majlis 0.4.1 → 0.4.3

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.
 
-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
+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.**
 
-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 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.
 
-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.
+If you find yourself wanting to debug your own fix, that's the signal to stop
+and write up what you learned.
+
+## 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
@@ -2294,19 +2305,20 @@ ${contextJson}
 
 ${taskPrompt}`;
   const turns = ROLE_MAX_TURNS[role] ?? 15;
-  console.log(`[majlis] Spawning ${role} agent (model: ${agentDef.model}, maxTurns: ${turns})...`);
+  console.log(`[${role}] Spawning (model: ${agentDef.model}, maxTurns: ${turns})...`);
   const { text: markdown, costUsd } = await runQuery({
     prompt,
     model: agentDef.model,
     tools: agentDef.tools,
     systemPrompt: agentDef.systemPrompt,
     cwd: root,
-    maxTurns: turns
+    maxTurns: turns,
+    label: role
   });
-  console.log(`[majlis] ${role} agent complete (cost: $${costUsd.toFixed(4)})`);
+  console.log(`[${role}] Complete (cost: $${costUsd.toFixed(4)})`);
   const artifactPath = writeArtifact(role, context, markdown, root);
   if (artifactPath) {
-    console.log(`[majlis] ${role} artifact written to ${artifactPath}`);
+    console.log(`[${role}] Artifact written to ${artifactPath}`);
   }
   const structured = await extractStructuredData(role, markdown);
   return { output: markdown, structured };
@@ -2323,20 +2335,21 @@ ${contextJson}
 
 ${taskPrompt}`;
   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...`);
+  console.log(`[synthesiser] Spawning (maxTurns: 5)...`);
   const { text: markdown, costUsd } = await runQuery({
     prompt,
     model: "opus",
     tools: ["Read", "Glob", "Grep"],
     systemPrompt,
     cwd: root,
-    maxTurns: 5
+    maxTurns: 5,
+    label: "synthesiser"
   });
-  console.log(`[majlis] Synthesiser complete (cost: $${costUsd.toFixed(4)})`);
-  const structured = await extractStructuredData("synthesiser", markdown);
-  return { output: markdown, structured };
+  console.log(`[synthesiser] Complete (cost: $${costUsd.toFixed(4)})`);
+  return { output: markdown, structured: { guidance: markdown } };
 }
 async function runQuery(opts) {
+  const tag = opts.label ?? "majlis";
   const conversation = (0, import_claude_agent_sdk2.query)({
     prompt: opts.prompt,
     options: {
@@ -2370,21 +2383,21 @@ async function runQuery(opts) {
           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}
+          process.stderr.write(`${DIM2}[${tag}]   ${CYAN2}${toolName}${RESET2}${DIM2}${detail}${RESET2}
 `);
         }
       }
       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 ? "..." : ""}${RESET2}
+          process.stderr.write(`${DIM2}[${tag}]   writing: ${preview}${preview.length >= 120 ? "..." : ""}${RESET2}
 `);
         }
       }
     } else if (message.type === "tool_progress") {
       const elapsed = Math.round(message.elapsed_time_seconds);
       if (elapsed > 0 && elapsed % 5 === 0) {
-        process.stderr.write(`${DIM2}[majlis]   ${message.tool_name} running (${elapsed}s)...${RESET2}
+        process.stderr.write(`${DIM2}[${tag}]   ${message.tool_name} running (${elapsed}s)...${RESET2}
 `);
       }
     } else if (message.type === "result") {
@@ -2392,7 +2405,7 @@ async function runQuery(opts) {
         costUsd = message.total_cost_usd;
       } else if (message.subtype === "error_max_turns") {
         costUsd = "total_cost_usd" in message ? message.total_cost_usd : 0;
-        console.warn(`[majlis] Agent hit max turns (${turnCount}). Returning partial output.`);
+        console.warn(`[${tag}] Hit max turns (${turnCount}). Returning partial output.`);
       } else {
         const errors = "errors" in message ? message.errors?.join("; ") ?? "Unknown error" : "Unknown error";
         throw new Error(`Agent query failed (${message.subtype}): ${errors}`);
@@ -2444,7 +2457,7 @@ function writeArtifact(role, context, markdown, projectRoot) {
   }
   const expSlug = context.experiment?.slug ?? "general";
   const existing = fs7.readdirSync(fullDir).filter((f) => f.endsWith(".md") && !f.startsWith("_"));
-  const nextNum = String(existing.length + 1).padStart(3, "0");
+  const nextNum = String(context.experiment?.id ?? existing.length + 1).padStart(3, "0");
   const filename = role === "builder" ? `${nextNum}-${expSlug}.md` : `${nextNum}-${role}-${expSlug}.md`;
   const target = path7.join(fullDir, filename);
   fs7.writeFileSync(target, markdown);
@@ -2460,13 +2473,13 @@ var init_spawn = __esm({
     init_parse();
     init_connection();
     ROLE_MAX_TURNS = {
-      builder: 15,
-      critic: 12,
-      adversary: 12,
-      verifier: 15,
-      compressor: 15,
-      reframer: 12,
-      scout: 12
+      builder: 50,
+      critic: 30,
+      adversary: 30,
+      verifier: 50,
+      compressor: 30,
+      reframer: 20,
+      scout: 20
     };
     DIM2 = "\x1B[2m";
     RESET2 = "\x1B[0m";
@@ -2482,9 +2495,19 @@ function worstGrade(grades) {
   return "sound";
 }
 async function resolve(db, exp, projectRoot) {
-  const grades = getVerificationsByExperiment(db, exp.id);
+  let grades = getVerificationsByExperiment(db, exp.id);
   if (grades.length === 0) {
-    throw new Error(`No verifications found for experiment ${exp.slug}. Run verify first.`);
+    warn(`No verification records for ${exp.slug}. Defaulting to weak.`);
+    insertVerification(
+      db,
+      exp.id,
+      "auto-default",
+      "weak",
+      null,
+      null,
+      "No structured verification output. Auto-defaulted to weak."
+    );
+    grades = getVerificationsByExperiment(db, exp.id);
   }
   const overallGrade = worstGrade(grades);
   switch (overallGrade) {
@@ -3230,7 +3253,16 @@ async function run(args) {
       continue;
     }
     info(`[Step ${stepCount}] ${exp.slug}: ${exp.status}`);
-    await next([exp.slug], false);
+    try {
+      await next([exp.slug], false);
+    } catch (err) {
+      const message = err instanceof Error ? err.message : String(err);
+      warn(`Step failed for ${exp.slug}: ${message}`);
+      try {
+        updateExperimentStatus(db, exp.id, "dead_end");
+      } catch {
+      }
+    }
   }
   if (stepCount >= MAX_STEPS) {
     warn(`Reached max steps (${MAX_STEPS}). Stopping autonomous mode.`);
@@ -3286,7 +3318,8 @@ ${deadEnds.map((d) => `- ${d.approach}: ${d.why_failed} [constraint: ${d.structu
 3. If NO \u2014 propose the SINGLE most promising next experiment hypothesis.
    - It must NOT repeat a dead-ended approach (check the dead-end registry!)
    - It should attack the weakest point revealed by synthesis/fragility
-   - It must be specific and actionable \u2014 name the exact code/function/mechanism to change
+   - 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"
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "majlis",
-  "version": "0.4.1",
+  "version": "0.4.3",
   "description": "Multi-agent workflow CLI for structured doubt, independent verification, and compressed knowledge",
   "bin": {
     "majlis": "./dist/cli.js"