topchester-ai 0.65.0 → 0.66.0

package/dist/bin.mjs

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { t as runTopchesterCli } from "./cli-2v1iJ0YF.mjs";
+import { t as runTopchesterCli } from "./cli-BS67FTk8.mjs";
 //#region src/bin.ts
 await runTopchesterCli();
 //#endregion

package/dist/{cli-2v1iJ0YF.mjs → cli-BS67FTk8.mjs}

@@ -21,6 +21,13 @@ import { diffWords } from "diff";
 import { highlight, supportsLanguage } from "cli-highlight";
 import { Client } from "@modelcontextprotocol/sdk/client/index.js";
 import { StdioClientTransport } from "@modelcontextprotocol/sdk/client/stdio.js";
+//#region src/agent/benchmark-profile.ts
+const BENCHMARK_PROFILES = ["terminal-bench"];
+function parseBenchmarkProfile(value) {
+	if (BENCHMARK_PROFILES.includes(value)) return value;
+	throw new Error(`Unsupported benchmark profile '${value}'. Supported profiles: ${BENCHMARK_PROFILES.join(", ")}.`);
+}
+//#endregion
 //#region src/auth/codex.ts
 const CODEX_CLIENT_ID = "app_EMoamEEZ73f0CkXaXp7hrann";
 const CODEX_ISSUER = "https://auth.openai.com";
@@ -638,9 +645,10 @@ async function validateBashPolicy(args, context) {
 	if (!cwd.allowed) return reject(command, cwd.reason, false);
 	const deniedRule = findMatchingBashRule(command, context.permissions?.deny ?? []);
 	if (deniedRule) return reject(command, `bash policy rejected '${command}' because it matches deny rule '${deniedRule}'.`, false);
+	const shell = resolveBashShell(context.permissions?.shell);
+	if (context.benchmarkProfile === "terminal-bench") return allow(command, cwd.path, realWorkspaceRoot, shell, "benchmark_terminal", "terminal-bench");
 	const destructive = getDestructiveReason(command);
 	if (destructive) return reject(command, `bash policy rejected '${command}' because it looks destructive: ${destructive}.`, false);
-	const shell = resolveBashShell(context.permissions?.shell);
 	const allowedExactRule = findExactBashRule(command, context.permissions?.allowExact ?? []);
 	if (allowedExactRule) return allow(command, cwd.path, realWorkspaceRoot, shell, "allow_exact", allowedExactRule);
 	const approvedRule = findExactBashRule(command, context.approvedCommands ?? []);
@@ -668,7 +676,7 @@ function getBashApprovalCandidates(command) {
 	};
 }
 function allow(command, cwd, workspaceRoot, shell, kind, matchedRule) {
-	const reason = kind === "allow_exact" ? `bash exact command allowed by '${matchedRule}'` : kind === "allow_prefix" ? `bash command allowed by prefix '${matchedRule}'` : `bash exact command approved by '${matchedRule}'`;
+	const reason = kind === "allow_exact" ? `bash exact command allowed by '${matchedRule}'` : kind === "benchmark_terminal" ? `bash command allowed by benchmark profile '${matchedRule}'` : kind === "allow_prefix" ? `bash command allowed by prefix '${matchedRule}'` : `bash exact command approved by '${matchedRule}'`;
 	return {
 		allowed: true,
 		reason,
@@ -758,14 +766,16 @@ const bashTool = defineTool({
 		pathEnv: context.pathEnv,
 		abortSignal: context.abortSignal,
 		permissions: context.config?.tools?.bash,
-		approvedCommands: context.bashApprovals?.allowExactCommands
+		approvedCommands: context.bashApprovals?.allowExactCommands,
+		benchmarkProfile: context.benchmarkProfile
 	})
 });
 async function runBashCommand(workspaceRoot, args, options = {}) {
 	const decision = await validateBashPolicy(args, {
 		workspaceRoot,
 		permissions: options.permissions,
-		approvedCommands: options.approvedCommands
+		approvedCommands: options.approvedCommands,
+		benchmarkProfile: options.benchmarkProfile
 	});
 	if (!decision.allowed) throw new Error(decision.reason);
 	const result = await runProcess({
@@ -1843,7 +1853,7 @@ 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. 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\":[]}}",
+	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, and remaining_issues must be empty. 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."),
@@ -12054,6 +12064,7 @@ async function executeToolCall(workspaceRoot, call, options = {}) {
 		subagents: options.subagents,
 		projectInstructions: options.projectInstructions,
 		currentUserMessage: options.currentUserMessage,
+		benchmarkProfile: options.benchmarkProfile,
 		eventSink: options.eventSink,
 		abortSignal: options.abortSignal,
 		toolCallId: options.toolCallId
@@ -13492,6 +13503,7 @@ var TopchesterAgentRuntime = class TopchesterAgentRuntime {
 							subagents,
 							projectInstructions: projectInstructionToolState,
 							currentUserMessage: message,
+							benchmarkProfile: options.benchmarkProfile,
 							abortSignal,
 							toolCallId: entry.toolCallId,
 							toolCatalog,
@@ -13510,7 +13522,7 @@ var TopchesterAgentRuntime = class TopchesterAgentRuntime {
 						const call = taskCalls[index];
 						const toolResult = taskResults[index];
 						for (const event of createInstructionContextEventsFromToolResult(toolResult, persistedProjectInstructionKeys)) yield event;
-						if (isSourceMutationResult(toolResult)) hasSourceMutation = true;
+						if (isSourceMutationResult(toolResult) || isBenchmarkMutationResult(toolResult, options.benchmarkProfile)) hasSourceMutation = true;
 						yield agentEvent.toolCall(call, formatToolCallMessage(call, toolResult), getToolCallDisplayDiff(toolResult));
 						const postHookRun = this.startPostToolUseHook(call, modelToolCalls[index]?.id, toolResult, session, abortSignal);
 						for await (const event of postHookRun.statusEvents) yield event;
@@ -13562,6 +13574,7 @@ var TopchesterAgentRuntime = class TopchesterAgentRuntime {
 						subagents,
 						projectInstructions: projectInstructionToolState,
 						currentUserMessage: message,
+						benchmarkProfile: options.benchmarkProfile,
 						abortSignal,
 						toolCallId: entry.toolCallId,
 						toolCatalog
@@ -13574,7 +13587,7 @@ var TopchesterAgentRuntime = class TopchesterAgentRuntime {
 						const call = parallelCalls[index];
 						const toolResult = parallelResults[index];
 						for (const event of createInstructionContextEventsFromToolResult(toolResult, persistedProjectInstructionKeys)) yield event;
-						if (isSourceMutationResult(toolResult)) hasSourceMutation = true;
+						if (isSourceMutationResult(toolResult) || isBenchmarkMutationResult(toolResult, options.benchmarkProfile)) hasSourceMutation = true;
 						yield agentEvent.toolCall(call, formatToolCallMessage(call, toolResult), getToolCallDisplayDiff(toolResult));
 						const postHookRun = this.startPostToolUseHook(call, modelToolCalls[index]?.id, toolResult, session, abortSignal);
 						for await (const event of postHookRun.statusEvents) yield event;
@@ -13639,6 +13652,7 @@ var TopchesterAgentRuntime = class TopchesterAgentRuntime {
 								subagents,
 								projectInstructions: projectInstructionToolState,
 								currentUserMessage: message,
+								benchmarkProfile: options.benchmarkProfile,
 								abortSignal,
 								toolCallId: toolCall.id,
 								toolCatalog,
@@ -13657,11 +13671,12 @@ var TopchesterAgentRuntime = class TopchesterAgentRuntime {
 						implementationTask,
 						hasSourceMutation,
 						hasOpenPlan: planTodoMode === "normal" && hasOpenTaskPlan(this.taskPlan.get()),
-						canUseSourceEditTool: canStillUseSourceEditTool(permissions)
+						canUseSourceEditTool: canStillUseSourceEditTool(permissions),
+						requireFinishTask
 					});
 					if (finishError) toolResult = createToolErrorResult(executableToolCall.tool, finishError);
 				}
-				if (isSourceMutationResult(toolResult)) hasSourceMutation = true;
+				if (isSourceMutationResult(toolResult) || isBenchmarkMutationResult(toolResult, options.benchmarkProfile)) hasSourceMutation = true;
 				yield agentEvent.toolCall(executableToolCall, formatToolCallMessage(executableToolCall, toolResult), getToolCallDisplayDiff(toolResult));
 				if (!isToolErrorResult(toolResult) && toolResult.tool === "plan_todo") {
 					planTodoUpdates += 1;
@@ -13870,7 +13885,8 @@ var TopchesterAgentRuntime = class TopchesterAgentRuntime {
 		const decision = await validateBashPolicy(parsed.data, {
 			workspaceRoot: this.context.workspaceRoot,
 			permissions: this.context.config.tools?.bash,
-			approvedCommands
+			approvedCommands,
+			benchmarkProfile: options.benchmarkProfile
 		});
 		if (decision.allowed) return {
 			cancelled: false,
@@ -14125,6 +14141,10 @@ function isSourceMutationResult(result) {
 	if (result.tool === "apply_patch") return result.changedFiles.some(isSourcePath);
 	return false;
 }
+function isBenchmarkMutationResult(result, benchmarkProfile) {
+	if (benchmarkProfile !== "terminal-bench" || isToolErrorResult(result)) return false;
+	return result.tool === "bash" && "workspaceMayHaveChanged" in result && result.workspaceMayHaveChanged === true;
+}
 function isSourcePath(path) {
 	if (!path) return false;
 	return !path.startsWith(".agents/") && !path.startsWith(".git/") && !path.startsWith("topchester-kb/");
@@ -14133,6 +14153,7 @@ function validateFinishTaskResult(result, state) {
 	if (isToolErrorResult(result) || result.tool !== "finish_task") return;
 	if (state.hasOpenPlan) return "finish_task rejected because the visible plan is still open. Close or update plan_todo first.";
 	if (state.implementationTask && state.canUseSourceEditTool && !state.hasSourceMutation) return "finish_task rejected because this appears to be an implementation task but no successful source-file edit has occurred. Use edit_file, write_file, or apply_patch first, unless you can prove no code change is required.";
+	if (state.requireFinishTask && result.remainingIssues.length > 0) return "finish_task rejected because benchmark mode cannot finish with known remaining issues. Continue implementing or validating until remaining_issues is empty.";
 }
 function isL1ContextDisabledByEnv() {
 	const value = process.env.TOPCHESTER_DISABLE_L1_CONTEXT?.trim().toLowerCase();
@@ -15710,12 +15731,14 @@ async function executeRunCommand(context, options) {
 		json: Boolean(options.json),
 		outputJson: options.outputJson,
 		timeoutMs,
-		dangerouslyAutoApprove: Boolean(options.dangerouslyAutoApprove)
+		dangerouslyAutoApprove: Boolean(options.dangerouslyAutoApprove),
+		benchmarkProfile: options.benchmarkProfile
 	}, "run started");
 	pushJson(jsonEvents, runId, session.sessionId, "run.started", {
 		workspaceRoot: runContext.workspaceRoot,
 		timeoutMs,
-		dangerouslyAutoApprove: Boolean(options.dangerouslyAutoApprove)
+		dangerouslyAutoApprove: Boolean(options.dangerouslyAutoApprove),
+		benchmarkProfile: options.benchmarkProfile
 	});
 	try {
 		if (!options.resume) await persistStartupMessages(session, runContext);
@@ -15772,7 +15795,8 @@ async function executeRunCommand(context, options) {
 			});
 			for await (const event of runtime.submitMessageStream(conversation, options.prompt, abortController.signal, {
 				session,
-				userApprovalMode: options.dangerouslyAutoApprove ? "auto_allow" : "interactive"
+				userApprovalMode: options.dangerouslyAutoApprove ? "auto_allow" : "interactive",
+				benchmarkProfile: options.benchmarkProfile
 			})) await applyRuntimeEvent({
 				event,
 				session,
@@ -16228,7 +16252,7 @@ function createTopchesterProgram() {
 			process.exitCode = 1;
 		}
 	});
-	program.command("run").description("run one prompt or slash command without opening the TUI").argument("<prompt...>", "prompt text or slash command").option("--model <model>", "override the agent.primary model for this run").option("--timeout <ms>", "timeout for the run in milliseconds", parsePositiveInteger).option("--json", "write JSONL run events to stdout").option("--output-json <path>", "write JSONL run events to a file").option("--dangerously-auto-approve", "auto-approve prompt-gated tool calls for this non-interactive run").action(async (promptParts, options) => {
+	program.command("run").description("run one prompt or slash command without opening the TUI").argument("<prompt...>", "prompt text or slash command").option("--model <model>", "override the agent.primary model for this run").option("--timeout <ms>", "timeout for the run in milliseconds", parsePositiveInteger).option("--json", "write JSONL run events to stdout").option("--output-json <path>", "write JSONL run events to a file").option("--dangerously-auto-approve", "auto-approve prompt-gated tool calls for this non-interactive run").option("--benchmark-profile <profile>", "enable an explicit benchmark runtime profile", parseBenchmarkProfile).action(async (promptParts, options) => {
 		const context = createContextFromOptions(program);
 		const globalOptions = program.opts();
 		try {
@@ -16239,7 +16263,8 @@ function createTopchesterProgram() {
 				json: options.json,
 				outputJson: options.outputJson,
 				resume: globalOptions.resume,
-				dangerouslyAutoApprove: options.dangerouslyAutoApprove
+				dangerouslyAutoApprove: options.dangerouslyAutoApprove,
+				benchmarkProfile: options.benchmarkProfile
 			});
 		} catch (error) {
 			console.error(formatStartupError(error));
@@ -16529,4 +16554,4 @@ function formatDryRunSyncStatus(status) {
 //#endregion
 export { runTopchesterCli as t };
 
-//# sourceMappingURL=cli-2v1iJ0YF.mjs.map
+//# sourceMappingURL=cli-BS67FTk8.mjs.map