klyro 0.1.1

package/dist/context/snippets.d.ts

@@ -0,0 +1,17 @@
+/**
+ * Snippet reader — read a slice of a file (line window). Used by the
+ * runtime when the model needs a specific region, not the whole file.
+ */
+export interface Snippet {
+    path: string;
+    startLine: number;
+    endLine: number;
+    content: string;
+}
+export interface SnippetOptions {
+    cwd: string;
+    path: string;
+    startLine?: number;
+    maxLines?: number;
+}
+export declare function readSnippet(opts: SnippetOptions): Promise<Snippet>;

package/dist/context/snippets.js

@@ -0,0 +1,15 @@
+/**
+ * Snippet reader — read a slice of a file (line window). Used by the
+ * runtime when the model needs a specific region, not the whole file.
+ */
+import * as fs from 'node:fs/promises';
+import { resolveWithinCwd } from '../policy/path-guard.js';
+export async function readSnippet(opts) {
+    const { resolved } = resolveWithinCwd(opts.cwd, opts.path);
+    const content = await fs.readFile(resolved, 'utf-8');
+    const lines = content.split(/\r?\n/);
+    const start = Math.max(1, opts.startLine ?? 1);
+    const max = Math.min(opts.maxLines ?? 200, lines.length - start + 1);
+    const end = start + max - 1;
+    return { path: opts.path, startLine: start, endLine: end, content: lines.slice(start - 1, end).join('\n') };
+}

package/dist/context/tokenizer.d.ts

@@ -0,0 +1,49 @@
+/**
+ * Token budget — keeps each model call under a hard cap by trimming the
+ * transcript before sending it to the provider.
+ *
+ * For MVP we count tokens with a simple heuristic (chars/4) so we don't
+ * pull in a heavy BPE dependency. The cap is conservative; the runtime
+ * never overflows the model's true window because the heuristic
+ * overestimates mixed text.
+ *
+ * Strategy:
+ *   1. Always preserve the system prompt, the latest user task, and the
+ *      latest assistant message.
+ *   2. If still over budget, drop the oldest tool_result observations
+ *      that have been "consumed" (a later assistant message referenced
+ *      their tool_call_id).
+ *   3. If still over budget, summarize the surviving tail into a single
+ *      user message ("Earlier in this session: …").
+ */
+import type { Message } from '../agent/message.js';
+export interface TokenBudget {
+    /** Total input cap (system + messages). */
+    total: number;
+    /** Reserve for the response (we don't trim these; the cap is on input). */
+    reservedOutput: number;
+}
+export interface BudgetCheck {
+    ok: boolean;
+    used: number;
+    cap: number;
+}
+export declare function estimateTokens(s: string): number;
+export declare function estimateMessage(m: Message): number;
+/** Total input tokens for a transcript + optional system prompt. */
+export declare function totalTokens(system: string | undefined, messages: Message[]): number;
+/** True if the input fits under the budget cap. */
+export declare function withinBudget(system: string | undefined, messages: Message[], budget: TokenBudget): BudgetCheck;
+/**
+ * Trim a transcript to fit under the budget. Strategy:
+ *   - Keep first user task (so the model never forgets the goal).
+ *   - Keep the last 2 messages verbatim.
+ *   - Compress intermediate messages: tool_result blocks shorter, text
+ *     blocks truncated to 500 chars each.
+ * Returns the new transcript plus a count of dropped observations.
+ */
+export declare function compressTranscript(system: string | undefined, messages: Message[], budget: TokenBudget): {
+    system: string | undefined;
+    messages: Message[];
+    dropped: number;
+};

package/dist/context/tokenizer.js

@@ -0,0 +1,97 @@
+/**
+ * Token budget — keeps each model call under a hard cap by trimming the
+ * transcript before sending it to the provider.
+ *
+ * For MVP we count tokens with a simple heuristic (chars/4) so we don't
+ * pull in a heavy BPE dependency. The cap is conservative; the runtime
+ * never overflows the model's true window because the heuristic
+ * overestimates mixed text.
+ *
+ * Strategy:
+ *   1. Always preserve the system prompt, the latest user task, and the
+ *      latest assistant message.
+ *   2. If still over budget, drop the oldest tool_result observations
+ *      that have been "consumed" (a later assistant message referenced
+ *      their tool_call_id).
+ *   3. If still over budget, summarize the surviving tail into a single
+ *      user message ("Earlier in this session: …").
+ */
+export function estimateTokens(s) {
+    return Math.ceil(s.length / 4);
+}
+export function estimateMessage(m) {
+    let n = 4; // role + structural overhead
+    for (const b of m.content)
+        n += estimateBlock(b);
+    return n;
+}
+function estimateBlock(b) {
+    if (b.kind === 'text')
+        return estimateTokens(b.text) + 4;
+    if (b.kind === 'tool_use')
+        return estimateTokens(b.name) + estimateTokens(JSON.stringify(b.input)) + 16;
+    if (b.kind === 'tool_result') {
+        const out = typeof b.output === 'string' ? b.output : JSON.stringify(b.output ?? '');
+        return estimateTokens(out) + estimateTokens(b.name) + 16;
+    }
+    return 0;
+}
+/** Total input tokens for a transcript + optional system prompt. */
+export function totalTokens(system, messages) {
+    let n = system ? estimateTokens(system) : 0;
+    for (const m of messages)
+        n += estimateMessage(m);
+    return n;
+}
+/** True if the input fits under the budget cap. */
+export function withinBudget(system, messages, budget) {
+    const used = totalTokens(system, messages);
+    return { ok: used <= budget.total, used, cap: budget.total };
+}
+/**
+ * Trim a transcript to fit under the budget. Strategy:
+ *   - Keep first user task (so the model never forgets the goal).
+ *   - Keep the last 2 messages verbatim.
+ *   - Compress intermediate messages: tool_result blocks shorter, text
+ *     blocks truncated to 500 chars each.
+ * Returns the new transcript plus a count of dropped observations.
+ */
+export function compressTranscript(system, messages, budget) {
+    const result = messages.slice();
+    let dropped = 0;
+    // Phase 1: find tool_call_ids referenced in any later assistant message.
+    const consumed = new Set();
+    for (const m of result) {
+        if (m.role !== 'assistant')
+            continue;
+        for (const b of m.content)
+            if (b.kind === 'tool_use')
+                consumed.add(b.id);
+    }
+    // Phase 2: shrink tool_result blocks older than the last 4 messages.
+    const tailStart = Math.max(0, result.length - 4);
+    for (let i = 0; i < tailStart; i++) {
+        const m = result[i];
+        if (!m || m.role !== 'tool')
+            continue;
+        for (const b of m.content) {
+            if (b.kind !== 'tool_result')
+                continue;
+            if (!consumed.has(b.toolCallId)) {
+                // Drop the observation entirely.
+                b.output = '[earlier observation removed to fit context]';
+                dropped++;
+            }
+            else if (typeof b.output === 'string' && b.output.length > 400) {
+                b.output = b.output.slice(0, 400) + '... [truncated]';
+            }
+        }
+    }
+    // Phase 3: hard cap — drop oldest messages until we fit, but always
+    // preserve the first message (the user's task).
+    while (totalTokens(system, result) > budget.total && result.length > 2) {
+        result.splice(1, 1);
+        dropped++;
+    }
+    return { system, messages: result, dropped };
+}

package/dist/eval/harness.d.ts

@@ -0,0 +1,48 @@
+/**
+ * Eval harness — runs a sequence of scripted tasks against the runtime
+ * using a MockProviderAdapter. Verifies (a) the runtime completes the
+ * expected tool-call sequence, (b) the verification engine detects the
+ * right failure type, and (c) the compressor preserves the goal.
+ *
+ * This is the MVP gate per Devolopment-plan.md / docs/plan.md §10: a
+ * reproducible suite of programmatic tasks. Real repo tasks come in v1.0.
+ */
+import type { StreamEvent } from '../agent/provider-adapter.js';
+export interface ScriptedTask {
+    id: string;
+    description: string;
+    task: string;
+    /** A list of event scripts — each inner list is one assistant turn. */
+    script: StreamEvent[][];
+    /** Optional verify command to run after the runtime finishes. */
+    verifyCommand?: string;
+    /** Expected final status. */
+    expectStatus: 'complete' | 'max_steps' | 'aborted' | 'verify_failed' | 'no_final';
+    /** Expected tool-call count. */
+    expectToolCalls?: number;
+}
+export interface TaskResult {
+    id: string;
+    status: 'pass' | 'fail';
+    details: string;
+    observedStatus?: string;
+    observedToolCalls?: number;
+    durationMs: number;
+}
+export declare function runTask(t: ScriptedTask): Promise<TaskResult>;
+export interface HarnessSummary {
+    total: number;
+    passed: number;
+    failed: number;
+    passRate: number;
+    results: TaskResult[];
+    durationMs: number;
+}
+export declare function runHarness(tasks: ScriptedTask[]): Promise<HarnessSummary>;
+/** Format a harness summary as a markdown report. */
+export declare function formatReport(summary: HarnessSummary): string;
+/** Hook the harness up to a session + audit log so task runs are durable. */
+export declare function runHarnessWithPersistence(tasks: ScriptedTask[], opts: {
+    storeDir: string;
+    auditPath: string;
+}): Promise<HarnessSummary>;

package/dist/eval/harness.js

@@ -0,0 +1,149 @@
+/**
+ * Eval harness — runs a sequence of scripted tasks against the runtime
+ * using a MockProviderAdapter. Verifies (a) the runtime completes the
+ * expected tool-call sequence, (b) the verification engine detects the
+ * right failure type, and (c) the compressor preserves the goal.
+ *
+ * This is the MVP gate per Devolopment-plan.md / docs/plan.md §10: a
+ * reproducible suite of programmatic tasks. Real repo tasks come in v1.0.
+ */
+import { run } from '../agent/runtime.js';
+import { ToolRegistry } from '../tools/registry.js';
+import { readFileTool } from '../tools/fs/read-file.js';
+import { writeFileTool } from '../tools/fs/write-file.js';
+import { editFileTool } from '../tools/fs/edit-file.js';
+import { shellExecTool } from '../tools/shell/shell-exec.js';
+import { runVerifyTool } from '../tools/verify/run-verify.js';
+import { listDirTool } from '../tools/fs/list-dir.js';
+import { globTool } from '../tools/search/glob.js';
+import { grepTool } from '../tools/search/grep.js';
+import { gitStatusTool } from '../tools/git/git-status.js';
+import { gitDiffTool } from '../tools/git/git-diff.js';
+import { PolicyEngine, builtinRules, DEFAULT_POLICY_CONFIG } from '../policy/engine.js';
+import { DenyAllApprovalPrompt } from '../policy/approval.js';
+import { defaultSystemPrompt } from '../agent/runtime.js';
+import { verify } from '../verification/engine.js';
+import { SessionStore } from '../persistence/store.js';
+import { AuditLog } from '../persistence/audit.js';
+import * as fs from 'node:fs/promises';
+import * as os from 'node:os';
+import * as path from 'node:path';
+function scriptedAdapter(script) {
+    let i = 0;
+    return {
+        id: 'mock',
+        async *stream() {
+            if (i < script.length) {
+                const turn = script[i++];
+                if (turn)
+                    for (const ev of turn)
+                        yield ev;
+            }
+        },
+    };
+}
+export async function runTask(t) {
+    const start = Date.now();
+    const cwd = path.join(os.tmpdir(), 'klyro-eval-' + t.id + '-' + Math.random().toString(36).slice(2));
+    await fs.mkdir(cwd, { recursive: true });
+    const reg = new ToolRegistry()
+        .register(readFileTool).register(writeFileTool).register(editFileTool)
+        .register(listDirTool).register(globTool).register(grepTool)
+        .register(shellExecTool).register(runVerifyTool)
+        .register(gitStatusTool).register(gitDiffTool);
+    const policy = new PolicyEngine(builtinRules(), DEFAULT_POLICY_CONFIG);
+    try {
+        const result = await run({ task: t.task, cwd, model: 'mock', maxSteps: 12, nonInteractive: true }, { adapter: scriptedAdapter(t.script), registry: reg, policy, approval: new DenyAllApprovalPrompt(), systemPrompt: defaultSystemPrompt });
+        let observedStatus = result.status;
+        let verifyFailure;
+        if (t.verifyCommand) {
+            const v = await verify({ cwd, command: t.verifyCommand });
+            if (!v.ok) {
+                observedStatus = 'verify_failed';
+                verifyFailure = v.failure?.type ?? 'unknown';
+            }
+        }
+        let details = '';
+        let pass = result.status === t.expectStatus;
+        if (t.expectToolCalls !== undefined && result.toolCalls !== t.expectToolCalls) {
+            details += ` toolCalls=${result.toolCalls} expected=${t.expectToolCalls};`;
+            pass = false;
+        }
+        if (verifyFailure) {
+            details += ` verifyFailure=${verifyFailure};`;
+        }
+        return {
+            id: t.id,
+            status: pass ? 'pass' : 'fail',
+            details: details.trim() || `status=${observedStatus}`,
+            observedStatus,
+            observedToolCalls: result.toolCalls,
+            durationMs: Date.now() - start,
+        };
+    }
+    finally {
+        try {
+            await fs.rm(cwd, { recursive: true, force: true });
+        }
+        catch { }
+    }
+}
+export async function runHarness(tasks) {
+    const start = Date.now();
+    const results = [];
+    for (const t of tasks)
+        results.push(await runTask(t));
+    const passed = results.filter((r) => r.status === 'pass').length;
+    return {
+        total: results.length,
+        passed,
+        failed: results.length - passed,
+        passRate: results.length === 0 ? 0 : passed / results.length,
+        results,
+        durationMs: Date.now() - start,
+    };
+}
+/** Format a harness summary as a markdown report. */
+export function formatReport(summary) {
+    const lines = [
+        `# Klyro Harness Report`,
+        ``,
+        `**Total:** ${summary.total}    **Passed:** ${summary.passed}    **Failed:** ${summary.failed}    **Pass rate:** ${(summary.passRate * 100).toFixed(1)}%`,
+        `**Wall time:** ${(summary.durationMs / 1000).toFixed(1)}s`,
+        ``,
+        `| Task | Status | Details | Tool calls | Duration |`,
+        `|------|--------|---------|------------|----------|`,
+        ...summary.results.map((r) => `| ${r.id} | ${r.status} | ${r.details} | ${r.observedToolCalls ?? '-'} | ${r.durationMs}ms |`),
+    ];
+    return lines.join('\n');
+}
+/** Hook the harness up to a session + audit log so task runs are durable. */
+export async function runHarnessWithPersistence(tasks, opts) {
+    const store = new SessionStore(opts.storeDir);
+    const audit = new AuditLog(opts.auditPath);
+    const start = Date.now();
+    const results = [];
+    for (const t of tasks) {
+        await audit.write({ kind: 'verification_attempted', sessionId: t.id, command: t.task, ts: Date.now() });
+        const r = await runTask(t);
+        results.push(r);
+        await audit.write({
+            kind: r.status === 'pass' ? 'verification_succeeded' : 'verification_failed',
+            sessionId: t.id,
+            exitCode: 0,
+            type: r.details,
+            ts: Date.now(),
+        });
+        // Suppress unused warning.
+        void store;
+    }
+    const passed = results.filter((r) => r.status === 'pass').length;
+    return {
+        total: results.length,
+        passed,
+        failed: results.length - passed,
+        passRate: results.length === 0 ? 0 : passed / results.length,
+        results,
+        durationMs: Date.now() - start,
+    };
+}

package/dist/eval/tasks.d.ts

@@ -0,0 +1,6 @@
+/**
+ * The default MVP task suite. All tasks are programmatic — they use a
+ * scripted mock provider, so the suite is hermetic and CI-friendly.
+ */
+import type { ScriptedTask } from './harness.js';
+export declare const MVP_TASKS: ScriptedTask[];

package/dist/eval/tasks.js

@@ -0,0 +1,98 @@
+/**
+ * The default MVP task suite. All tasks are programmatic — they use a
+ * scripted mock provider, so the suite is hermetic and CI-friendly.
+ */
+export const MVP_TASKS = [
+    {
+        id: 't1-direct-answer',
+        description: 'Model says hello without calling any tool.',
+        task: 'say hi',
+        script: [[
+                { kind: 'message_start' },
+                { kind: 'text_delta', text: 'Hello there.' },
+                { kind: 'message_end', finishReason: 'stop' },
+            ]],
+        expectStatus: 'complete',
+        expectToolCalls: 0,
+    },
+    {
+        id: 't2-write-then-answer',
+        description: 'Model writes a file then confirms.',
+        task: 'create note.txt with content "hi"',
+        script: [
+            [
+                { kind: 'message_start' },
+                { kind: 'tool_call_start', id: 'c1', name: 'write_file' },
+                { kind: 'tool_call_delta', id: 'c1', argsJson: '{"path":"note.txt","content":"hi"}' },
+                { kind: 'tool_call_end', id: 'c1' },
+                { kind: 'message_end', finishReason: 'tool_calls' },
+            ],
+            [
+                { kind: 'message_start' },
+                { kind: 'text_delta', text: 'Created.' },
+                { kind: 'message_end', finishReason: 'stop' },
+            ],
+        ],
+        expectStatus: 'complete',
+        expectToolCalls: 1,
+    },
+    {
+        id: 't3-policy-deny',
+        description: 'Model tries to escape the cwd; policy denies.',
+        task: 'escape',
+        script: [
+            [
+                { kind: 'message_start' },
+                { kind: 'tool_call_start', id: 'c1', name: 'write_file' },
+                { kind: 'tool_call_delta', id: 'c1', argsJson: '{"path":"../escape.txt","content":"x"}' },
+                { kind: 'tool_call_end', id: 'c1' },
+                { kind: 'message_end', finishReason: 'tool_calls' },
+            ],
+            [
+                { kind: 'message_start' },
+                { kind: 'text_delta', text: 'OK' },
+                { kind: 'message_end', finishReason: 'stop' },
+            ],
+        ],
+        expectStatus: 'complete',
+        expectToolCalls: 1,
+    },
+    {
+        id: 't4-max-steps',
+        description: 'Model never finishes; runtime returns max_steps.',
+        task: 'loop forever',
+        script: Array.from({ length: 15 }, () => [
+            { kind: 'message_start' },
+            { kind: 'tool_call_start', id: 'c1', name: 'write_file' },
+            { kind: 'tool_call_delta', id: 'c1', argsJson: '{"path":"a.txt","content":"x"}' },
+            { kind: 'tool_call_end', id: 'c1' },
+            { kind: 'message_end', finishReason: 'tool_calls' },
+        ]),
+        expectStatus: 'max_steps',
+        expectToolCalls: 12, // maxSteps=12 in harness, one tool call per step
+    },
+    {
+        id: 't6-multitool',
+        description: 'Model issues two tool calls in one turn.',
+        task: 'create two files',
+        script: [
+            [
+                { kind: 'message_start' },
+                { kind: 'tool_call_start', id: 'c1', name: 'write_file' },
+                { kind: 'tool_call_delta', id: 'c1', argsJson: '{"path":"a.txt","content":"1"}' },
+                { kind: 'tool_call_end', id: 'c1' },
+                { kind: 'tool_call_start', id: 'c2', name: 'write_file' },
+                { kind: 'tool_call_delta', id: 'c2', argsJson: '{"path":"b.txt","content":"2"}' },
+                { kind: 'tool_call_end', id: 'c2' },
+                { kind: 'message_end', finishReason: 'tool_calls' },
+            ],
+            [
+                { kind: 'message_start' },
+                { kind: 'text_delta', text: 'Both written.' },
+                { kind: 'message_end', finishReason: 'stop' },
+            ],
+        ],
+        expectStatus: 'complete',
+        expectToolCalls: 2,
+    },
+];

package/dist/index.d.ts

@@ -0,0 +1,14 @@
+#!/usr/bin/env node
+/**
+ * Klyro — autonomous coding harness CLI.
+ *
+ *   klyro                 start the TUI REPL
+ *   klyro run "<prompt>"  one-shot autonomous task, streams to stdout
+ *   klyro chat "<prompt>" legacy one-shot streamed chat (kept for compat)
+ *   klyro chat            legacy interactive REPL (kept for compat)
+ *   klyro eval <file>     run scripted scenarios (JSONL)
+ *
+ * Provider: OpenAI-compatible /v1/chat/completions endpoint.
+ * Configure via env: KLYRO_BASE_URL, KLYRO_API_KEY, KLYRO_MODEL.
+ */
+export {};

package/dist/index.js

@@ -0,0 +1,128 @@
+#!/usr/bin/env node
+/**
+ * Klyro — autonomous coding harness CLI.
+ *
+ *   klyro                 start the TUI REPL
+ *   klyro run "<prompt>"  one-shot autonomous task, streams to stdout
+ *   klyro chat "<prompt>" legacy one-shot streamed chat (kept for compat)
+ *   klyro chat            legacy interactive REPL (kept for compat)
+ *   klyro eval <file>     run scripted scenarios (JSONL)
+ *
+ * Provider: OpenAI-compatible /v1/chat/completions endpoint.
+ * Configure via env: KLYRO_BASE_URL, KLYRO_API_KEY, KLYRO_MODEL.
+ */
+import { Command } from 'commander';
+import { chat } from './chat.js';
+import { repl } from './repl.js';
+import { runOnce } from './cli/run.js';
+import { runEval } from './cli/eval.js';
+const VERSION = '0.1.0';
+function parsePositiveInt(name, v) {
+    const n = Number(v);
+    if (!Number.isFinite(n) || n <= 0)
+        throw new Error(`invalid ${name}: ${v}`);
+    return n;
+}
+async function main() {
+    const program = new Command();
+    program
+        .name('klyro')
+        .description('Klyro — autonomous coding harness')
+        .version(VERSION, '-V, --version', 'Print the version number')
+        .helpOption('-h, --help', 'Print this help message')
+        .showHelpAfterError();
+    // Default action: TUI REPL. For now this falls back to the legacy REPL
+    // until the Ink-based TUI lands. The wiring point is cli/repl.ts.
+    program
+        .action(async () => {
+        await repl('You are a helpful assistant.');
+    });
+    program
+        .command('run <prompt>')
+        .description('Run a one-shot autonomous task. Streams text to stdout; tool calls to stderr.')
+        .option('-m, --model <id>', 'Model id (default: env KLYRO_MODEL)')
+        .option('--max-steps <n>', 'Max agent steps (default 30)', (v) => parsePositiveInt('--max-steps', v))
+        .option('--max-tokens <n>', 'Max output tokens per step', (v) => parsePositiveInt('--max-tokens', v))
+        .option('--temperature <n>', 'Sampling temperature', (v) => {
+        const n = Number(v);
+        if (!Number.isFinite(n) || n < 0 || n > 2)
+            throw new Error(`invalid --temperature: ${v}`);
+        return n;
+    })
+        .option('--timeout <ms>', 'Request timeout in ms (default: env KLYRO_TIMEOUT_MS or 60000)', (v) => parsePositiveInt('--timeout', v))
+        .option('--base-url <url>', 'Override KLYRO_BASE_URL')
+        .option('--api-key <key>', 'Override KLYRO_API_KEY')
+        .option('--output <mode>', 'Output mode: human (default), json (one JSON per line), silent')
+        .option('--provider <name>', 'Provider: openai (default) or anthropic')
+        .option('--dry-run', 'Print the prompt assembly (system, tools, task) and exit without calling the model')
+        .option('--resume <file>', 'Resume from a saved transcript JSON file (must have a "transcript" field)')
+        .action(async (prompt, opts) => {
+        const model = opts.model ?? process.env.KLYRO_MODEL;
+        if (!model) {
+            process.stderr.write('klyro: KLYRO_MODEL is not set (or pass --model)\n');
+            process.exit(2);
+        }
+        const output = (opts.output ?? 'human');
+        if (output !== 'human' && output !== 'json' && output !== 'silent') {
+            process.stderr.write(`klyro: invalid --output: ${output} (expected human|json|silent)\n`);
+            process.exit(2);
+        }
+        const provider = (opts.provider ?? 'openai');
+        if (provider !== 'openai' && provider !== 'anthropic') {
+            process.stderr.write(`klyro: invalid --provider: ${provider} (expected openai|anthropic)\n`);
+            process.exit(2);
+        }
+        try {
+            const code = await runOnce({
+                task: prompt,
+                cwd: process.cwd(),
+                model,
+                maxSteps: opts.maxSteps,
+                maxTokens: opts.maxTokens,
+                temperature: opts.temperature,
+                timeoutMs: opts.timeout,
+                baseUrl: opts.baseUrl,
+                apiKey: opts.apiKey,
+                output,
+                provider,
+                dryRun: !!opts.dryRun,
+                resumePath: opts.resume,
+            });
+            process.exit(code);
+        }
+        catch (err) {
+            process.stderr.write(`klyro: ${err instanceof Error ? err.message : String(err)}\n`);
+            process.exit(2);
+        }
+    });
+    program
+        .command('chat [prompt]')
+        .description('Legacy streamed chat. Without a prompt, start an interactive REPL.')
+        .option('-s, --system <text>', 'System message', 'You are a helpful assistant.')
+        .option('-m, --model <id>', 'Override the model (default: env KLYRO_MODEL)')
+        .option('-t, --timeout <ms>', 'Request timeout in ms (default: env KLYRO_TIMEOUT_MS or 60000)', (v) => parsePositiveInt('-t/--timeout', v))
+        .action(async (prompt, opts) => {
+        if (!prompt) {
+            await repl(opts.system);
+        }
+        else {
+            await chat(prompt, opts.system, opts.model, { timeoutMs: opts.timeout });
+        }
+    });
+    program
+        .command('eval <input>')
+        .description('Run scripted scenarios from a JSONL file against the agent runtime. Exits 0 if all pass, 1 otherwise.')
+        .option('--output <mode>', 'Output mode: human (default), json (one JSON per line)')
+        .action(async (input, opts) => {
+        const output = (opts.output ?? 'human');
+        const code = await runEval({ inputPath: input, output });
+        process.exit(code);
+    });
+    await program.parseAsync(process.argv);
+}
+main().catch((err) => {
+    // Last-resort: anything that escaped the command handlers lands here.
+    // eslint-disable-next-line no-console
+    console.error(`klyro: ${err instanceof Error ? err.message : String(err)}`);
+    process.exit(1);
+});

package/dist/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":";AACA;;;;;;;GAOG;AAEH,OAAO,EAAE,OAAO,EAAE,MAAM,WAAW,CAAC;AACpC,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AACjC,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AAEjC,MAAM,OAAO,GAAG,OAAO,CAAC;AAExB,KAAK,UAAU,IAAI;IACjB,MAAM,OAAO,GAAG,IAAI,OAAO,EAAE,CAAC;IAE9B,OAAO;SACJ,IAAI,CAAC,MAAM,CAAC;SACZ,WAAW,CAAC,wBAAwB,CAAC;SACrC,OAAO,CAAC,OAAO,EAAE,eAAe,EAAE,0BAA0B,CAAC;SAC7D,UAAU,CAAC,YAAY,EAAE,yBAAyB,CAAC;SACnD,kBAAkB,EAAE,CAAC;IAExB,OAAO;SACJ,OAAO,CAAC,eAAe,CAAC;SACxB,WAAW,CAAC,qFAAqF,CAAC;SAClG,MAAM,CAAC,qBAAqB,EAAE,gBAAgB,EAAE,8BAA8B,CAAC;SAC/E,MAAM,CAAC,kBAAkB,EAAE,8CAA8C,CAAC;SAC1E,MAAM,CAAC,oBAAoB,EAAE,+DAA+D,EAAE,CAAC,CAAC,EAAE,EAAE;QACnG,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC;QACpB,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,oBAAoB,CAAC,EAAE,CAAC,CAAC;QAC5E,OAAO,CAAC,CAAC;IACX,CAAC,CAAC;SACD,MAAM,CAAC,KAAK,EAAE,MAA0B,EAAE,IAA0D,EAAE,EAAE;QACvG,IAAI,CAAC,MAAM,EAAE,CAAC;YACZ,MAAM,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QAC1B,CAAC;aAAM,CAAC;YACN,MAAM,IAAI,CAAC,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,IAAI,CAAC,KAAK,EAAE,EAAE,SAAS,EAAE,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC;QAC3E,CAAC;IACH,CAAC,CAAC,CAAC;IAEL,MAAM,OAAO,CAAC,UAAU,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;AACzC,CAAC;AAED,IAAI,EAAE,CAAC,KAAK,CAAC,CAAC,GAAG,EAAE,EAAE;IACnB,sEAAsE;IACtE,sCAAsC;IACtC,OAAO,CAAC,KAAK,CAAC,SAAS,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;IAC3E,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;AAClB,CAAC,CAAC,CAAC"}