tinker-agent 1.6.0 → 1.8.0

package/src/events/stdout-event-printer.ts

@@ -202,6 +202,7 @@ function formatToolRawResult(call: ToolCall, raw: ToolRawResult): string[] {
       return optionalLine(formatDiff(call, raw));
     case "bash":
     case "task_output":
+    case "task_input":
       return optionalLine(formatBashResult(call, raw));
     case "task_list":
     case "task_stop":
@@ -278,7 +279,7 @@ function formatDiff(call: ToolCall, raw: ToolRawResult): string | undefined {
 }
 
 function formatBashResult(_call: ToolCall, raw: ToolRawResult): string | undefined {
-  if (raw.kind !== "bash" && raw.kind !== "task_output") {
+  if (raw.kind !== "bash" && raw.kind !== "task_output" && raw.kind !== "task_input") {
     return undefined;
   }
 
@@ -348,7 +349,11 @@ function formatToolLine(prefix: string, call: ToolCall): string {
     return `${prefix} name=${call.name}\n`;
   }
 
-  if (call.name === "TaskOutput" || call.name === "TaskStop") {
+  if (
+    call.name === "TaskOutput" ||
+    call.name === "TaskInput" ||
+    call.name === "TaskStop"
+  ) {
     const taskId = toolTaskId(call);
     return `${prefix} name=${call.name}${taskId === undefined ? "" : ` task=${taskId}`}\n`;
   }

package/src/mcp/mcp-manager.ts

@@ -1,15 +1,24 @@
 import { Client } from "@modelcontextprotocol/sdk/client/index.js";
+import { tmpdir } from "node:os";
+import path from "node:path";
+import process from "node:process";
+import { pathToFileURL } from "node:url";
 import {
   StdioClientTransport,
   getDefaultEnvironment,
 } from "@modelcontextprotocol/sdk/client/stdio.js";
 import type { RuntimeSessionContext } from "../agent/runtime-session";
+import { ListRootsRequestSchema } from "@modelcontextprotocol/sdk/types.js";
 import type { ToolExecutor } from "../tools/types";
 import type { McpConfig, McpServerConfig } from "./mcp-config";
 import { createMcpToolExecutor } from "./mcp-tool-executor";
 
 const STDERR_TAIL_MAX_CHARS = 2_000;
 
+function temporaryDirectoryEnvironment(): Record<string, string> {
+  return process.platform === "win32" ? { TEMP: tmpdir() } : { TMPDIR: tmpdir() };
+}
+
 export type McpClientConnection = {
   client: Client;
   close(): Promise<void>;
@@ -18,6 +27,7 @@ export type McpClientConnection = {
 export type McpClientFactory = (
   serverName: string,
   serverConfig: McpServerConfig,
+  workspaceRoot: string,
 ) => Promise<McpClientConnection>;
 
 export type McpManager = {
@@ -37,6 +47,7 @@ export type McpServerInventory = {
 
 export type CreateMcpManagerOptions = {
   config: McpConfig;
+  workspaceRoot: string;
   runtimeSession: RuntimeSessionContext;
   clientFactory?: McpClientFactory;
   timeoutMs?: number;
@@ -70,7 +81,11 @@ export async function createMcpManager(
       let tools;
 
       try {
-        connection = await clientFactory(serverName, serverConfig);
+        connection = await clientFactory(
+          serverName,
+          serverConfig,
+          options.workspaceRoot,
+        );
       } catch (error) {
         await options.runtimeSession.append({
           type: "mcp.server.failed",
@@ -227,11 +242,16 @@ async function closeConnections(
 async function stdioClientFactory(
   serverName: string,
   serverConfig: McpServerConfig,
+  workspaceRoot: string,
 ): Promise<McpClientConnection> {
   const transport = new StdioClientTransport({
     command: serverConfig.command,
     args: serverConfig.args,
-    env: { ...getDefaultEnvironment(), ...serverConfig.env },
+    env: {
+      ...getDefaultEnvironment(),
+      ...temporaryDirectoryEnvironment(),
+      ...serverConfig.env,
+    },
     cwd: serverConfig.cwd,
     stderr: "pipe",
   });
@@ -241,7 +261,18 @@ async function stdioClientFactory(
     stderrTail = (stderrTail + chunk.toString("utf8")).slice(-STDERR_TAIL_MAX_CHARS);
   });
 
-  const client = new Client({ name: "tinker", version: "0.1.0" });
+  const client = new Client(
+    { name: "tinker", version: "0.1.0" },
+    { capabilities: { roots: { listChanged: false } } },
+  );
+  client.setRequestHandler(ListRootsRequestSchema, () => ({
+    roots: [
+      {
+        uri: pathToFileURL(workspaceRoot).href,
+        name: path.basename(workspaceRoot),
+      },
+    ],
+  }));
 
   try {
     await client.connect(transport);

package/src/model/fake-model-client.ts

@@ -247,6 +247,9 @@ export class FakeModelClient implements ModelClient {
     if (this.mode === "pty-background-task") {
       return this.ptyBackgroundTask(input, prepared, options);
     }
+    if (this.mode === "pty-interactive-terminal") {
+      return this.ptyInteractiveTerminal(input, prepared, options);
+    }
     if (this.mode === "pty-resume") {
       return this.ptyResume(input, prepared, options);
     }
@@ -658,6 +661,87 @@ export class FakeModelClient implements ModelClient {
     return textOutput(prepared, "PTY_BACKGROUND_STOPPED");
   }
 
+  private ptyInteractiveTerminal(
+    input: ModelRequestInput,
+    prepared: PreparedModelRequest,
+    options: ModelRequestOptions,
+  ): ModelRequestOutput {
+    requireTools(input, ["Bash", "TaskOutput", "TaskInput"]);
+    const prompt = lastUserMessage(input.messages);
+    if (prompt === "PTY_INTERACTIVE_FOLLOWUP") {
+      requireMessage(input.messages, "assistant", "PTY_INTERACTIVE_DONE");
+      return textOutput(prepared, "PTY_INTERACTIVE_FOLLOWUP_DONE");
+    }
+    if (prompt !== "PTY_INTERACTIVE_TERMINAL" && prompt !== "PTY_INTERACTIVE_QUIT") {
+      throw new Error(
+        `Unexpected pty-interactive-terminal prompt: ${JSON.stringify(prompt)}.`,
+      );
+    }
+
+    const tools = toolMessagesAfterLastUser(input.messages);
+    const bash = tools.find((message) => message.name === "Bash");
+    if (bash === undefined) {
+      return toolCallOutput(prepared, options, "Bash", {
+        command: "python3 -q",
+        description: "Start interactive Python fixture",
+        tty: true,
+        timeout: 25,
+      });
+    }
+    if (!bash.content.includes("taskId=") || !bash.content.includes("tty=true")) {
+      throw new Error("PTY Bash task did not return an interactive task ID.");
+    }
+    const taskId = requireObservationValue(bash.content, "taskId");
+
+    const outputs = tools.filter((message) => message.name === "TaskOutput");
+    const output = outputs.at(-1);
+    if (output === undefined || !output.content.includes(">>>")) {
+      if (outputs.length >= 20) {
+        throw new Error("Interactive Python fixture did not show its prompt.");
+      }
+      return toolCallOutput(prepared, options, "TaskOutput", {
+        task_id: taskId,
+      });
+    }
+
+    const inputs = tools.filter((message) => message.name === "TaskInput");
+    if (inputs.length === 0) {
+      return toolCallOutput(prepared, options, "TaskInput", {
+        task_id: taskId,
+        chars:
+          prompt === "PTY_INTERACTIVE_QUIT"
+            ? "import os; print('PTY_INTERACTIVE_PID=' + str(os.getpid()))\n"
+            : "print(6 * 7)\n",
+        wait_ms: 250,
+      });
+    }
+
+    const latestInput = inputs.at(-1);
+    const expected = prompt === "PTY_INTERACTIVE_QUIT" ? "PTY_INTERACTIVE_PID=" : "42";
+    if (!latestInput?.content.includes(expected)) {
+      if (inputs.length >= 20) {
+        throw new Error(`Interactive Python fixture did not show ${expected}.`);
+      }
+      return toolCallOutput(prepared, options, "TaskInput", {
+        task_id: taskId,
+        chars: "",
+        wait_ms: 250,
+      });
+    }
+
+    if (prompt === "PTY_INTERACTIVE_QUIT") {
+      return textOutput(prepared, "PTY_INTERACTIVE_RUNNING");
+    }
+    if (!inputs.some((message) => message.content.includes("status=completed"))) {
+      return toolCallOutput(prepared, options, "TaskInput", {
+        task_id: taskId,
+        chars: "exit()\n",
+        wait_ms: 500,
+      });
+    }
+    return textOutput(prepared, "PTY_INTERACTIVE_DONE");
+  }
+
   private ptyResume(
     input: ModelRequestInput,
     prepared: PreparedModelRequest,

package/src/observation/observation-builder.ts

@@ -11,6 +11,7 @@ import type {
   ReadFileRawResult,
   RecallRawResult,
   SkillRawResult,
+  TaskInputRawResult,
   TaskListRawResult,
   TaskOutputRawResult,
   TaskStopRawResult,
@@ -51,6 +52,8 @@ export class ObservationBuilder {
         return { content: renderTaskListObservation(input.raw) };
       case "task_output":
         return { content: renderTaskOutputObservation(input.raw) };
+      case "task_input":
+        return { content: renderTaskInputObservation(input.raw) };
       case "task_stop":
         return { content: renderTaskStopObservation(input.raw) };
       case "web_search":
@@ -331,8 +334,11 @@ function renderBashObservation(raw: BashRawResult): string {
       `timeoutMs=${raw.timeoutMs ?? 0}`,
       `command=${raw.command}`,
       `cwd=${raw.cwd}`,
+      `tty=${raw.tty}`,
       `outputFilePath=${raw.outputFilePath}`,
-      "Use Read on outputFilePath to inspect current output.",
+      raw.tty
+        ? "Use TaskOutput to inspect the current terminal screen and TaskInput to interact."
+        : "Use TaskOutput to inspect current output.",
     ].join("\n");
   }
 
@@ -342,8 +348,11 @@ function renderBashObservation(raw: BashRawResult): string {
       `taskId=${raw.taskId}`,
       `command=${raw.command}`,
       `cwd=${raw.cwd}`,
+      `tty=${raw.tty}`,
       `outputFilePath=${raw.outputFilePath}`,
-      "Use Read on outputFilePath to inspect current output.",
+      raw.tty
+        ? "Use TaskOutput to inspect the current terminal screen and TaskInput to interact."
+        : "Use TaskOutput to inspect current output.",
     ].join("\n");
   }
 
@@ -353,6 +362,7 @@ function renderBashObservation(raw: BashRawResult): string {
     `exitCode=${raw.exitCode ?? "null"}`,
     `status=${raw.status}`,
     `cwd=${raw.cwd}`,
+    `tty=${raw.tty}`,
     `outputFilePath=${raw.outputFilePath}`,
     `outputBytes=${raw.outputBytes}`,
     `outputLines=${raw.outputLines}`,
@@ -362,8 +372,9 @@ function renderBashObservation(raw: BashRawResult): string {
       ? undefined
       : `returnCodeInterpretation=${raw.returnCodeInterpretation}`,
     raw.error === undefined ? undefined : `error=${raw.error}`,
-    "preview:",
-    raw.preview,
+    raw.tty ? `screen=${raw.screenColumns ?? 80}x${raw.screenRows ?? 24}` : undefined,
+    raw.tty ? "current screen:" : "preview:",
+    raw.tty ? (raw.screen ?? "") : raw.preview,
   ]
     .filter((line): line is string => line !== undefined)
     .join("\n");
@@ -387,23 +398,48 @@ function renderTaskOutputObservation(raw: TaskOutputRawResult): string {
     return `TaskOutput failed for ${raw.taskId || "(unknown task ID)"}: ${raw.error ?? "Unknown error."}`;
   }
 
+  const terminalScreen = raw.task.tty;
   return [
     "Task output retrieved.",
     `taskId=${raw.taskId}`,
     `status=${raw.task.status}`,
     `command=${raw.task.command}`,
+    `tty=${terminalScreen}`,
     `outputFilePath=${raw.outputFilePath}`,
     `outputBytes=${raw.outputBytes ?? 0}`,
     `outputLines=${raw.outputLines ?? 0}`,
     `truncated=${raw.truncated ?? false}`,
     raw.omittedLines === undefined ? undefined : `omittedLines=${raw.omittedLines}`,
-    "preview:",
-    raw.preview ?? "",
+    terminalScreen
+      ? `screen=${raw.screenColumns ?? 80}x${raw.screenRows ?? 24}`
+      : undefined,
+    terminalScreen ? "current screen:" : "preview:",
+    terminalScreen ? (raw.screen ?? "") : (raw.preview ?? ""),
   ]
     .filter((line): line is string => line !== undefined)
     .join("\n");
 }
 
+function renderTaskInputObservation(raw: TaskInputRawResult): string {
+  if (!raw.ok) {
+    return `TaskInput failed for ${raw.taskId || "(unknown task ID)"}: ${raw.error}`;
+  }
+
+  return [
+    "Terminal input sent.",
+    `taskId=${raw.taskId}`,
+    `status=${raw.status}`,
+    `writtenBytes=${raw.writtenBytes}`,
+    `waitedMs=${raw.waitedMs}`,
+    `screen=${raw.screenColumns}x${raw.screenRows}`,
+    `outputFilePath=${raw.outputFilePath}`,
+    `outputBytes=${raw.outputBytes}`,
+    `outputLines=${raw.outputLines}`,
+    "current screen:",
+    raw.screen,
+  ].join("\n");
+}
+
 function renderTaskStopObservation(raw: TaskStopRawResult): string {
   if (!raw.ok || raw.task === undefined) {
     return `TaskStop failed for ${raw.taskId || "(unknown task ID)"}: ${raw.error ?? "Unknown error."}`;
@@ -428,6 +464,7 @@ function renderTaskSummary(task: TaskListRawResult["tasks"][number]): string {
     `taskId=${task.taskId}`,
     `description=${task.description}`,
     `status=${task.status}`,
+    `tty=${task.tty}`,
     `startedAt=${task.startedAt}`,
     task.endedAt === undefined ? undefined : `endedAt=${task.endedAt}`,
     task.exitCode === undefined ? undefined : `exitCode=${task.exitCode}`,

package/src/session/session-catalog.ts

@@ -34,6 +34,15 @@ export class SessionCatalog {
   }
 
   async list(currentSessionId?: SessionId): Promise<readonly SessionSummary[]> {
+    const summaries = await this.scan(currentSessionId);
+    return Object.freeze(summaries.slice(0, this.input.limit ?? 20));
+  }
+
+  async listAll(currentSessionId?: SessionId): Promise<readonly SessionSummary[]> {
+    return Object.freeze(await this.scan(currentSessionId));
+  }
+
+  private async scan(currentSessionId?: SessionId): Promise<SessionSummary[]> {
     const workspaceRoot = await this.workspaceRootPromise;
     const sessionsRoot = path.join(workspaceRoot, ".tinker", "sessions");
     let entries;
@@ -63,17 +72,14 @@ export class SessionCatalog {
       );
     }
 
-    return Object.freeze(
-      summaries
-        .filter(
-          (summary) =>
-            summary.turnCount > 0 ||
-            summary.status === "incomplete" ||
-            summary.status === "unavailable",
-        )
-        .sort((left, right) => right.updatedAt.localeCompare(left.updatedAt))
-        .slice(0, this.input.limit ?? 20),
-    );
+    return summaries
+      .filter(
+        (summary) =>
+          summary.turnCount > 0 ||
+          summary.status === "incomplete" ||
+          summary.status === "unavailable",
+      )
+      .sort((left, right) => right.updatedAt.localeCompare(left.updatedAt));
   }
 
   async get(

package/src/session/session-store.ts

@@ -4844,6 +4844,7 @@ export function decodeStoredToolRawResult(value: unknown): ToolRawResult {
       "bash",
       "task_list",
       "task_output",
+      "task_input",
       "task_stop",
       "web_search",
       "web_fetch",