haansi 0.1.51 → 0.1.52

package/dist/haansi.js

@@ -95,7 +95,7 @@ var require_package = __commonJS({
   "package.json"(exports2, module2) {
     module2.exports = {
       name: "haansi",
-      version: "0.1.51",
+      version: "0.1.52",
       description: "Haansi CLI - Session collector and MCP server for Claude Code & OpenAI Codex",
       bin: {
         haansi: "./dist/haansi.js"
@@ -85064,6 +85064,7 @@ function parseReplayArgs(argv) {
     familyId: "",
     agentCmd: "",
     testCmd: "",
+    setupCmd: null,
     attempts: 3,
     arms: [...ARMS],
     limit: null,
@@ -85082,6 +85083,7 @@ function parseReplayArgs(argv) {
     if (a === "--family") options.familyId = next();
     else if (a === "--agent-cmd") options.agentCmd = next();
     else if (a === "--test-cmd") options.testCmd = next();
+    else if (a === "--setup-cmd") options.setupCmd = next() || null;
     else if (a === "--attempts") options.attempts = Number(next());
     else if (a === "--limit") options.limit = Number(next());
     else if (a === "--cutoff") options.cutoff = next();
@@ -85173,6 +85175,25 @@ async function replayArm(deps, options, task, arm, playbook) {
   );
   if (add.code !== 0) throw new Error(`could not check out ${task.baseSha}`);
   try {
+    const label = `${task.name ?? task.processId} [${arm}]`;
+    const since = (from) => `${Math.round((deps.now() - from) / 1e3)}s`;
+    deps.log(`\u2026 ${label}: checked out ${task.baseSha.slice(0, 9)}`);
+    if (options.setupCmd) {
+      const at = deps.now();
+      deps.log(`\u2026 ${label}: setting up`);
+      const setup = await deps.shell(
+        options.setupCmd,
+        dir,
+        {},
+        options.timeoutMs
+      );
+      if (setup.code !== 0) {
+        throw new Error(
+          `setup failed after ${since(at)}: ${setup.output.trim().split("\n").slice(-3).join(" | ")}`
+        );
+      }
+      deps.log(`\u2026 ${label}: set up in ${since(at)}`);
+    }
     const promptFile = (0, import_node_path9.join)(dir, ".haansi-replay", "prompt.md");
     if (arm === "haansi" && playbook)
       deps.writeFile(
@@ -85183,6 +85204,10 @@ async function replayArm(deps, options, task, arm, playbook) {
     for (let attempt = 1; attempt <= options.attempts; attempt++) {
       const prompt2 = replayPrompt(task, arm, feedback);
       deps.writeFile(promptFile, prompt2);
+      const agentAt = deps.now();
+      deps.log(
+        `\u2026 ${label}: attempt ${attempt}/${options.attempts}, running the agent`
+      );
       await deps.shell(
         options.agentCmd,
         dir,
@@ -85196,12 +85221,21 @@ async function replayArm(deps, options, task, arm, playbook) {
       );
       if (checkout.code !== 0)
         throw new Error("could not check out the fix's tests");
+      const testAt = deps.now();
+      deps.log(`\u2026 ${label}: agent done in ${since(agentAt)}, grading`);
       const test = await deps.shell(
         options.testCmd,
         dir,
         { HAANSI_REPLAY_TEST_FILES: files.join(" ") },
         options.timeoutMs
       );
+      deps.log(
+        `\u2026 ${label}: tests ${test.code === 0 ? "passed" : "failed"} in ${since(testAt)}`
+      );
+      if (test.code !== 0) {
+        for (const line of test.output.trim().split("\n").slice(-FAILURE_LINES))
+          deps.log(`    ${line}`);
+      }
       if (test.code === 0) {
         return {
           processId: task.processId,
@@ -85376,7 +85410,7 @@ Baseline: ${passed("baseline")}/${ran("baseline")} passed \xB7 With Haansi: ${pa
     (0, import_node_fs10.rmSync)(work, { recursive: true, force: true });
   }
 }
-var import_node_child_process4, import_node_fs10, import_node_os9, import_node_path9, SHA, ARMS, FEEDBACK_CHARS, USAGE;
+var import_node_child_process4, import_node_fs10, import_node_os9, import_node_path9, SHA, ARMS, FEEDBACK_CHARS, FAILURE_LINES, USAGE;
 var init_replay = __esm({
   "src/commands/replay.ts"() {
     "use strict";
@@ -85389,13 +85423,14 @@ var init_replay = __esm({
     SHA = /^[0-9a-f]{7,40}$/i;
     ARMS = ["baseline", "haansi"];
     FEEDBACK_CHARS = 6e3;
+    FAILURE_LINES = 8;
     USAGE = `Usage: haansi replay run --family <id> --agent-cmd "<cmd>" --test-cmd "<cmd>"
-                         [--attempts 3] [--arms baseline,haansi] [--limit n]
+                         [--setup-cmd "<cmd>"] [--attempts 3] [--arms baseline,haansi] [--limit n]
                          [--cutoff <date>] [--timeout-min 30]
                          [--agent-name <name>] [--model <name>] [--run-ref <url>] [--dry-run]
 
 Run from a clone of the repository. Example (GitHub Actions):
-  haansi replay run --family "$FAMILY_ID" \\
+  haansi replay run --family "$FAMILY_ID" --setup-cmd 'npm ci' \\
     --agent-cmd 'claude -p "$HAANSI_REPLAY_PROMPT" --permission-mode acceptEdits' \\
     --test-cmd 'npx vitest run $HAANSI_REPLAY_TEST_FILES' \\
     --agent-name claude-code --model claude-sonnet-5-5`;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "haansi",
-  "version": "0.1.51",
+  "version": "0.1.52",
   "description": "Haansi CLI - Session collector and MCP server for Claude Code & OpenAI Codex",
   "bin": {
     "haansi": "./dist/haansi.js"