klyro 1.0.10 → 1.0.12

package/dist/cli/eval.d.ts

@@ -79,6 +79,8 @@ export interface EvalResult {
         notes: string;
         skipped: boolean;
     };
+    /** Isolated workdir the scenario ran in (tmp unless --cwd). Debugging aid. */
+    workDir?: string;
 }
 export interface RunEvalOptions {
     inputPath: string;
@@ -90,10 +92,16 @@ export interface RunEvalOptions {
     model?: string;
     /** Live model id for grading `judge.rubric` (env endpoint + key required). */
     judgeModel?: string;
+    /**
+     * Shared workdir for JSONL scenarios. When omitted each scenario runs in
+     * a fresh tmp dir (deleted afterwards) so scripted tool calls can never
+     * touch the caller's directory. Pass explicitly to inspect artifacts.
+     */
+    cwd?: string;
 }
 export declare function runEval(opts: RunEvalOptions): Promise<number>;
 export declare function scriptedAdapterFromSpec(spec: Array<Array<unknown[]>> | undefined): ProviderAdapter;
 export declare function runScenario(sc: EvalScenario, judgeOpts?: {
     adapter: ProviderAdapter;
     model: string;
-}): Promise<EvalResult>;
+}, workDir?: string): Promise<EvalResult>;

package/dist/cli/eval.js

@@ -40,6 +40,9 @@
  * otherwise.
  */
 import * as fs from 'node:fs';
+import * as fsp from 'node:fs/promises';
+import * as os from 'node:os';
+import * as path from 'node:path';
 import * as readline from 'node:readline/promises';
 import { stdin as input, stdout, stderr } from 'node:process';
 import { run } from '../agent/runtime.js';
@@ -153,7 +156,7 @@ export async function runEval(opts) {
     const results = [];
     for (const sc of scenarios) {
         const start = Date.now();
-        const r = await runScenario(sc, judgeAdapter && opts.judgeModel ? { adapter: judgeAdapter, model: opts.judgeModel } : undefined);
+        const r = await runScenario(sc, judgeAdapter && opts.judgeModel ? { adapter: judgeAdapter, model: opts.judgeModel } : undefined, opts.cwd);
         r.durationMs = Date.now() - start;
         results.push(r);
         if (opts.output === 'json') {
@@ -242,35 +245,53 @@ function tupleToEvent(tuple) {
             throw new Error(`scriptedAdapterFromSpec: unknown event kind: ${kind}`);
     }
 }
-export async function runScenario(sc, judgeOpts) {
+export async function runScenario(sc, judgeOpts, workDir) {
     const failures = [];
     const model = sc.model ?? 'mock';
     const adapter = scriptedAdapterFromSpec(sc.scripted_events);
     const registry = builtinRegistry();
     const policy = new PolicyEngine(builtinRules(), DEFAULT_POLICY_CONFIG);
-    const result = await run({
-        task: sc.task,
-        cwd: process.cwd(),
-        model,
-        maxSteps: sc.maxSteps,
-        maxTokens: sc.maxTokens,
-        nonInteractive: true,
-        ...(sc.verify
-            ? {
-                verify: {
-                    enabled: true,
-                    ...(sc.verify.command !== undefined ? { command: sc.verify.command } : {}),
-                    ...(sc.verify.mode !== undefined ? { mode: sc.verify.mode } : {}),
-                },
+    // Isolation (fix: scripted tool calls must never run in the caller's
+    // directory — a JSONL scenario writing a.txt/b.txt used to pollute it).
+    // Explicit workDir is shared as-is (inspect artifacts); otherwise each
+    // scenario gets a fresh tmp dir that is removed afterwards.
+    const owned = !workDir;
+    const cwd = workDir ?? await fsp.mkdtemp(path.join(os.tmpdir(), 'klyro-eval-jsonl-'));
+    await fsp.mkdir(cwd, { recursive: true });
+    let result;
+    try {
+        result = await run({
+            task: sc.task,
+            cwd,
+            model,
+            maxSteps: sc.maxSteps,
+            maxTokens: sc.maxTokens,
+            nonInteractive: true,
+            ...(sc.verify
+                ? {
+                    verify: {
+                        enabled: true,
+                        ...(sc.verify.command !== undefined ? { command: sc.verify.command } : {}),
+                        ...(sc.verify.mode !== undefined ? { mode: sc.verify.mode } : {}),
+                    },
+                }
+                : {}),
+        }, {
+            adapter,
+            registry,
+            policy,
+            approval: new DenyAllApprovalPrompt(),
+            systemPrompt: ({ cwd }) => `You are Klyro. cwd=${cwd}.`,
+        });
+    }
+    finally {
+        if (owned) {
+            try {
+                await fsp.rm(cwd, { recursive: true, force: true });
             }
-            : {}),
-    }, {
-        adapter,
-        registry,
-        policy,
-        approval: new DenyAllApprovalPrompt(),
-        systemPrompt: ({ cwd }) => `You are Klyro. cwd=${cwd}.`,
-    });
+            catch { /* ignore */ }
+        }
+    }
     const exp = sc.expect ?? {};
     if (exp.status !== undefined && result.status !== exp.status) {
         failures.push(`status: expected ${exp.status}, got ${result.status}`);
@@ -311,6 +332,7 @@ export async function runScenario(sc, judgeOpts) {
         toolCalls: result.toolCalls,
         text: result.finalText,
         durationMs: 0,
+        workDir: cwd,
         ...(judge ? { judge } : {}),
     };
 }

package/dist/cli/repl.js

@@ -25,7 +25,7 @@ import { parseUnifiedDiff } from '../tui/diff-parser.js';
 import { parse } from './slash/parser.js';
 import { resolveProvider, providerHelp, lastProviderError } from '../providers.js';
 import { readVersion } from '../version.js';
-import { MouseFilter, MOUSE_ENABLE, MOUSE_DISABLE, PASTE_ENABLE, PASTE_DISABLE, PasteFilter, createReadWrapper } from '../tui/mouse.js';
+import { MouseFilter, MOUSE_ENABLE, MOUSE_DISABLE, PASTE_ENABLE, PASTE_DISABLE, PasteFilter, createReadWrapper, isMouseReportingEnabled } from '../tui/mouse.js';
 import { inferProviderFromBaseURL } from '../agent/registry.js';
 import { getDefaultSessionStore } from '../persistence/session.js';
 import { buildSystemPrompt, parseImageInput } from '../context/system-prompt.js';
@@ -372,13 +372,20 @@ export async function startRepl(opts = {}) {
     // ── Full-screen takeover like OpenCode — always when klyro in TTY (user explicitly wants it)
     // Scroll now works correctly via internal viewport, not native terminal scroll
     const isAltScreen = useTui && !!process.stdout.isTTY && process.env.KLYRO_NO_ALT !== '1';
+    // Mouse reporting is opt-in (KLYRO_MOUSE=1): when on, the terminal sends
+    // clicks/wheel to the app (wheel scrolls ±3 lines) but native text
+    // selection and right-click paste stop working. Default off so select to
+    // copy and right-click paste work out of the box; bracketed paste
+    // (keyboard paste) is unaffected and always enabled below.
+    const mouseReporting = isAltScreen && isMouseReportingEnabled();
     const enterAlt = () => {
         if (!isAltScreen)
             return;
         try {
             process.stdout.write('\x1b[?1049h\x1b[?25l'); // alt screen + hide cursor
             process.stdout.write('\x1b[H\x1b[2J'); // home + clear
-            process.stdout.write(MOUSE_ENABLE); // wheel events (SGR), see tui/mouse.ts
+            if (mouseReporting)
+                process.stdout.write(MOUSE_ENABLE); // wheel events (SGR), see tui/mouse.ts
         }
         catch { /* ignore */ }
     };
@@ -386,7 +393,8 @@ export async function startRepl(opts = {}) {
         if (!isAltScreen)
             return;
         try {
-            process.stdout.write(MOUSE_DISABLE);
+            if (mouseReporting)
+                process.stdout.write(MOUSE_DISABLE);
             process.stdout.write('\x1b[?25h\x1b[?1049l'); // show cursor + leave alt
         }
         catch { /* ignore */ }
@@ -2681,8 +2689,8 @@ export async function startRepl(opts = {}) {
                         '  Enter send · Shift+Enter newline · Tab complete slash · Esc drop queued / Esc×2 cancel run',
                         '  Ctrl+C cancel (1st) / quit (2nd) · Ctrl+O expand last tool group · Ctrl+G jump bottom',
                         '  PgUp/PgDn or Ctrl+U/Ctrl+D half-page · Ctrl+Home/End top/bottom · Home/End jump · Space jump to unread',
-                        '  Ctrl+B/F page · Shift/Ctrl+↑/↓ line · ↑/↓ history (with text) else scroll · wheel ±3 lines',
-                        '  Shift+drag selects · /vim toggles vim input mode · /keymap <note> saves a display note',
+                        '  Ctrl+B/F page · Shift/Ctrl+↑/↓ line · ↑/↓ input history · PgUp/Dn scroll (KLYRO_MOUSE=1 adds wheel ±3 lines)',
+                        '  Text selection/copy and right-click paste work natively · Shift+Enter newline · /vim toggles vim input mode · /keymap <note> saves a display note',
                     ].join('\n'),
                 });
                 return;

package/dist/index.js

@@ -390,6 +390,7 @@ async function main() {
         .option('--parallel <n>', 'Parallelism (default 1)', (v) => parsePositiveInt('--parallel', v))
         .option('--model <id>', 'Model for eval')
         .option('--judge-model <id>', 'Live model id for grading judge.rubric (needs endpoint + key)')
+        .option('--cwd <path>', 'Shared scenario workdir (default: isolated tmp per scenario)')
         .action(async (input, opts) => {
         const output = (opts.output ?? 'human');
         if (opts.suite) {
@@ -400,7 +401,7 @@ async function main() {
             process.stderr.write('klyro eval: missing input (provide <input> or --suite)\n');
             process.exit(2);
         }
-        const code = await runEval({ inputPath: input, output, suite: opts.suite, filter: opts.filter, runs: opts.runs, parallel: opts.parallel, model: opts.model, judgeModel: opts.judgeModel });
+        const code = await runEval({ inputPath: input, output, suite: opts.suite, filter: opts.filter, runs: opts.runs, parallel: opts.parallel, model: opts.model, judgeModel: opts.judgeModel, cwd: opts.cwd });
         process.exit(code);
     });
     program

package/dist/tui/app.js

@@ -1052,16 +1052,16 @@ export function App(props) {
             void props.onSlash({ kind: 'quit' });
             return;
         }
-        // Contextual ↑/↓ (§8.3): text in buffer (or browsing) → history;
-        // empty buffer → scroll viewport one line.
+        // Contextual ↑/↓ (§8.3): history always wins when entries exist —
+        // empty input + ↑ recalls the last prompt (standard REPL behavior),
+        // typing filters by prefix is unnecessary so plain recall applies;
+        // the viewport scrolls only when there is no history to show.
         if (key.upArrow && !key.shift && !key.ctrl) {
-            if (input.trim() !== '' || histIdx !== null) {
-                if (history.length > 0) {
-                    const next = histIdx === null ? history.length - 1 : Math.max(0, histIdx - 1);
-                    setHistIdx(next);
-                    setInput(history[next] ?? '');
-                    setVimCursor(null);
-                }
+            if (history.length > 0) {
+                const next = histIdx === null ? history.length - 1 : Math.max(0, histIdx - 1);
+                setHistIdx(next);
+                setInput(history[next] ?? '');
+                setVimCursor(null);
                 return;
             }
             if (isFullscreen && maxTop > 0) {

package/dist/tui/app.test.js

@@ -247,6 +247,7 @@ describe('App', () => {
         expect(lastFrame() ?? '').toContain('first recallable prompt');
     });
     it('↑ on empty input scrolls one line instead of history', async () => {
+        // No prompts submitted yet → history is empty → viewport scrolls.
         const { stdin, lastFrame } = render(_jsx(App, { initialModel: "m", maxSteps: 10, cwd: "/test", onPrompt: async () => { }, onSlash: async () => { }, isFullscreen: true, initialTranscript: makeInitialTranscript(25) }));
         await new Promise((r) => setTimeout(r, 50));
         stdin.write(KEY_HOME);
@@ -256,6 +257,30 @@ describe('App', () => {
         await new Promise((r) => setTimeout(r, 30));
         expect(lastFrame() ?? '').toMatch(/MSG-00-tag/);
     });
+    it('↑ on empty input recalls history newest-first; ↓ browses back (user report)', async () => {
+        const onPrompt = vi.fn(async () => { });
+        const { stdin, lastFrame } = render(_jsx(App, { initialModel: "m", maxSteps: 10, cwd: "/test", onPrompt: onPrompt, onSlash: async () => { } }));
+        stdin.write('hi');
+        await new Promise((r) => setTimeout(r, 20));
+        stdin.write('\x0d');
+        await new Promise((r) => setTimeout(r, 50));
+        stdin.write('second prompt');
+        await new Promise((r) => setTimeout(r, 20));
+        stdin.write('\x0d');
+        await new Promise((r) => setTimeout(r, 50));
+        expect(onPrompt).toHaveBeenCalledTimes(2);
+        // Input is empty after each submit; plain ↑ must recall, not scroll.
+        stdin.write('\x1b[A');
+        await new Promise((r) => setTimeout(r, 30));
+        expect(lastFrame() ?? '').toContain('second prompt');
+        stdin.write('\x1b[A');
+        await new Promise((r) => setTimeout(r, 30));
+        expect(lastFrame() ?? '').toContain('hi');
+        // ↓ walks back toward newer entries.
+        stdin.write('\x1b[B');
+        await new Promise((r) => setTimeout(r, 30));
+        expect(lastFrame() ?? '').toContain('second prompt');
+    });
     it('/c shows top-6 suggestions and Tab completes', async () => {
         const { stdin, lastFrame } = render(_jsx(App, { initialModel: "m", maxSteps: 10, cwd: "/test", onPrompt: async () => { }, onSlash: async () => { } }));
         stdin.write('/c');

package/dist/tui/mouse.d.ts

@@ -34,6 +34,16 @@ export declare class MouseFilter {
 }
 export declare const MOUSE_ENABLE = "\u001B[?1000h\u001B[?1006h";
 export declare const MOUSE_DISABLE = "\u001B[?1000l\u001B[?1006l";
+/**
+ * Whether the TUI may request terminal mouse reporting (wheel scrolling).
+ *
+ * Default OFF: with button reporting enabled the terminal routes
+ * selection clicks and right-click paste to the app (which swallows them),
+ * so native select-to-copy and right-click-paste break. Native selection
+ * works out of the box; set `KLYRO_MOUSE=1` to opt into wheel scrolling
+ * (Shift+drag still selects natively in most terminals).
+ */
+export declare function isMouseReportingEnabled(env?: Readonly<Record<string, string | undefined>>): boolean;
 export declare const PASTE_START = "\u001B[200~";
 export declare const PASTE_END = "\u001B[201~";
 export declare const PASTE_ENABLE = "\u001B[?2004h";

package/dist/tui/mouse.js

@@ -92,6 +92,18 @@ export class MouseFilter {
 }
 export const MOUSE_ENABLE = '\x1b[?1000h\x1b[?1006h'; // button events + SGR coords
 export const MOUSE_DISABLE = '\x1b[?1000l\x1b[?1006l';
+/**
+ * Whether the TUI may request terminal mouse reporting (wheel scrolling).
+ *
+ * Default OFF: with button reporting enabled the terminal routes
+ * selection clicks and right-click paste to the app (which swallows them),
+ * so native select-to-copy and right-click-paste break. Native selection
+ * works out of the box; set `KLYRO_MOUSE=1` to opt into wheel scrolling
+ * (Shift+drag still selects natively in most terminals).
+ */
+export function isMouseReportingEnabled(env = process.env) {
+    return env.KLYRO_MOUSE === '1';
+}
 export const PASTE_START = '\x1b[200~';
 export const PASTE_END = '\x1b[201~';
 export const PASTE_ENABLE = '\x1b[?2004h'; // bracketed paste: terminal wraps pastes

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "klyro",
-  "version": "1.0.10",
+  "version": "1.0.12",
   "description": "Klyro — autonomous coding harness CLI that streams from any OpenAI-compatible or Anthropic LLM endpoint.",
   "type": "module",
   "main": "dist/index.js",