majlis 0.4.0 → 0.4.2

package/dist/cli.js

@@ -522,24 +522,35 @@ Before building:
 3. Check docs/classification/ for problem taxonomy
 4. Check docs/experiments/ for prior work
 
-## Scope Constraint (CRITICAL)
+Read as much code as you need to understand the problem. Reading is free \u2014 spend
+as many turns as necessary on Read, Grep, and Glob to build full context before
+you touch anything.
+
+## The Rule: ONE Change, Then Document
+
+You make ONE code change per cycle. Not two, not "one more quick fix." ONE.
+
+The sequence:
+1. **Read and understand** \u2014 read synthesis, dead-ends, source code. Take your time.
+2. **Write the experiment doc FIRST** \u2014 before coding, fill in the Approach section
+   with what you plan to do and why. This ensures there is always a record.
+3. **Implement ONE focused change** \u2014 a single coherent edit to the codebase.
+4. **Run the benchmark ONCE** \u2014 observe the result.
+5. **Update the experiment doc** \u2014 fill in Results and Metrics with what happened.
+6. **Output the majlis-json block** \u2014 your structured decisions.
+7. **STOP.**
 
-You get ONE attempt per cycle. Your job is:
-1. Read and diagnose \u2014 understand the problem thoroughly
-2. Form ONE hypothesis about what to fix
-3. Implement ONE focused change (not a multi-step debug session)
-4. Run the benchmark ONCE to see the result
-5. Update the experiment doc in docs/experiments/ \u2014 fill in Approach, Results, and Metrics sections. This is NOT optional.
-6. Output the structured majlis-json block with your decisions
-7. STOP
+If your change doesn't work, document what happened and STOP. Do NOT try to fix it.
+Do NOT iterate. Do NOT "try one more thing." The adversary, critic, and verifier
+exist to diagnose what went wrong. The cycle comes back to you with their insights.
 
-Do NOT iterate. Do NOT try multiple approaches. Do NOT debug your own fix.
-If your change doesn't work, document why and let the cycle continue \u2014
-the adversary, critic, and verifier will help diagnose what went wrong.
-The cycle will come back to you with their insights.
+If you find yourself wanting to debug your own fix, that's the signal to stop
+and write up what you learned.
 
-If you find yourself wanting to "try one more thing," that's the signal to stop
-and write up what you learned. The other agents exist precisely for this reason.
+## Off-limits (DO NOT modify)
+- \`fixtures/\` \u2014 test data, ground truth, STL files. Read-only.
+- \`scripts/benchmark.py\` \u2014 the measurement tool. Never change how you're measured.
+- \`.majlis/\` \u2014 framework config. Not your concern.
 
 ## During building:
 - Tag EVERY decision: proof / test / strong-consensus / consensus / analogy / judgment
@@ -2293,13 +2304,15 @@ ${contextJson}
 \`\`\`
 
 ${taskPrompt}`;
-  console.log(`[majlis] Spawning ${role} agent (model: ${agentDef.model})...`);
+  const turns = ROLE_MAX_TURNS[role] ?? 15;
+  console.log(`[majlis] Spawning ${role} agent (model: ${agentDef.model}, maxTurns: ${turns})...`);
   const { text: markdown, costUsd } = await runQuery({
     prompt,
     model: agentDef.model,
     tools: agentDef.tools,
     systemPrompt: agentDef.systemPrompt,
-    cwd: root
+    cwd: root,
+    maxTurns: turns
   });
   console.log(`[majlis] ${role} agent complete (cost: $${costUsd.toFixed(4)})`);
   const artifactPath = writeArtifact(role, context, markdown, root);
@@ -2320,14 +2333,15 @@ ${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.";
+  const systemPrompt = 'You are a Synthesis Agent. Be concrete: which decisions failed, which assumptions broke, what constraints must the next approach satisfy. CRITICAL: Your LAST line of output MUST be a <!-- majlis-json --> block. The framework parses this programmatically \u2014 if you omit it, the pipeline breaks. Format: <!-- majlis-json {"guidance": "your guidance here"} -->';
   console.log(`[majlis] Spawning synthesiser micro-agent...`);
   const { text: markdown, costUsd } = await runQuery({
     prompt,
     model: "opus",
     tools: ["Read", "Glob", "Grep"],
     systemPrompt,
-    cwd: root
+    cwd: root,
+    maxTurns: 5
   });
   console.log(`[majlis] Synthesiser complete (cost: $${costUsd.toFixed(4)})`);
   const structured = await extractStructuredData("synthesiser", markdown);
@@ -2347,7 +2361,7 @@ async function runQuery(opts) {
       cwd: opts.cwd,
       permissionMode: "bypassPermissions",
       allowDangerouslySkipPermissions: true,
-      maxTurns: 30,
+      maxTurns: opts.maxTurns ?? 15,
       persistSession: false,
       settingSources: ["project"]
     }
@@ -2447,7 +2461,7 @@ function writeArtifact(role, context, markdown, projectRoot) {
   fs7.writeFileSync(target, markdown);
   return target;
 }
-var fs7, path7, import_claude_agent_sdk2, DIM2, RESET2, CYAN2;
+var fs7, path7, import_claude_agent_sdk2, ROLE_MAX_TURNS, DIM2, RESET2, CYAN2;
 var init_spawn = __esm({
   "src/agents/spawn.ts"() {
     "use strict";
@@ -2456,6 +2470,15 @@ var init_spawn = __esm({
     import_claude_agent_sdk2 = require("@anthropic-ai/claude-agent-sdk");
     init_parse();
     init_connection();
+    ROLE_MAX_TURNS = {
+      builder: 50,
+      critic: 12,
+      adversary: 12,
+      verifier: 15,
+      compressor: 15,
+      reframer: 12,
+      scout: 12
+    };
     DIM2 = "\x1B[2m";
     RESET2 = "\x1B[0m";
     CYAN2 = "\x1B[36m";
@@ -3122,6 +3145,8 @@ async function executeStep(step, exp, root) {
       break;
     case "compressed" /* COMPRESSED */:
       await cycle("compress", []);
+      updateExperimentStatus(getDb(root), exp.id, "compressed");
+      info(`Experiment ${exp.slug} compressed.`);
       break;
     case "reframed" /* REFRAMED */:
       updateExperimentStatus(getDb(root), exp.id, "reframed");
@@ -3268,14 +3293,20 @@ ${deadEnds.map((d) => `- ${d.approach}: ${d.why_failed} [constraint: ${d.structu
 
 ## Your Task
 1. Assess: based on the metrics and synthesis, has the goal been met? Be specific.
-2. If YES \u2014 output: <!-- majlis-json {"goal_met": true, "hypothesis": null} -->
+2. If YES \u2014 output the JSON block below with goal_met: true.
 3. If NO \u2014 propose the SINGLE most promising next experiment hypothesis.
-   - It must NOT repeat a dead-ended approach
+   - It must NOT repeat a dead-ended approach (check the dead-end registry!)
    - It should attack the weakest point revealed by synthesis/fragility
-   - It should be specific and actionable (not vague)
-   - Output: <!-- majlis-json {"goal_met": false, "hypothesis": "your hypothesis here"} -->
+   - It must be specific and actionable \u2014 name the function or mechanism to change
+   - Do NOT reference specific line numbers \u2014 they shift between experiments
+   - The hypothesis should be a single sentence describing what to do, e.g.:
+     "Activate addSeamEdges() in the runEdgeFirst pipeline for full-revolution cylinder faces"
+
+CRITICAL: Your LAST line of output MUST be EXACTLY this format (on its own line, nothing after it):
+<!-- majlis-json {"goal_met": false, "hypothesis": "your single-sentence hypothesis here"} -->
 
-IMPORTANT: You MUST output the <!-- majlis-json --> block. This is how the framework reads your decision.`
+If the goal is met:
+<!-- majlis-json {"goal_met": true, "hypothesis": null} -->`
   }, root);
   const structured = result.structured;
   if (structured?.goal_met === true) {
@@ -3284,9 +3315,27 @@ IMPORTANT: You MUST output the <!-- majlis-json --> block. This is how the frame
   if (structured?.hypothesis) {
     return structured.hypothesis;
   }
-  const match = result.output.match(/hypothesis["\s:]+([^"}\n]+)/i);
-  if (match) return match[1].trim();
-  warn("Planner did not return a structured hypothesis. Using goal as fallback.");
+  const jsonMatch = result.output.match(/"hypothesis"\s*:\s*"([^"]+)"/);
+  if (jsonMatch && jsonMatch[1].length > 10) return jsonMatch[1].trim();
+  const blockMatch = result.output.match(/<!--\s*majlis-json\s*(\{[\s\S]*?\})\s*-->/);
+  if (blockMatch) {
+    try {
+      const parsed = JSON.parse(blockMatch[1]);
+      if (parsed.goal_met === true) return null;
+      if (parsed.hypothesis) return parsed.hypothesis;
+    } catch {
+    }
+  }
+  warn("Planner did not return structured output. Retrying with focused prompt...");
+  const retry = await spawnSynthesiser({
+    taskPrompt: `Based on this analysis, output ONLY a single-line JSON block:
+
+${result.output.slice(-2e3)}
+
+<!-- majlis-json {"goal_met": false, "hypothesis": "your hypothesis"} -->`
+  }, root);
+  if (retry.structured?.hypothesis) return retry.structured.hypothesis;
+  warn("Could not extract hypothesis. Using goal as fallback.");
   return goal;
 }
 function createNewExperiment(db, root, hypothesis) {
@@ -3312,6 +3361,8 @@ function createNewExperiment(db, root, hypothesis) {
     warn(`Could not create branch ${branch} \u2014 continuing without git branch.`);
   }
   const exp = createExperiment(db, finalSlug, branch, hypothesis, null, null);
+  updateExperimentStatus(db, exp.id, "reframed");
+  exp.status = "reframed";
   const docsDir = path13.join(root, "docs", "experiments");
   const templatePath = path13.join(docsDir, "_TEMPLATE.md");
   if (fs13.existsSync(templatePath)) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "majlis",
-  "version": "0.4.0",
+  "version": "0.4.2",
   "description": "Multi-agent workflow CLI for structured doubt, independent verification, and compressed knowledge",
   "bin": {
     "majlis": "./dist/cli.js"