majlis 0.4.3 → 0.4.5

package/dist/cli.js

@@ -2054,7 +2054,8 @@ var init_types = __esm({
     TRANSITIONS = {
       ["classified" /* CLASSIFIED */]: ["reframed" /* REFRAMED */, "building" /* BUILDING */],
       ["reframed" /* REFRAMED */]: ["building" /* BUILDING */],
-      ["building" /* BUILDING */]: ["built" /* BUILT */],
+      ["building" /* BUILDING */]: ["built" /* BUILT */, "building" /* BUILDING */],
+      // self-loop for retry after truncation
       ["built" /* BUILT */]: ["challenged" /* CHALLENGED */, "doubted" /* DOUBTED */],
       ["challenged" /* CHALLENGED */]: ["doubted" /* DOUBTED */, "verifying" /* VERIFYING */],
       ["doubted" /* DOUBTED */]: ["challenged" /* CHALLENGED */, "scouted" /* SCOUTED */, "verifying" /* VERIFYING */],
@@ -2306,7 +2307,7 @@ ${contextJson}
 ${taskPrompt}`;
   const turns = ROLE_MAX_TURNS[role] ?? 15;
   console.log(`[${role}] Spawning (model: ${agentDef.model}, maxTurns: ${turns})...`);
-  const { text: markdown, costUsd } = await runQuery({
+  const { text: markdown, costUsd, truncated } = await runQuery({
     prompt,
     model: agentDef.model,
     tools: agentDef.tools,
@@ -2315,13 +2316,13 @@ ${taskPrompt}`;
     maxTurns: turns,
     label: role
   });
-  console.log(`[${role}] Complete (cost: $${costUsd.toFixed(4)})`);
+  console.log(`[${role}] Complete (cost: $${costUsd.toFixed(4)}${truncated ? ", TRUNCATED" : ""})`);
   const artifactPath = writeArtifact(role, context, markdown, root);
   if (artifactPath) {
     console.log(`[${role}] Artifact written to ${artifactPath}`);
   }
   const structured = await extractStructuredData(role, markdown);
-  return { output: markdown, structured };
+  return { output: markdown, structured, truncated };
 }
 async function spawnSynthesiser(context, projectRoot) {
   const root = projectRoot ?? findProjectRoot() ?? process.cwd();
@@ -2336,7 +2337,7 @@ ${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(`[synthesiser] Spawning (maxTurns: 5)...`);
-  const { text: markdown, costUsd } = await runQuery({
+  const { text: markdown, costUsd, truncated } = await runQuery({
     prompt,
     model: "opus",
     tools: ["Read", "Glob", "Grep"],
@@ -2346,9 +2347,58 @@ ${taskPrompt}`;
     label: "synthesiser"
   });
   console.log(`[synthesiser] Complete (cost: $${costUsd.toFixed(4)})`);
-  return { output: markdown, structured: { guidance: markdown } };
+  return { output: markdown, structured: { guidance: markdown }, truncated };
+}
+async function spawnRecovery(role, partialOutput, context, projectRoot) {
+  const root = projectRoot ?? findProjectRoot() ?? process.cwd();
+  const expSlug = context.experiment?.slug ?? "unknown";
+  console.log(`[recovery] Cleaning up after truncated ${role} for ${expSlug}...`);
+  const expDocPath = path7.join(
+    root,
+    "docs",
+    "experiments",
+    `${String(context.experiment?.id ?? 0).padStart(3, "0")}-${expSlug}.md`
+  );
+  const templatePath = path7.join(root, "docs", "experiments", "_TEMPLATE.md");
+  const template = fs7.existsSync(templatePath) ? fs7.readFileSync(templatePath, "utf-8") : "";
+  const currentDoc = fs7.existsSync(expDocPath) ? fs7.readFileSync(expDocPath, "utf-8") : "";
+  const prompt = `The ${role} agent was truncated (hit max turns) while working on experiment "${expSlug}".
+
+Here is the partial agent output (reasoning + tool calls):
+<partial_output>
+${partialOutput.slice(-3e3)}
+</partial_output>
+
+Here is the current experiment doc:
+<current_doc>
+${currentDoc}
+</current_doc>
+
+Here is the template that the experiment doc should follow:
+<template>
+${template}
+</template>
+
+Your job: Write a CLEAN experiment doc to ${expDocPath} using the Write tool.
+- Keep any valid content from the current doc
+- Fill in what you can infer from the partial output
+- Mark incomplete sections with "[TRUNCATED \u2014 ${role} did not finish]"
+- The doc MUST have the <!-- majlis-json --> block, even if decisions are empty
+- Do NOT include agent reasoning or thinking \u2014 only structured experiment content
+- Be concise. This is cleanup, not new work.`;
+  const { text: _markdown } = await runQuery({
+    prompt,
+    model: "haiku",
+    tools: ["Read", "Write"],
+    systemPrompt: `You are a Recovery Agent. You clean up experiment docs after truncated agent runs. Write clean, structured docs. Never include agent reasoning or monologue.`,
+    cwd: root,
+    maxTurns: 5,
+    label: "recovery"
+  });
+  console.log(`[recovery] Cleanup complete for ${expSlug}.`);
 }
 async function runQuery(opts) {
+  let truncated = false;
   const tag = opts.label ?? "majlis";
   const conversation = (0, import_claude_agent_sdk2.query)({
     prompt: opts.prompt,
@@ -2404,6 +2454,7 @@ async function runQuery(opts) {
       if (message.subtype === "success") {
         costUsd = message.total_cost_usd;
       } else if (message.subtype === "error_max_turns") {
+        truncated = true;
         costUsd = "total_cost_usd" in message ? message.total_cost_usd : 0;
         console.warn(`[${tag}] Hit max turns (${turnCount}). Returning partial output.`);
       } else {
@@ -2412,7 +2463,7 @@ async function runQuery(opts) {
       }
     }
   }
-  return { text: textParts.join("\n\n"), costUsd };
+  return { text: textParts.join("\n\n"), costUsd, truncated };
 }
 function formatToolDetail(toolName, input) {
   switch (toolName) {
@@ -2446,19 +2497,14 @@ function writeArtifact(role, context, markdown, projectRoot) {
   };
   const dir = dirMap[role];
   if (!dir) return null;
+  if (role === "builder" || role === "compressor") return null;
   const fullDir = path7.join(projectRoot, dir);
   if (!fs7.existsSync(fullDir)) {
     fs7.mkdirSync(fullDir, { recursive: true });
   }
-  if (role === "compressor") {
-    const target2 = path7.join(fullDir, "current.md");
-    fs7.writeFileSync(target2, markdown);
-    return target2;
-  }
   const expSlug = context.experiment?.slug ?? "general";
-  const existing = fs7.readdirSync(fullDir).filter((f) => f.endsWith(".md") && !f.startsWith("_"));
-  const nextNum = String(context.experiment?.id ?? existing.length + 1).padStart(3, "0");
-  const filename = role === "builder" ? `${nextNum}-${expSlug}.md` : `${nextNum}-${role}-${expSlug}.md`;
+  const nextNum = String(context.experiment?.id ?? 1).padStart(3, "0");
+  const filename = `${nextNum}-${role}-${expSlug}.md`;
   const target = path7.join(fullDir, filename);
   fs7.writeFileSync(target, markdown);
   return target;
@@ -2588,6 +2634,10 @@ function gitRevert(branch, cwd) {
       encoding: "utf-8"
     }).trim();
     if (currentBranch === branch) {
+      try {
+        (0, import_node_child_process3.execSync)("git checkout -- .", { cwd, encoding: "utf-8", stdio: ["pipe", "pipe", "pipe"] });
+      } catch {
+      }
       (0, import_node_child_process3.execSync)("git checkout main 2>/dev/null || git checkout master", {
         cwd,
         encoding: "utf-8",
@@ -2690,8 +2740,17 @@ ${builderGuidance}
 Build the experiment: ${exp.hypothesis}` : `Build the experiment: ${exp.hypothesis}`
   }, root);
   ingestStructuredOutput(db, exp.id, result.structured);
-  updateExperimentStatus(db, exp.id, "built");
-  success(`Build complete for ${exp.slug}. Run \`majlis doubt\` or \`majlis challenge\` next.`);
+  if (result.truncated && !result.structured) {
+    warn(`Builder was truncated (hit max turns) without producing structured output.`);
+    await spawnRecovery("builder", result.output, {
+      experiment: { id: exp.id, slug: exp.slug, hypothesis: exp.hypothesis, status: "building", sub_type: exp.sub_type, builder_guidance: null }
+    }, root);
+    warn(`Experiment stays at 'building'. Run \`majlis build\` to retry or \`majlis revert\` to abandon.`);
+  } else {
+    gitCommitBuild(exp, root);
+    updateExperimentStatus(db, exp.id, "built");
+    success(`Build complete for ${exp.slug}. Run \`majlis doubt\` or \`majlis challenge\` next.`);
+  }
 }
 async function doChallenge(db, exp, root) {
   transition(exp.status, "challenged" /* CHALLENGED */);
@@ -2707,8 +2766,12 @@ async function doChallenge(db, exp, root) {
     taskPrompt: `Construct adversarial test cases for experiment ${exp.slug}: ${exp.hypothesis}`
   }, root);
   ingestStructuredOutput(db, exp.id, result.structured);
-  updateExperimentStatus(db, exp.id, "challenged");
-  success(`Challenge complete for ${exp.slug}. Run \`majlis doubt\` or \`majlis verify\` next.`);
+  if (result.truncated && !result.structured) {
+    warn(`Adversary was truncated without structured output. Experiment stays at current status.`);
+  } else {
+    updateExperimentStatus(db, exp.id, "challenged");
+    success(`Challenge complete for ${exp.slug}. Run \`majlis doubt\` or \`majlis verify\` next.`);
+  }
 }
 async function doDoubt(db, exp, root) {
   transition(exp.status, "doubted" /* DOUBTED */);
@@ -2725,8 +2788,12 @@ async function doDoubt(db, exp, root) {
     taskPrompt: `Doubt the work in experiment ${exp.slug}: ${exp.hypothesis}. Produce a doubt document with evidence for each doubt.`
   }, root);
   ingestStructuredOutput(db, exp.id, result.structured);
-  updateExperimentStatus(db, exp.id, "doubted");
-  success(`Doubt pass complete for ${exp.slug}. Run \`majlis challenge\` or \`majlis verify\` next.`);
+  if (result.truncated && !result.structured) {
+    warn(`Critic was truncated without structured output. Experiment stays at current status.`);
+  } else {
+    updateExperimentStatus(db, exp.id, "doubted");
+    success(`Doubt pass complete for ${exp.slug}. Run \`majlis challenge\` or \`majlis verify\` next.`);
+  }
 }
 async function doScout(db, exp, root) {
   transition(exp.status, "scouted" /* SCOUTED */);
@@ -2774,6 +2841,10 @@ async function doVerify(db, exp, root) {
     taskPrompt: `Verify experiment ${exp.slug}: ${exp.hypothesis}. Check provenance and content. Test the ${doubts.length} doubt(s) and any adversarial challenges.`
   }, root);
   ingestStructuredOutput(db, exp.id, result.structured);
+  if (result.truncated && !result.structured) {
+    warn(`Verifier was truncated without structured output. Experiment stays at 'verifying'.`);
+    return;
+  }
   if (result.structured?.doubt_resolutions) {
     for (const dr of result.structured.doubt_resolutions) {
       if (dr.doubt_id && dr.resolution) {
@@ -2795,6 +2866,23 @@ async function doCompress(db, root) {
   recordCompression(db, sessionCount, sizeBefore, sizeAfter);
   success(`Compression complete. Synthesis: ${sizeBefore}B \u2192 ${sizeAfter}B`);
 }
+function gitCommitBuild(exp, cwd) {
+  try {
+    (0, import_node_child_process4.execSync)('git add -A -- ":!.majlis/"', { cwd, encoding: "utf-8", stdio: ["pipe", "pipe", "pipe"] });
+    const diff = (0, import_node_child_process4.execSync)("git diff --cached --stat", { cwd, encoding: "utf-8", stdio: ["pipe", "pipe", "pipe"] }).trim();
+    if (!diff) {
+      info("No code changes to commit.");
+      return;
+    }
+    const msg = `EXP-${String(exp.id).padStart(3, "0")}: ${exp.slug}
+
+${exp.hypothesis ?? ""}`;
+    (0, import_node_child_process4.execSync)(`git commit -m ${JSON.stringify(msg)}`, { cwd, encoding: "utf-8", stdio: ["pipe", "pipe", "pipe"] });
+    info(`Committed builder changes on ${exp.branch}.`);
+  } catch {
+    warn("Could not auto-commit builder changes \u2014 commit manually before resolving.");
+  }
+}
 function resolveExperimentArg(db, args) {
   const slugArg = args.filter((a) => !a.startsWith("--"))[0];
   let exp;
@@ -2849,12 +2937,13 @@ function ingestStructuredOutput(db, experimentId, structured) {
     info(`Ingested ${structured.challenges.length} challenge(s)`);
   }
 }
-var fs9, path9;
+var fs9, path9, import_node_child_process4;
 var init_cycle = __esm({
   "src/commands/cycle.ts"() {
     "use strict";
     fs9 = __toESM(require("fs"));
     path9 = __toESM(require("path"));
+    import_node_child_process4 = require("child_process");
     init_connection();
     init_queries();
     init_machine();
@@ -3281,7 +3370,7 @@ async function deriveNextHypothesis(goal, root, db) {
   let metricsOutput = "";
   if (config.metrics?.command) {
     try {
-      metricsOutput = (0, import_node_child_process4.execSync)(config.metrics.command, {
+      metricsOutput = (0, import_node_child_process5.execSync)(config.metrics.command, {
         cwd: root,
         encoding: "utf-8",
         timeout: 6e4,
@@ -3372,7 +3461,7 @@ function createNewExperiment(db, root, hypothesis) {
   const paddedNum = String(num).padStart(3, "0");
   const branch = `exp/${paddedNum}-${finalSlug}`;
   try {
-    (0, import_node_child_process4.execSync)(`git checkout -b ${branch}`, {
+    (0, import_node_child_process5.execSync)(`git checkout -b ${branch}`, {
       cwd: root,
       encoding: "utf-8",
       stdio: ["pipe", "pipe", "pipe"]
@@ -3421,13 +3510,13 @@ function loadConfig7(projectRoot) {
   }
   return JSON.parse(fs13.readFileSync(configPath, "utf-8"));
 }
-var fs13, path13, import_node_child_process4;
+var fs13, path13, import_node_child_process5;
 var init_run = __esm({
   "src/commands/run.ts"() {
     "use strict";
     fs13 = __toESM(require("fs"));
     path13 = __toESM(require("path"));
-    import_node_child_process4 = require("child_process");
+    import_node_child_process5 = require("child_process");
     init_connection();
     init_queries();
     init_machine();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "majlis",
-  "version": "0.4.3",
+  "version": "0.4.5",
   "description": "Multi-agent workflow CLI for structured doubt, independent verification, and compressed knowledge",
   "bin": {
     "majlis": "./dist/cli.js"