA,wBAAgB,kCAAkC,CAAC,UAAU,EAAE,MAAM,GAAG,MAAM,CAe7E;AAED,wBAAgB,gCAAgC,CAC9C,UAAU,EAAE,MAAM,EAClB,UAAU,EAAE,MAAM,GACjB,MAAM,CAOR;AAED,wBAAgB,8BAA8B,CAAC,mBAAmB,EAAE,MAAM,GAAG,MAAM,CAmBlF;AAED,wBAAgB,4BAA4B,CAC1C,WAAW,EAAE,MAAM,EACnB,UAAU,EAAE,MAAM,GAAG,SAAS,EAC9B,aAAa,EAAE,MAAM,GACpB,MAAM,CAQR"}
+{"version":3,"file":"sessionCreatePrompts.d.ts","sourceRoot":"","sources":["../../../src/cli/commands/sessionCreatePrompts.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,sBAAsB,EAAE,MAAM,qBAAqB,CAAC;AAElE,MAAM,WAAW,0BAA0B;IACzC,QAAQ,EAAE,MAAM,CAAC;IACjB,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,EAAE,MAAM,CAAC;IACpB,GAAG,EAAE,MAAM,CAAC;IACZ,SAAS,EAAE,MAAM,CAAC;IAClB,mBAAmB,CAAC,EAAE,MAAM,CAAC;IAC7B,mBAAmB,EAAE,sBAAsB,EAAE,CAAC;IAC9C,kBAAkB,EAAE,OAAO,CAAC;IAC5B,kBAAkB,EAAE,MAAM,CAAC;IAC3B,oBAAoB,EAAE,OAAO,GAAG,MAAM,GAAG,QAAQ,CAAC;IAClD,kBAAkB,EAAE,OAAO,CAAC;IAC5B,QAAQ,EAAE,KAAK,CAAC;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;CAClD;AA+ED,wBAAgB,kCAAkC,CAChD,UAAU,EAAE,MAAM,EAClB,OAAO,EAAE,0BAA0B,GAClC,MAAM,CAkBR;AAED,wBAAgB,gCAAgC,CAC9C,UAAU,EAAE,MAAM,EAClB,UAAU,EAAE,MAAM,GACjB,MAAM,CAOR;AAED,wBAAgB,8BAA8B,CAC5C,mBAAmB,EAAE,MAAM,EAC3B,OAAO,EAAE,0BAA0B,GAClC,MAAM,CAuBR;AAED,wBAAgB,iCAAiC,CAC/C,WAAW,EAAE,MAAM,EACnB,UAAU,EAAE,MAAM,GAAG,SAAS,EAC9B,aAAa,EAAE,MAAM,GACpB,MAAM,CAWR;AAED,wBAAgB,4BAA4B,CAAC,IAAI,EAAE;IACjD,WAAW,EAAE,MAAM,CAAC;IACpB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,aAAa,EAAE,MAAM,CAAC;IACtB,gBAAgB,EAAE,MAAM,CAAC;IACzB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,cAAc,CAAC,EAAE,KAAK,CAAC;QACrB,QAAQ,EAAE,MAAM,CAAC;QACjB,IAAI,EAAE,MAAM,CAAC;QACb,KAAK,CAAC,EAAE,MAAM,CAAC;QACf,OAAO,CAAC,EAAE,MAAM,CAAC;KAClB,CAAC,CAAC;IACH,cAAc,CAAC,EAAE,MAAM,CAAC;CACzB,GAAG,MAAM,CA2CT"}