majlis 0.1.0 → 0.3.0

package/README.md

@@ -2,7 +2,7 @@
 
 Multi-agent workflow CLI for structured doubt, independent verification, and compressed knowledge.
 
-This is the CLI package. For full documentation, see the [root README](../../README.md).
+This is the CLI package. For full documentation, see the [root README](https://github.com/raihaan123/majlis#readme).
 
 ## Installation
 

package/dist/cli.js

@@ -501,9 +501,9 @@ var init_init = __esm({
       },
       models: {
         builder: "opus",
-        critic: "sonnet",
-        adversary: "sonnet",
-        verifier: "sonnet",
+        critic: "opus",
+        adversary: "opus",
+        verifier: "opus",
         reframer: "opus",
         compressor: "opus"
       }
@@ -546,7 +546,7 @@ At the end of your work, include a <!-- majlis-json --> block with your decision
 \`\`\``,
       critic: `---
 name: critic
-model: sonnet
+model: opus
 tools: [Read, Glob, Grep]
 ---
 You are the Critic. You practise constructive doubt.
@@ -578,7 +578,7 @@ Write to docs/doubts/NNN-against-experiment-NNN.md
 -->`,
       adversary: `---
 name: adversary
-model: sonnet
+model: opus
 tools: [Read, Glob, Grep]
 ---
 You are the Adversary. You do NOT review code for bugs.
@@ -606,7 +606,7 @@ Write to docs/challenges/NNN-against-experiment-NNN.md
 -->`,
       verifier: `---
 name: verifier
-model: sonnet
+model: opus
 tools: [Read, Glob, Grep, Bash]
 ---
 You are the Verifier. Perform dual verification:
@@ -687,7 +687,7 @@ You may NOT write code, make decisions, or run experiments.
 -->`,
       scout: `---
 name: scout
-model: sonnet
+model: opus
 tools: [Read, Glob, Grep, WebSearch]
 ---
 You are the Scout. You practise rihla \u2014 travel in search of knowledge.
@@ -2094,7 +2094,7 @@ var init_types2 = __esm({
 });
 
 // src/agents/parse.ts
-function extractStructuredData(role, markdown) {
+async function extractStructuredData(role, markdown) {
   const tier1 = extractMajlisJsonBlock(markdown);
   if (tier1) {
     const parsed = tryParseJson(tier1);
@@ -2109,7 +2109,7 @@ function extractStructuredData(role, markdown) {
     return tier2;
   }
   console.warn(`[majlis] Regex fallback insufficient for ${role}. Using Haiku extraction.`);
-  const tier3 = extractViaHaiku(role, markdown);
+  const tier3 = await extractViaHaiku(role, markdown);
   if (tier3) return tier3;
   console.error(
     `[majlis] FAILED to extract structured data from ${role} output. State machine will continue but data is missing. Manual review required.`
@@ -2185,22 +2185,36 @@ function extractViaPatterns(role, markdown) {
   if (doubts.length > 0) result.doubts = doubts;
   return result;
 }
-function extractViaHaiku(role, markdown) {
+async function extractViaHaiku(role, markdown) {
   try {
     const truncated = markdown.length > 8e3 ? markdown.slice(0, 8e3) + "\n[truncated]" : markdown;
     const prompt = `Extract all decisions, evidence levels, grades, doubts, and guidance from this ${role} document as JSON. Follow this schema exactly: ${EXTRACTION_SCHEMA}
 
 Document:
 ${truncated}`;
-    const result = (0, import_node_child_process3.execSync)(
-      `claude --print --model haiku --output-format json -p ${JSON.stringify(prompt)}`,
-      {
-        encoding: "utf-8",
-        timeout: 3e4,
-        stdio: ["pipe", "pipe", "pipe"]
+    const conversation = (0, import_claude_agent_sdk.query)({
+      prompt,
+      options: {
+        model: "haiku",
+        tools: [],
+        systemPrompt: "You are a JSON extraction assistant. Output only valid JSON matching the requested schema. No markdown, no explanation, just JSON.",
+        permissionMode: "bypassPermissions",
+        allowDangerouslySkipPermissions: true,
+        maxTurns: 1,
+        persistSession: false
       }
-    );
-    return tryParseJson(result.trim());
+    });
+    let resultText = "";
+    for await (const message of conversation) {
+      if (message.type === "assistant") {
+        for (const block of message.message.content) {
+          if (block.type === "text") {
+            resultText += block.text;
+          }
+        }
+      }
+    }
+    return tryParseJson(resultText.trim());
   } catch (err) {
     console.warn(`[majlis] Haiku extraction failed for ${role}: ${err instanceof Error ? err.message : String(err)}`);
     return null;
@@ -2209,12 +2223,12 @@ ${truncated}`;
 function hasData(output) {
   return !!(output.decisions && output.decisions.length > 0 || output.grades && output.grades.length > 0 || output.doubts && output.doubts.length > 0 || output.guidance);
 }
-var import_node_child_process3;
+var import_claude_agent_sdk;
 var init_parse = __esm({
   "src/agents/parse.ts"() {
     "use strict";
     init_types2();
-    import_node_child_process3 = require("child_process");
+    import_claude_agent_sdk = require("@anthropic-ai/claude-agent-sdk");
   }
 });
 
@@ -2233,7 +2247,7 @@ function loadAgentDefinition(role, projectRoot) {
   const frontmatter = frontmatterMatch[1];
   const body = frontmatterMatch[2].trim();
   const name = extractYamlField(frontmatter, "name") ?? role;
-  const model = extractYamlField(frontmatter, "model") ?? "sonnet";
+  const model = extractYamlField(frontmatter, "model") ?? "opus";
   const toolsStr = extractYamlField(frontmatter, "tools") ?? "[]";
   const tools = toolsStr.replace(/[\[\]]/g, "").split(",").map((t) => t.trim()).filter(Boolean);
   return { name, model, tools, systemPrompt: body };
@@ -2242,125 +2256,139 @@ function extractYamlField(yaml, field) {
   const match = yaml.match(new RegExp(`^${field}:\\s*(.+)$`, "m"));
   return match ? match[1].trim() : null;
 }
-function writeTempContext(context) {
-  const tmpDir = os.tmpdir();
-  const tmpFile = path7.join(tmpDir, `majlis-context-${Date.now()}.json`);
-  fs7.writeFileSync(tmpFile, JSON.stringify(context, null, 2));
-  return tmpFile;
-}
 async function spawnAgent(role, context, projectRoot) {
   const agentDef = loadAgentDefinition(role, projectRoot);
-  const contextFile = writeTempContext(context);
   const root = projectRoot ?? findProjectRoot() ?? process.cwd();
   const taskPrompt = context.taskPrompt ?? `Perform your role as ${agentDef.name}.`;
-  const prompt = `Read the context at ${contextFile}. ${taskPrompt}`;
-  const args = [
-    "--print",
-    "--output-format",
-    "stream-json",
-    "--model",
-    agentDef.model,
-    "--allowedTools",
-    ...agentDef.tools,
-    "--append-system-prompt",
-    agentDef.systemPrompt,
-    "-p",
-    prompt
-  ];
+  const contextJson = JSON.stringify(context, null, 2);
+  const prompt = `Here is your context:
+
+\`\`\`json
+${contextJson}
+\`\`\`
+
+${taskPrompt}`;
   console.log(`[majlis] Spawning ${role} agent (model: ${agentDef.model})...`);
-  const output = await runClaude(args, root);
-  const markdown = parseStreamJsonOutput(output);
+  const { text: markdown, costUsd } = await runQuery({
+    prompt,
+    model: agentDef.model,
+    tools: agentDef.tools,
+    systemPrompt: agentDef.systemPrompt,
+    cwd: root
+  });
+  console.log(`[majlis] ${role} agent complete (cost: $${costUsd.toFixed(4)})`);
   const artifactPath = writeArtifact(role, context, markdown, root);
   if (artifactPath) {
     console.log(`[majlis] ${role} artifact written to ${artifactPath}`);
   }
-  const structured = extractStructuredData(role, markdown);
-  try {
-    fs7.unlinkSync(contextFile);
-  } catch {
-  }
+  const structured = await extractStructuredData(role, markdown);
   return { output: markdown, structured };
 }
 async function spawnSynthesiser(context, projectRoot) {
-  const contextFile = writeTempContext(context);
   const root = projectRoot ?? findProjectRoot() ?? process.cwd();
-  const prompt = `Read the context at ${contextFile}. ${context.taskPrompt ?? "Synthesise the findings into actionable builder guidance."}`;
-  const args = [
-    "--print",
-    "--output-format",
-    "stream-json",
-    "--model",
-    "sonnet",
-    "--allowedTools",
-    "Read",
-    "Glob",
-    "Grep",
-    "--append-system-prompt",
-    "You are a Synthesis Agent. Your job is to take a verification report, confirmed doubts, and adversarial test results, and compress them into specific, actionable guidance for the builder's next attempt. Be concrete: which decisions failed, which assumptions broke, what constraints must the next approach satisfy. Output a 'guidance' field in JSON wrapped in a <!-- majlis-json --> block.",
-    "-p",
-    prompt
-  ];
+  const contextJson = JSON.stringify(context, null, 2);
+  const taskPrompt = context.taskPrompt ?? "Synthesise the findings into actionable builder guidance.";
+  const prompt = `Here is your context:
+
+\`\`\`json
+${contextJson}
+\`\`\`
+
+${taskPrompt}`;
+  const systemPrompt = "You are a Synthesis Agent. Your job is to take a verification report, confirmed doubts, and adversarial test results, and compress them into specific, actionable guidance for the builder's next attempt. Be concrete: which decisions failed, which assumptions broke, what constraints must the next approach satisfy. Output a 'guidance' field in JSON wrapped in a <!-- majlis-json --> block.";
   console.log(`[majlis] Spawning synthesiser micro-agent...`);
-  const output = await runClaude(args, root);
-  const markdown = parseStreamJsonOutput(output);
-  const structured = extractStructuredData("synthesiser", markdown);
-  try {
-    fs7.unlinkSync(contextFile);
-  } catch {
-  }
+  const { text: markdown, costUsd } = await runQuery({
+    prompt,
+    model: "opus",
+    tools: ["Read", "Glob", "Grep"],
+    systemPrompt,
+    cwd: root
+  });
+  console.log(`[majlis] Synthesiser complete (cost: $${costUsd.toFixed(4)})`);
+  const structured = await extractStructuredData("synthesiser", markdown);
   return { output: markdown, structured };
 }
-function runClaude(args, cwd) {
-  return new Promise((resolve2, reject) => {
-    const proc = (0, import_node_child_process4.spawn)("claude", args, {
-      cwd,
-      stdio: ["pipe", "pipe", "pipe"],
-      env: { ...process.env }
-    });
-    let stdout = "";
-    let stderr = "";
-    proc.stdout.on("data", (chunk) => {
-      stdout += chunk.toString();
-    });
-    proc.stderr.on("data", (chunk) => {
-      stderr += chunk.toString();
-    });
-    proc.on("close", (code) => {
-      if (code !== 0) {
-        reject(new Error(`claude exited with code ${code}: ${stderr}`));
-      } else {
-        resolve2(stdout);
-      }
-    });
-    proc.on("error", (err) => {
-      reject(new Error(`Failed to spawn claude: ${err.message}`));
-    });
+async function runQuery(opts) {
+  const conversation = (0, import_claude_agent_sdk2.query)({
+    prompt: opts.prompt,
+    options: {
+      model: opts.model,
+      tools: opts.tools,
+      systemPrompt: {
+        type: "preset",
+        preset: "claude_code",
+        append: opts.systemPrompt
+      },
+      cwd: opts.cwd,
+      permissionMode: "bypassPermissions",
+      allowDangerouslySkipPermissions: true,
+      maxTurns: 50,
+      persistSession: false,
+      settingSources: ["project"]
+    }
   });
-}
-function parseStreamJsonOutput(raw) {
-  const parts = [];
-  for (const line of raw.split("\n")) {
-    if (!line.trim()) continue;
-    try {
-      const event = JSON.parse(line);
-      if (event.type === "assistant" && event.message?.content) {
-        for (const block of event.message.content) {
-          if (block.type === "text") {
-            parts.push(block.text);
-          }
+  const textParts = [];
+  let costUsd = 0;
+  let turnCount = 0;
+  for await (const message of conversation) {
+    if (message.type === "assistant") {
+      turnCount++;
+      let hasText = false;
+      for (const block of message.message.content) {
+        if (block.type === "text") {
+          textParts.push(block.text);
+          hasText = true;
+        } else if (block.type === "tool_use") {
+          const toolName = block.name ?? "tool";
+          const input = block.input ?? {};
+          const detail = formatToolDetail(toolName, input);
+          process.stderr.write(`${DIM2}[majlis]   ${CYAN2}${toolName}${RESET2}${DIM2}${detail}${RESET2}
+`);
         }
-      } else if (event.type === "content_block_delta" && event.delta?.text) {
-        parts.push(event.delta.text);
-      } else if (event.type === "result" && event.result) {
-        if (typeof event.result === "string") {
-          parts.push(event.result);
+      }
+      if (hasText) {
+        const preview = textParts[textParts.length - 1].slice(0, 120).replace(/\n/g, " ").trim();
+        if (preview) {
+          process.stderr.write(`${DIM2}[majlis]   writing: ${preview}${preview.length >= 120 ? "..." : ""}${RESET2}
+`);
         }
       }
-    } catch {
-      if (line.trim()) parts.push(line);
+    } else if (message.type === "tool_progress") {
+      const elapsed = Math.round(message.elapsed_time_seconds);
+      if (elapsed > 0 && elapsed % 5 === 0) {
+        process.stderr.write(`${DIM2}[majlis]   ${message.tool_name} running (${elapsed}s)...${RESET2}
+`);
+      }
+    } else if (message.type === "result") {
+      if (message.subtype === "success") {
+        costUsd = message.total_cost_usd;
+      } else {
+        const errors = "errors" in message ? message.errors?.join("; ") ?? "Unknown error" : "Unknown error";
+        throw new Error(`Agent query failed (${message.subtype}): ${errors}`);
+      }
     }
   }
-  return parts.join("");
+  return { text: textParts.join("\n\n"), costUsd };
+}
+function formatToolDetail(toolName, input) {
+  switch (toolName) {
+    case "Read":
+      return input.file_path ? ` ${input.file_path}` : "";
+    case "Write":
+      return input.file_path ? ` \u2192 ${input.file_path}` : "";
+    case "Edit":
+      return input.file_path ? ` ${input.file_path}` : "";
+    case "Glob":
+      return input.pattern ? ` ${input.pattern}` : "";
+    case "Grep":
+      return input.pattern ? ` /${input.pattern}/` : "";
+    case "Bash":
+      return input.command ? ` $ ${input.command.slice(0, 80)}` : "";
+    case "WebSearch":
+      return input.query ? ` "${input.query}"` : "";
+    default:
+      return "";
+  }
 }
 function writeArtifact(role, context, markdown, projectRoot) {
   const dirMap = {
@@ -2391,16 +2419,18 @@ function writeArtifact(role, context, markdown, projectRoot) {
   fs7.writeFileSync(target, markdown);
   return target;
 }
-var fs7, path7, os, import_node_child_process4;
+var fs7, path7, import_claude_agent_sdk2, DIM2, RESET2, CYAN2;
 var init_spawn = __esm({
   "src/agents/spawn.ts"() {
     "use strict";
     fs7 = __toESM(require("fs"));
     path7 = __toESM(require("path"));
-    os = __toESM(require("os"));
-    import_node_child_process4 = require("child_process");
+    import_claude_agent_sdk2 = require("@anthropic-ai/claude-agent-sdk");
     init_parse();
     init_connection();
+    DIM2 = "\x1B[2m";
+    RESET2 = "\x1B[0m";
+    CYAN2 = "\x1B[36m";
   }
 });
 
@@ -2479,7 +2509,7 @@ async function resolve(db, exp, projectRoot) {
 }
 function gitMerge(branch, cwd) {
   try {
-    (0, import_node_child_process5.execSync)(`git merge ${branch} --no-ff -m "Merge experiment branch ${branch}"`, {
+    (0, import_node_child_process3.execSync)(`git merge ${branch} --no-ff -m "Merge experiment branch ${branch}"`, {
       cwd,
       encoding: "utf-8",
       stdio: ["pipe", "pipe", "pipe"]
@@ -2490,12 +2520,12 @@ function gitMerge(branch, cwd) {
 }
 function gitRevert(branch, cwd) {
   try {
-    const currentBranch = (0, import_node_child_process5.execSync)("git rev-parse --abbrev-ref HEAD", {
+    const currentBranch = (0, import_node_child_process3.execSync)("git rev-parse --abbrev-ref HEAD", {
       cwd,
       encoding: "utf-8"
     }).trim();
     if (currentBranch === branch) {
-      (0, import_node_child_process5.execSync)("git checkout main 2>/dev/null || git checkout master", {
+      (0, import_node_child_process3.execSync)("git checkout main 2>/dev/null || git checkout master", {
         cwd,
         encoding: "utf-8",
         stdio: ["pipe", "pipe", "pipe"]
@@ -2517,7 +2547,7 @@ ${gaps}
 `;
   fs8.writeFileSync(fragPath, content + entry);
 }
-var fs8, path8, import_node_child_process5;
+var fs8, path8, import_node_child_process3;
 var init_resolve = __esm({
   "src/resolve.ts"() {
     "use strict";
@@ -2526,7 +2556,7 @@ var init_resolve = __esm({
     init_types();
     init_queries();
     init_spawn();
-    import_node_child_process5 = require("child_process");
+    import_node_child_process3 = require("child_process");
     init_format();
   }
 });
@@ -3233,8 +3263,8 @@ async function main() {
       case "history":
       case "circuit-breakers":
       case "check-commit": {
-        const { query: query2 } = await Promise.resolve().then(() => (init_query(), query_exports));
-        await query2(command, rest, isJson);
+        const { query: query4 } = await Promise.resolve().then(() => (init_query(), query_exports));
+        await query4(command, rest, isJson);
         break;
       }
       case "build":

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "majlis",
-  "version": "0.1.0",
+  "version": "0.3.0",
   "description": "Multi-agent workflow CLI for structured doubt, independent verification, and compressed knowledge",
   "bin": {
     "majlis": "./dist/cli.js"
@@ -10,6 +10,7 @@
     "test": "tsx --test src/test/*.test.ts"
   },
   "dependencies": {
+    "@anthropic-ai/claude-agent-sdk": "^0.2.42",
     "better-sqlite3": "^11.0.0"
   },
   "devDependencies": {