@visorcraft/idlehands 4.2.2 → 4.2.3

package/dist/{anton-DSS3RNo7.js → anton-CmB9XtGt.js}

@@ -94,71 +94,84 @@ function buildDirectTaskPrompt(task) {
 	].join("\n");
 }
 function buildDiscoveryPrompt(task, taskFile, planFilePath) {
-	return `You are running PRE-FLIGHT for an autonomous coding orchestrator.
+	return `You are a coding assistant running discovery for a task orchestrator.
 
-## RULES
-- DO NOT implement code changes. DO NOT modify source files.
-- You may ONLY write to the plan file specified below.
-- Read the codebase to understand what needs to change.
+## YOUR JOB
+Analyze the codebase and write a detailed implementation plan to a file.
+You will NOT implement anything yourself — a separate agent will do that using your plan.
 
 ## TASK
 From task file: ${taskFile}
 Task: ${task}
 
-## DECISION FLOW
+## WHAT YOU MUST DO
 
-**If task is ALREADY COMPLETE** (you are CERTAIN it's done after checking with tools):
-→ Return ONLY: {"status":"complete","filename":""}
+1. Use the read tool to examine relevant source files, configs, tests, and docs.
+2. Understand what currently exists and what needs to change.
+3. Use the write tool to create the plan file at: ${planFilePath}
 
-**If task is INCOMPLETE or UNCERTAIN**:
-1. Read relevant source files to understand the current state
-2. Write a detailed implementation plan to: ${planFilePath}
-3. The plan MUST include:
-   - Task description (verbatim from above)
-   - What needs to change and why
-   - Implementation approach with specific steps
-   - Files to modify/create
-   - How to verify (test commands)
-4. Verify the file exists and is non-empty
-5. Return ONLY: {"status":"incomplete","filename":"${planFilePath}"}
+The plan file MUST contain:
+- The task description
+- What currently exists (files you read, current behavior)
+- What specifically needs to change and why
+- Step-by-step implementation instructions (specific enough that another agent can follow them)
+- Which files to modify or create, with the exact changes needed
+- How to verify the changes work (test commands, expected output)
 
-## OUTPUT
-Return ONLY the JSON object. No markdown fences. No explanation.`;
-}
-function buildReviewPrompt(planFilePath) {
-	return `Review this implementation plan and improve it:
-${planFilePath}
+4. After writing the plan file, return ONLY this JSON:
+{"status":"incomplete","filename":"${planFilePath}"}
 
-Treat it as written by a junior developer. Look for:
-- Missing edge cases
-- Opportunities to reuse existing code
-- Unclear or ambiguous steps
-- Missing test scenarios
+If you are CERTAIN the task is already complete (you verified with tools), return:
+{"status":"complete","filename":""}
 
-Update the SAME file in-place with your improvements.
+## IMPORTANT
+- You MUST use tools (read, exec, write). Do not skip straight to returning JSON.
+- Do not modify any source files other than the plan file above.
+- The quality of your plan directly determines whether the task succeeds.`;
+}
+function buildReviewPrompt(planFilePath) {
+	return `You are reviewing an implementation plan written by another agent.
 
-After review, return ONLY:
-{"status":"ready","filename":"${planFilePath}"}
+1. Use the read tool to open: ${planFilePath}
+2. Read the plan carefully. Check for:
+   - Missing edge cases or unclear steps
+   - Opportunities to reuse existing code
+   - Missing test scenarios or verification steps
+   - Whether the file paths and change descriptions are specific enough for another agent to follow
+3. Use the write tool to update the SAME file in-place with your improvements.
+4. After updating, return: {"status":"ready","filename":"${planFilePath}"}
 
-Do not return status=ready unless the file exists and has content.
-Return JSON only. No markdown fences. No commentary.`;
+Do not skip reading the file. Do not return JSON without first reading and updating the plan.`;
 }
 function buildImplementationPrompt(task, planFilePath) {
 	return [
-		"You are implementing a task from a managed checklist.",
-		"A planning agent has already analyzed the codebase and written a detailed spec.",
+		`Complete this task: ${task}`,
 		"",
-		`Task: ${task}`,
+		`An implementation plan is available at: ${planFilePath}`,
+		"Read it first, then follow its instructions.",
 		"",
-		`Implementation plan: ${planFilePath}`,
+		"You MUST use the edit or write tool to modify the actual source files.",
+		"Describing changes in text or showing diffs without applying them does NOT count.",
+		"Do not ask for confirmation — apply the changes now.",
 		"",
-		"Rules:",
-		"1) Read the plan file FIRST — it contains the implementation approach.",
-		"2) Follow the plan precisely. Make the code changes specified.",
-		"3) Run the verification steps from the plan.",
-		"4) Return a concise completion summary.",
+		"After making the edits, run any verification steps from the plan.",
+		"Then return a brief summary of what you changed."
+	].join("\n");
+}
+function buildImplementationRetryPrompt(task, planFilePath) {
+	return [
+		"Your previous implementation attempt made zero repository changes.",
+		`Task: ${task}`,
+		`Plan file: ${planFilePath}`,
+		"",
+		"You must now perform actual file edits in this turn.",
+		"Required sequence:",
+		"1) Read the plan file.",
+		"2) Use edit/write to modify repository files.",
+		"3) Optionally run verification.",
+		"4) Return a short summary.",
 		"",
-		"Do not deviate from the plan unless you find a clear error in it."
+		"If you only provide analysis/diff text without edit/write tool calls, this attempt fails."
 	].join("\n");
 }
 function formatDuration(ms) {
@@ -225,7 +238,7 @@ async function loadAntonConfig() {
 	}
 }
 async function runAgentTask(args) {
-	const { agentCliCommand } = await import("./agent-via-gateway-CldlRczJ.js").then((n) => n.n);
+	const { agentCliCommand } = await import("./agent-via-gateway-BN0HMYd7.js").then((n) => n.n);
 	const extraSystemPrompt = args.workspaceDir ? `Your working directory is: ${args.workspaceDir}\nAll file paths are relative to this directory. Use this as your cwd for all operations.` : void 0;
 	const result = await agentCliCommand({
 		message: args.message,
@@ -326,18 +339,18 @@ function normalizeDiscoveryFilename(filename, expectedPlanFile) {
 	return path.resolve(trimmed);
 }
 function buildDiscoveryRepairPrompt(task, taskFile, planFilePath) {
-	return `RETRY: your previous preflight response did not produce a valid plan file.
+	return `Your previous attempt did not produce a valid plan file. Try again.
 
-You MUST complete these steps exactly:
-1) Write the implementation plan to: ${planFilePath}
-2) Verify the file exists and is non-empty
-3) Return ONLY strict JSON: {"status":"incomplete","filename":"${planFilePath}"}
+Task file: ${taskFile}
+Task: ${task}
 
-If task is already complete, return ONLY: {"status":"complete","filename":""}
+You MUST:
+1. Use the read tool to examine the relevant source files in this repository.
+2. Use the write tool to create a detailed implementation plan at: ${planFilePath}
+   The plan must include: what exists now, what needs to change, specific file edits, and verification steps.
+3. After writing the file, return: {"status":"incomplete","filename":"${planFilePath}"}
 
-Do not include THOUGHT, markdown fences, or extra commentary.
-Task file: ${taskFile}
-Task: ${task}`;
+Do not skip tool usage. Do not return JSON without first writing the plan file.`;
 }
 async function tryPersistPlanFallback(params) {
 	const planText = (((typeof params.parsed?.planMarkdown === "string" ? params.parsed.planMarkdown : void 0) ?? (typeof params.parsed?.plan === "string" ? params.parsed.plan : void 0))?.trim() || extractPlanMarkdownFromText(params.rawText) || "").trim();
@@ -738,8 +751,29 @@ async function runAnton(args) {
 					deps: args.deps,
 					workspaceDir: args.workspaceDir
 				});
-				const changedAfter = await getGitChangedFileCount(gitCwd, changeIgnores);
-				if (mode === "preflight" && changedBeforeTask !== null && changedAfter !== null && changedAfter <= changedBeforeTask) throw new Error("Implementation made no repository changes; refusing to mark task complete");
+				let changedAfter = await getGitChangedFileCount(gitCwd, changeIgnores);
+				if (mode === "preflight" && changedBeforeTask !== null && changedAfter !== null && changedAfter <= changedBeforeTask) {
+					const retrySessionId = `anton-impl-retry-${Date.now()}-${taskNum}`;
+					await notify({
+						phase: "task_agent_spawned",
+						index: taskNum,
+						total: pending.length,
+						task: `${task.text} (implementation retry)`,
+						sessionId: retrySessionId
+					});
+					await runAgentTask({
+						message: buildImplementationRetryPrompt(task.text, planFile ?? ""),
+						sessionId: retrySessionId,
+						timeout: defaultTimeout,
+						agent: args.agent,
+						to: args.to,
+						runtime: args.runtime,
+						deps: args.deps,
+						workspaceDir: args.workspaceDir
+					});
+					changedAfter = await getGitChangedFileCount(gitCwd, changeIgnores);
+					if (changedAfter !== null && changedAfter <= changedBeforeTask) throw new Error("Implementation made no repository changes after retry; refusing to mark task complete");
+				}
 				const updated = markTaskDone(await fs.readFile(filePath, "utf8"), task.line);
 				await fs.writeFile(filePath, updated, "utf8");
 				completed += 1;

package/dist/{anton-CuLeFevj.js → anton-Dn_Pc4P0.js}

@@ -1,5 +1,5 @@
-import "./paths-BYNVLNi_.js";
-import { t as CONFIG_DIR } from "./utils-BB3tnxio.js";
+import "./paths-DQpv9a3Q.js";
+import { t as CONFIG_DIR } from "./utils-D9wGoiU9.js";
 import fs from "node:fs/promises";
 import path from "node:path";
 import { promisify } from "node:util";
@@ -88,71 +88,84 @@ function buildDirectTaskPrompt(task) {
 	].join("\n");
 }
 function buildDiscoveryPrompt(task, taskFile, planFilePath) {
-	return `You are running PRE-FLIGHT for an autonomous coding orchestrator.
+	return `You are a coding assistant running discovery for a task orchestrator.
 
-## RULES
-- DO NOT implement code changes. DO NOT modify source files.
-- You may ONLY write to the plan file specified below.
-- Read the codebase to understand what needs to change.
+## YOUR JOB
+Analyze the codebase and write a detailed implementation plan to a file.
+You will NOT implement anything yourself — a separate agent will do that using your plan.
 
 ## TASK
 From task file: ${taskFile}
 Task: ${task}
 
-## DECISION FLOW
+## WHAT YOU MUST DO
 
-**If task is ALREADY COMPLETE** (you are CERTAIN it's done after checking with tools):
-→ Return ONLY: {"status":"complete","filename":""}
+1. Use the read tool to examine relevant source files, configs, tests, and docs.
+2. Understand what currently exists and what needs to change.
+3. Use the write tool to create the plan file at: ${planFilePath}
 
-**If task is INCOMPLETE or UNCERTAIN**:
-1. Read relevant source files to understand the current state
-2. Write a detailed implementation plan to: ${planFilePath}
-3. The plan MUST include:
-   - Task description (verbatim from above)
-   - What needs to change and why
-   - Implementation approach with specific steps
-   - Files to modify/create
-   - How to verify (test commands)
-4. Verify the file exists and is non-empty
-5. Return ONLY: {"status":"incomplete","filename":"${planFilePath}"}
+The plan file MUST contain:
+- The task description
+- What currently exists (files you read, current behavior)
+- What specifically needs to change and why
+- Step-by-step implementation instructions (specific enough that another agent can follow them)
+- Which files to modify or create, with the exact changes needed
+- How to verify the changes work (test commands, expected output)
 
-## OUTPUT
-Return ONLY the JSON object. No markdown fences. No explanation.`;
-}
-function buildReviewPrompt(planFilePath) {
-	return `Review this implementation plan and improve it:
-${planFilePath}
+4. After writing the plan file, return ONLY this JSON:
+{"status":"incomplete","filename":"${planFilePath}"}
 
-Treat it as written by a junior developer. Look for:
-- Missing edge cases
-- Opportunities to reuse existing code
-- Unclear or ambiguous steps
-- Missing test scenarios
+If you are CERTAIN the task is already complete (you verified with tools), return:
+{"status":"complete","filename":""}
 
-Update the SAME file in-place with your improvements.
+## IMPORTANT
+- You MUST use tools (read, exec, write). Do not skip straight to returning JSON.
+- Do not modify any source files other than the plan file above.
+- The quality of your plan directly determines whether the task succeeds.`;
+}
+function buildReviewPrompt(planFilePath) {
+	return `You are reviewing an implementation plan written by another agent.
 
-After review, return ONLY:
-{"status":"ready","filename":"${planFilePath}"}
+1. Use the read tool to open: ${planFilePath}
+2. Read the plan carefully. Check for:
+   - Missing edge cases or unclear steps
+   - Opportunities to reuse existing code
+   - Missing test scenarios or verification steps
+   - Whether the file paths and change descriptions are specific enough for another agent to follow
+3. Use the write tool to update the SAME file in-place with your improvements.
+4. After updating, return: {"status":"ready","filename":"${planFilePath}"}
 
-Do not return status=ready unless the file exists and has content.
-Return JSON only. No markdown fences. No commentary.`;
+Do not skip reading the file. Do not return JSON without first reading and updating the plan.`;
 }
 function buildImplementationPrompt(task, planFilePath) {
 	return [
-		"You are implementing a task from a managed checklist.",
-		"A planning agent has already analyzed the codebase and written a detailed spec.",
+		`Complete this task: ${task}`,
 		"",
-		`Task: ${task}`,
+		`An implementation plan is available at: ${planFilePath}`,
+		"Read it first, then follow its instructions.",
 		"",
-		`Implementation plan: ${planFilePath}`,
+		"You MUST use the edit or write tool to modify the actual source files.",
+		"Describing changes in text or showing diffs without applying them does NOT count.",
+		"Do not ask for confirmation — apply the changes now.",
 		"",
-		"Rules:",
-		"1) Read the plan file FIRST — it contains the implementation approach.",
-		"2) Follow the plan precisely. Make the code changes specified.",
-		"3) Run the verification steps from the plan.",
-		"4) Return a concise completion summary.",
+		"After making the edits, run any verification steps from the plan.",
+		"Then return a brief summary of what you changed."
+	].join("\n");
+}
+function buildImplementationRetryPrompt(task, planFilePath) {
+	return [
+		"Your previous implementation attempt made zero repository changes.",
+		`Task: ${task}`,
+		`Plan file: ${planFilePath}`,
+		"",
+		"You must now perform actual file edits in this turn.",
+		"Required sequence:",
+		"1) Read the plan file.",
+		"2) Use edit/write to modify repository files.",
+		"3) Optionally run verification.",
+		"4) Return a short summary.",
 		"",
-		"Do not deviate from the plan unless you find a clear error in it."
+		"If you only provide analysis/diff text without edit/write tool calls, this attempt fails."
 	].join("\n");
 }
 function formatDuration(ms) {
@@ -212,14 +225,14 @@ async function antonStop(runtime) {
 /** Load Anton config from the IdleHands config file. */
 async function loadAntonConfig() {
 	try {
-		const { loadConfig } = await import("./model-selection-C88a_9g6.js").then((n) => n.I);
+		const { loadConfig } = await import("./model-selection-C5hI1dum.js").then((n) => n.I);
 		return loadConfig().anton ?? {};
 	} catch {
 		return {};
 	}
 }
 async function runAgentTask(args) {
-	const { agentCliCommand } = await import("./agent-via-gateway-Bf6TEh2k.js");
+	const { agentCliCommand } = await import("./agent-via-gateway-BsTP4WyI.js");
 	const extraSystemPrompt = args.workspaceDir ? `Your working directory is: ${args.workspaceDir}\nAll file paths are relative to this directory. Use this as your cwd for all operations.` : void 0;
 	const result = await agentCliCommand({
 		message: args.message,
@@ -320,18 +333,18 @@ function normalizeDiscoveryFilename(filename, expectedPlanFile) {
 	return path.resolve(trimmed);
 }
 function buildDiscoveryRepairPrompt(task, taskFile, planFilePath) {
-	return `RETRY: your previous preflight response did not produce a valid plan file.
+	return `Your previous attempt did not produce a valid plan file. Try again.
 
-You MUST complete these steps exactly:
-1) Write the implementation plan to: ${planFilePath}
-2) Verify the file exists and is non-empty
-3) Return ONLY strict JSON: {"status":"incomplete","filename":"${planFilePath}"}
+Task file: ${taskFile}
+Task: ${task}
 
-If task is already complete, return ONLY: {"status":"complete","filename":""}
+You MUST:
+1. Use the read tool to examine the relevant source files in this repository.
+2. Use the write tool to create a detailed implementation plan at: ${planFilePath}
+   The plan must include: what exists now, what needs to change, specific file edits, and verification steps.
+3. After writing the file, return: {"status":"incomplete","filename":"${planFilePath}"}
 
-Do not include THOUGHT, markdown fences, or extra commentary.
-Task file: ${taskFile}
-Task: ${task}`;
+Do not skip tool usage. Do not return JSON without first writing the plan file.`;
 }
 async function tryPersistPlanFallback(params) {
 	const planText = (((typeof params.parsed?.planMarkdown === "string" ? params.parsed.planMarkdown : void 0) ?? (typeof params.parsed?.plan === "string" ? params.parsed.plan : void 0))?.trim() || extractPlanMarkdownFromText(params.rawText) || "").trim();
@@ -592,7 +605,7 @@ async function runAnton(args) {
 	const reviewTimeout = antonCfg.reviewTimeoutSec ?? taskTimeout;
 	const preflightMaxRetries = antonCfg.preflightMaxRetries ?? 2;
 	const planDir = antonCfg.planDir ? path.resolve(antonCfg.planDir) : path.resolve(path.dirname(filePath), ".agents", "tasks");
-	const { loadConfig } = await import("./model-selection-C88a_9g6.js").then((n) => n.I);
+	const { loadConfig } = await import("./model-selection-C5hI1dum.js").then((n) => n.I);
 	const cfg = loadConfig();
 	const defaultTimeout = String(Number.isFinite(args.timeoutSec) && (args.timeoutSec ?? 0) > 0 ? args.timeoutSec : cfg.agents?.defaults?.timeoutSeconds ?? taskTimeout);
 	await acquireLock(Boolean(args.force));
@@ -732,8 +745,29 @@ async function runAnton(args) {
 					deps: args.deps,
 					workspaceDir: args.workspaceDir
 				});
-				const changedAfter = await getGitChangedFileCount(gitCwd, changeIgnores);
-				if (mode === "preflight" && changedBeforeTask !== null && changedAfter !== null && changedAfter <= changedBeforeTask) throw new Error("Implementation made no repository changes; refusing to mark task complete");
+				let changedAfter = await getGitChangedFileCount(gitCwd, changeIgnores);
+				if (mode === "preflight" && changedBeforeTask !== null && changedAfter !== null && changedAfter <= changedBeforeTask) {
+					const retrySessionId = `anton-impl-retry-${Date.now()}-${taskNum}`;
+					await notify({
+						phase: "task_agent_spawned",
+						index: taskNum,
+						total: pending.length,
+						task: `${task.text} (implementation retry)`,
+						sessionId: retrySessionId
+					});
+					await runAgentTask({
+						message: buildImplementationRetryPrompt(task.text, planFile ?? ""),
+						sessionId: retrySessionId,
+						timeout: defaultTimeout,
+						agent: args.agent,
+						to: args.to,
+						runtime: args.runtime,
+						deps: args.deps,
+						workspaceDir: args.workspaceDir
+					});
+					changedAfter = await getGitChangedFileCount(gitCwd, changeIgnores);
+					if (changedAfter !== null && changedAfter <= changedBeforeTask) throw new Error("Implementation made no repository changes after retry; refusing to mark task complete");
+				}
 				const updated = markTaskDone(await fs.readFile(filePath, "utf8"), task.line);
 				await fs.writeFile(filePath, updated, "utf8");
 				completed += 1;