v0.6.0 — TUI planner cockpit, planner off claude -p, web tools

CHANGELOG.md

@@ -1,5 +1,19 @@
 # Changelog
 
+## 0.6.0 — 2026-06-18
+
+### Added
+
+- **TUI Planner Cockpit (`/plan <goal>`)** — the TUI is no longer just chat + a read-only monitor. `/plan` runs the Planner to preview a decomposition, gates on human approval (`y` / `n` / `edit` to drop sub-tasks), then dispatches into the daemon loop; progress shows in the Tasks tab. Available in both the blessed TUI (`chat`) and the readline chat. (INT-1572)
+- `POST /api/plan/dispatch` — dual-path dispatch: with Linear configured it creates a parent issue + dependency-wired sub-issues and triggers a heartbeat (reusing the autonomous decomposition engine); otherwise it falls back to running each sub-task through the exec pipeline.
+- **Web tools in the agentic loop** — `web_fetch` (keyless: URL → readable text) and `web_search` (pluggable backend: Tavily/Brave when `TAVILY_KEY`/`BRAVE_SEARCH_KEY` is set, else a keyless DuckDuckGo fallback) are now exposed to every adapter (openrouter/gpt/local), restoring the web capability the `claude -p` harness used to provide. Enabled by default (`webTools` option); disabled for the SWE-bench harness to keep the benchmark honest. (INT-1573)
+
+### Changed
+
+- **Planner migrated off `claude -p`** — `runPlanner` now runs through the OpenSwarm agentic loop via the configured adapter (read-only, multi-turn) instead of shelling out to `claude -p --max-turns 1`. Completes the INT-1420 `claude -p` removal, drops the claude-binary dependency, and lets the planner read the codebase before decomposing. `PlannerResult` contract unchanged.
+- Extracted `createSubIssuesWithDependencies()` from the autonomous runner so the `/plan` endpoint and `decomposeTask` share one sub-issue/dependency engine (no logic fork).
+- Extracted `startExecTask()` in the web server so `POST /api/exec` and the `/plan` fallback share one exec-task lifecycle.
+
 ## 0.5.0 — 2026-06-11
 
 ### Added

benchmarks/sweBench.ts

@@ -174,6 +174,9 @@ async function solveOne(inst: SweInstance, model: string): Promise<{ pred: Recor
         model: diagModel,
         timeoutMs: 900_000,
         maxTurns: 50,
+        // Benchmark integrity — no web tools, or the model could just search up
+        // the instance's GitHub issue / gold patch instead of diagnosing.
+        webTools: false,
         onLog: process.env.SWE_VERBOSE ? (l) => console.log(`    [diag] ${l}`) : () => {},
       });
       // 진단자가 실수로 수정했어도 구현 단계는 깨끗한 베이스에서 시작
@@ -214,6 +217,8 @@ async function solveOne(inst: SweInstance, model: string): Promise<{ pred: Recor
       // run_tests.sh = docker cp + in-container pytest — the 30s default times
       // out into a silent no-output failure the model reads as a broken env.
       bashTimeoutMs: 240_000,
+      // Benchmark integrity — no web search of the gold patch / GitHub issue.
+      webTools: false,
       onLog: process.env.SWE_VERBOSE ? (l) => console.log(`    ${l}`) : () => {},
     });
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@intrect/openswarm",
-  "version": "0.5.0",
+  "version": "0.6.0",
   "description": "Autonomous AI agent orchestrator — Claude, GPT, Codex, and local models (Ollama/LMStudio/llama.cpp)",
   "license": "GPL-3.0",
   "type": "module",

src/adapters/agenticLoop.ts

@@ -7,6 +7,7 @@
 // ============================================
 
 import { TOOL_DEFINITIONS, executeToolCalls, createReadCache, type ToolCall, type ToolResult, type ToolDefinition } from './tools.js';
+import { WEB_TOOL_DEFINITIONS } from './webTools.js';
 import type { CliRunResult } from './types.js';
 
 // ============ 토큰 카운팅 (VEGA token_count.py 이식) ============
@@ -115,6 +116,8 @@ export interface AgenticLoopOptions {
   protectedFiles?: string[];
   /** bash tool timeout — docker-based tests need minutes (default 30s) */
   bashTimeoutMs?: number;
+  /** Expose web_fetch + web_search tools (default true). Disabled e.g. for SWE-bench integrity. */
+  webTools?: boolean;
 }
 
 /** 루프 실행 결과 */
@@ -160,6 +163,7 @@ export async function runAgenticLoop(options: AgenticLoopOptions): Promise<Agent
     nudgeMaxOnNoEdit = 0,
     protectedFiles,
     bashTimeoutMs,
+    webTools = true,
   } = options;
 
   const startTime = Date.now();
@@ -180,7 +184,9 @@ export async function runAgenticLoop(options: AgenticLoopOptions): Promise<Agent
     `or absolute paths under this root. Do NOT use "/" or a bare repo name — those are outside the project and will be rejected.\n\n`;
   messages.push({ role: 'user', content: cwdNote + prompt });
 
-  const tools = enableTools ? TOOL_DEFINITIONS : [];
+  const tools = enableTools
+    ? (webTools ? [...TOOL_DEFINITIONS, ...WEB_TOOL_DEFINITIONS] : TOOL_DEFINITIONS)
+    : [];
   const readCache = createReadCache(); // 루프 단위 read 캐시 (중복 read 차단)
   let toolCallCount = 0;
   let editToolCount = 0; // edit_file/write_file 호출 수 (no-edit 가드용)

src/adapters/gpt.ts

@@ -82,6 +82,7 @@ export class GptCliAdapter implements CliAdapter {
       nudgeMaxOnNoEdit: options.nudgeMaxOnNoEdit,
       protectedFiles: options.protectedFiles,
       bashTimeoutMs: options.bashTimeoutMs,
+      webTools: options.webTools,
     };
 
     try {

src/adapters/local.ts

@@ -160,6 +160,7 @@ export class LocalModelAdapter implements CliAdapter {
       nudgeMaxOnNoEdit: options.nudgeMaxOnNoEdit,
       protectedFiles: options.protectedFiles,
       bashTimeoutMs: options.bashTimeoutMs,
+      webTools: options.webTools,
     };
 
     try {

src/adapters/openrouter.ts

@@ -101,6 +101,7 @@ export class OpenRouterCliAdapter implements CliAdapter {
       nudgeMaxOnNoEdit: options.nudgeMaxOnNoEdit,
       protectedFiles: options.protectedFiles,
       bashTimeoutMs: options.bashTimeoutMs,
+      webTools: options.webTools,
     };
 
     try {

src/adapters/tools.ts

@@ -9,6 +9,7 @@ import fs from 'node:fs/promises';
 import { execFile } from 'node:child_process';
 import { promisify } from 'node:util';
 import path from 'node:path';
+import { webFetch, webSearch } from './webTools.js';
 
 const execFileAsync = promisify(execFile);
 
@@ -394,6 +395,16 @@ export async function executeTool(
         }
       }
 
+      case 'web_fetch': {
+        const text = await webFetch(args.url);
+        return { tool_call_id: callId, content: text, is_error: text.startsWith('Invalid URL') || text.startsWith('Fetch ') };
+      }
+
+      case 'web_search': {
+        const text = await webSearch(args.query, args.max_results);
+        return { tool_call_id: callId, content: text, is_error: text.startsWith('Search failed') || text.startsWith('Invalid query') };
+      }
+
       default:
         return { tool_call_id: callId, content: `Unknown tool: ${name}`, is_error: true };
     }

src/adapters/types.ts

@@ -60,6 +60,8 @@ export interface CliRunOptions {
   protectedFiles?: string[];
   /** bash tool timeout in ms (default 30s). Raise for docker-based tests that take minutes. */
   bashTimeoutMs?: number;
+  /** Expose web_fetch + web_search tools (default true). Set false for SWE-bench integrity. */
+  webTools?: boolean;
 }
 
 /**

src/adapters/webTools.test.ts

@@ -0,0 +1,93 @@
+import { afterEach, describe, expect, it, vi } from 'vitest';
+import { webFetch, webSearch, searchBackend, WEB_TOOL_DEFINITIONS } from './webTools.js';
+
+afterEach(() => {
+  vi.unstubAllGlobals();
+  vi.unstubAllEnvs();
+  vi.clearAllMocks();
+});
+
+describe('WEB_TOOL_DEFINITIONS', () => {
+  it('exposes exactly web_fetch and web_search', () => {
+    expect(WEB_TOOL_DEFINITIONS.map((t) => t.function.name)).toEqual(['web_fetch', 'web_search']);
+  });
+});
+
+describe('webFetch', () => {
+  it('strips HTML to readable text', async () => {
+    vi.stubGlobal('fetch', vi.fn(async () =>
+      new Response(
+        '<html><body><h1>Hi</h1><script>bad()</script><p>world &amp; co</p></body></html>',
+        { status: 200, headers: { 'content-type': 'text/html' } },
+      ),
+    ));
+    const out = await webFetch('https://example.com');
+    expect(out).toContain('Hi');
+    expect(out).toContain('world & co');
+    expect(out).not.toContain('<');
+    expect(out).not.toContain('bad()');
+  });
+
+  it('rejects a non-http URL without fetching', async () => {
+    const f = vi.fn();
+    vi.stubGlobal('fetch', f);
+    const out = await webFetch('ftp://x');
+    expect(out).toContain('Invalid URL');
+    expect(f).not.toHaveBeenCalled();
+  });
+
+  it('reports an HTTP error rather than throwing', async () => {
+    vi.stubGlobal('fetch', vi.fn(async () => new Response('nope', { status: 404, statusText: 'Not Found' })));
+    const out = await webFetch('https://example.com/x');
+    expect(out).toContain('404');
+  });
+});
+
+describe('webSearch — backend selection', () => {
+  it('defaults to duckduckgo with no keys', () => {
+    expect(searchBackend()).toBe('duckduckgo');
+  });
+
+  it('prefers Tavily when TAVILY_KEY is set', async () => {
+    vi.stubEnv('TAVILY_KEY', 'tk');
+    expect(searchBackend()).toBe('tavily');
+    const f = vi.fn(async () =>
+      new Response(JSON.stringify({ results: [{ title: 'T', url: 'https://t', content: 'snip' }] }), { status: 200 }),
+    );
+    vi.stubGlobal('fetch', f);
+    const out = await webSearch('q', 3);
+    expect(String(f.mock.calls[0][0])).toContain('api.tavily.com');
+    expect(out).toContain('T');
+    expect(out).toContain('https://t');
+  });
+
+  it('uses Brave when BRAVE_SEARCH_KEY is set', async () => {
+    vi.stubEnv('BRAVE_SEARCH_KEY', 'bk');
+    expect(searchBackend()).toBe('brave');
+    const f = vi.fn(async () =>
+      new Response(JSON.stringify({ web: { results: [{ title: 'B', url: 'https://b', description: 'd' }] } }), { status: 200 }),
+    );
+    vi.stubGlobal('fetch', f);
+    const out = await webSearch('q');
+    expect(String(f.mock.calls[0][0])).toContain('brave.com');
+    expect(out).toContain('B');
+  });
+
+  it('parses keyless DuckDuckGo HTML results', async () => {
+    const html =
+      '<a class="result__a" href="//duckduckgo.com/l/?uddg=https%3A%2F%2Fex.com%2Fa">Result A</a>' +
+      '<a class="result__snippet">snippet a</a>';
+    vi.stubGlobal('fetch', vi.fn(async () => new Response(html, { status: 200 })));
+    const out = await webSearch('q', 5);
+    expect(out).toContain('Result A');
+    expect(out).toContain('https://ex.com/a');
+    expect(out).toContain('snippet a');
+  });
+
+  it('returns an error string (does not throw) on backend failure', async () => {
+    vi.stubGlobal('fetch', vi.fn(async () => { throw new Error('net down'); }));
+    const out = await webSearch('q');
+    expect(out).toContain('Search failed');
+    expect(out).toContain('TAVILY_KEY'); // keyless hint
+  });
+});