topchester-ai 0.64.0 → 0.65.0

package/dist/bin.mjs

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { t as runTopchesterCli } from "./cli-CnI4QA2l.mjs";
+import { t as runTopchesterCli } from "./cli-2v1iJ0YF.mjs";
 //#region src/bin.ts
 await runTopchesterCli();
 //#endregion

package/dist/{cli-CnI4QA2l.mjs → cli-2v1iJ0YF.mjs}

@@ -1843,7 +1843,7 @@ async function deleteWorkspaceFile(workspaceRoot, path) {
 const finishTaskTool = defineTool({
 	name: "finish_task",
 	description: "Finish the current task only after the requested work has actually been completed with tools. Do not use this to claim edits, file reads, or validation that did not happen.",
-	prompt: "finish_task: complete the task with a brief final response only after tool results prove the work is done. For implementation tasks, do not call finish_task until source files were changed by edit_file, write_file, apply_patch, or another mutating tool, unless no code change is truly required. Example: {\"tool\":\"finish_task\",\"args\":{\"final_response\":\"Changed src/foo.ts and ran pnpm test foo.test.ts.\",\"files_changed\":[\"src/foo.ts\"],\"validation\":[\"pnpm test foo.test.ts\"],\"remaining_issues\":[]}}",
+	prompt: "finish_task: complete the task with a brief final response only after tool results prove the work is done. In benchmark or require-finish mode, this is the only valid terminal action; normal assistant messages are progress notes and do not finish the task. For implementation tasks, do not call finish_task until source files were changed by edit_file, write_file, apply_patch, or another mutating tool, unless no code change is truly required. Example: {\"tool\":\"finish_task\",\"args\":{\"final_response\":\"Changed src/foo.ts and ran pnpm test foo.test.ts.\",\"files_changed\":[\"src/foo.ts\"],\"validation\":[\"pnpm test foo.test.ts\"],\"remaining_issues\":[]}}",
 	argsSchema: z.object({
 		final_response: z.string().describe("Brief final response for the user."),
 		files_changed: z.array(z.string()).optional().describe("Workspace-relative files actually changed by successful edit tools."),
@@ -12855,6 +12855,22 @@ function formatNoEditCompletionFailure(draftAnswer) {
 	const trimmedDraft = draftAnswer.trim();
 	return ["I could not complete the implementation because no successful source-file edit occurred in this turn.", trimmedDraft ? `The model tried to finish with this unsupported summary:\n${trimmedDraft}` : ""].filter(Boolean).join("\n");
 }
+function formatFinishTaskRequiredRepairInstruction(protocol, draftAnswer) {
+	const toolInstruction = protocol === "text-xml" ? "Reply now with only one XML tool call for the next implementation or validation step, or call finish_task if the work is complete." : protocol === "text-json" ? "Reply now with only one tool JSON object for the next implementation or validation step, or call finish_task if the work is complete." : "Use the available tool calling path now for the next implementation or validation step, or call finish_task if the work is complete.";
+	const trimmedDraft = draftAnswer.trim();
+	return [
+		"This run requires finish_task as the only terminal action.",
+		"A normal assistant message is treated as a progress note, not completion.",
+		"Do not summarize next steps in prose. Use tools to perform the next step.",
+		"Call finish_task only when the requested work is complete and its files_changed, validation, and remaining_issues fields are accurate.",
+		toolInstruction,
+		trimmedDraft ? `Previous progress note that did not finish the task:\n${trimmedDraft}` : ""
+	].filter(Boolean).join("\n");
+}
+function formatFinishTaskRequiredFailure(draftAnswer) {
+	const trimmedDraft = draftAnswer.trim();
+	return ["I stopped because this run requires finish_task, but the model kept replying with normal assistant messages instead of using tools.", trimmedDraft ? `Last progress note:\n${trimmedDraft}` : ""].filter(Boolean).join("\n");
+}
 function getTextToolCallSources(protocol) {
 	return protocol === "text-xml" ? ["text-xml"] : protocol === "text-json" ? ["text-json"] : ["text-json", "text-xml"];
 }
@@ -13153,7 +13169,9 @@ const DEFAULT_MAX_TOOL_CALLS_PER_TURN = 75;
 const MAX_TOOL_CALLS_PER_TURN_ENV = "TOPCHESTER_MAX_TOOL_CALLS_PER_TURN";
 const PLAN_TODO_MODE_ENV = "TOPCHESTER_PLAN_TODO_MODE";
 const MAX_PLAN_TODO_UPDATES_PER_TURN_ENV = "TOPCHESTER_MAX_PLAN_TODO_UPDATES_PER_TURN";
+const REQUIRE_FINISH_TASK_ENV = "TOPCHESTER_REQUIRE_FINISH_TASK";
 const DEFAULT_COMPACT_MAX_PLAN_TODO_UPDATES_PER_TURN = 3;
+const MAX_FINISH_TASK_REQUIRED_REPAIRS = 3;
 const DEFAULT_TASK_CONCURRENCY = 3;
 var TopchesterAgentRuntime = class TopchesterAgentRuntime {
 	context;
@@ -13280,12 +13298,14 @@ var TopchesterAgentRuntime = class TopchesterAgentRuntime {
 		let requestedPlanClosure = false;
 		let invalidToolCallRepairs = 0;
 		let noEditCompletionRepairs = 0;
+		let finishTaskRequiredRepairs = 0;
 		let hasSourceMutation = false;
 		const maxToolCallsPerTurn = readMaxToolCallsPerTurn();
 		const planTodoMode = readPlanTodoMode();
 		const maxPlanTodoUpdatesPerTurn = readMaxPlanTodoUpdatesPerTurn(planTodoMode);
 		let planTodoUpdates = 0;
 		const implementationTask = isImplementationTaskRequest(message);
+		const requireFinishTask = isFinishTaskRequiredByEnv() && isToolAllowed(permissions, "finish_task");
 		const projectInstructionToolState = { shownSourceKeys: /* @__PURE__ */ new Set() };
 		const persistedProjectInstructionKeys = /* @__PURE__ */ new Set();
 		try {
@@ -13396,6 +13416,18 @@ var TopchesterAgentRuntime = class TopchesterAgentRuntime {
 						yield agentEvent.status("ready");
 						return;
 					}
+					if (requireFinishTask) {
+						if (finishTaskRequiredRepairs < MAX_FINISH_TASK_REQUIRED_REPAIRS) {
+							finishTaskRequiredRepairs += 1;
+							nextPrompt = `${nextPrompt}\n\n${formatFinishTaskRequiredRepairInstruction(result.toolProtocol, finalText)}`;
+							continue;
+						}
+						const finalMessage = formatFinishTaskRequiredFailure(finalText);
+						yield agentEvent.assistantMessage(finalMessage, formatAgentMessageMeta(result.modelId, totalDurationMs, tokenUsageTotals));
+						for await (const event of this.streamStopHookEvents(session, finalMessage, "failed", abortSignal)) yield event;
+						yield agentEvent.status("ready");
+						return;
+					}
 					if (planTodoMode === "normal" && hasOpenTaskPlan(plan)) {
 						if (!requestedPlanClosure) {
 							requestedPlanClosure = true;
@@ -14046,6 +14078,10 @@ function readMaxToolCallsPerTurn() {
 function readPlanTodoMode() {
 	return process.env[PLAN_TODO_MODE_ENV]?.trim().toLowerCase() === "compact" ? "compact" : "normal";
 }
+function isFinishTaskRequiredByEnv() {
+	const raw = process.env[REQUIRE_FINISH_TASK_ENV]?.trim().toLowerCase();
+	return raw === "1" || raw === "true" || raw === "yes" || raw === "on";
+}
 function readMaxPlanTodoUpdatesPerTurn(mode) {
 	const raw = process.env[MAX_PLAN_TODO_UPDATES_PER_TURN_ENV]?.trim();
 	if (!raw) return mode === "compact" ? DEFAULT_COMPACT_MAX_PLAN_TODO_UPDATES_PER_TURN : void 0;
@@ -16493,4 +16529,4 @@ function formatDryRunSyncStatus(status) {
 //#endregion
 export { runTopchesterCli as t };
 
-//# sourceMappingURL=cli-CnI4QA2l.mjs.map
+//# sourceMappingURL=cli-2v1iJ0YF.mjs.map