haansi 0.1.56 → 0.1.58

package/dist/haansi.js

@@ -95,7 +95,7 @@ var require_package = __commonJS({
   "package.json"(exports2, module2) {
     module2.exports = {
       name: "haansi",
-      version: "0.1.56",
+      version: "0.1.58",
       description: "Haansi CLI - Session collector and MCP server for Claude Code & OpenAI Codex",
       bin: {
         haansi: "./dist/haansi.js"
@@ -19220,6 +19220,24 @@ var init_exclusions = __esm({
         model: ["operational", "The model name the replay ran with."],
         run_ref: ["not-personal", "A link to the customer's CI run."]
       }),
+      ...exchangeExclusions(INTEL, "replay_sessions", {
+        excluded_process_ids: [
+          "join-key",
+          "process_registry.process_id of the replayed workstream, matched by equality."
+        ],
+        excluded_family_ids: [
+          "join-key",
+          "problem_families ids holding the replayed workstream, matched by equality."
+        ]
+      }),
+      ...exchangeExclusions(INTEL, "replay_session_calls", {
+        server: ["operational", "The MCP server's name, e.g. haansi-skills."],
+        tool: ["operational", "The MCP tool's name, e.g. search_skills."],
+        returned_ids: [
+          "join-key",
+          "Ids of the solutions, skills and revisions a tool returned."
+        ]
+      }),
       ...exchangeExclusions(INTEL, "replay_task_results", {
         inputs_fingerprint: [
           "derived",
@@ -19228,7 +19246,15 @@ var init_exclusions = __esm({
         arm: ["operational", "Closed CHECK vocabulary (baseline, haansi)."],
         agent: ["operational", "The agent CLI's name, e.g. claude-code."],
         model: ["operational", "The model name the replay ran with."],
-        run_ref: ["not-personal", "A link to the customer's CI run."]
+        run_ref: ["not-personal", "A link to the customer's CI run."],
+        haansi_used: [
+          "operational",
+          "Closed CHECK vocabulary (used, partly, not_used)."
+        ],
+        haansi_judge_tag: [
+          "not-personal",
+          "The prompt version that judged it (replay-use-judge-vN)."
+        ]
       }),
       ...exchangeExclusions(INTEL, "replay_tasks", {
         process_id: [
@@ -19315,7 +19341,11 @@ var init_exclusions = __esm({
           "The CI's setup and test commands an org admin or CI owner typed, runtime versions, the runner image and the CLI version."
         ],
         calibration_runs: ["operational", "Exit codes of the test runs."],
-        calibration_run_ref: ["not-personal", "A link to the customer's CI run."]
+        calibration_run_ref: ["not-personal", "A link to the customer's CI run."],
+        review_state: [
+          "operational",
+          "Closed CHECK vocabulary (approved, excluded)."
+        ]
       }),
       ...exchangeExclusions(INTEL, "problem_family_members", {
         process_id: [
@@ -26583,6 +26613,12 @@ var init_table_registry = __esm({
       entry("intelligence", INTELLIGENCE_DB, "replay_task_results", "direct", {
         scopingColumn: "organization_id"
       }),
+      entry("intelligence", INTELLIGENCE_DB, "replay_sessions", "direct", {
+        scopingColumn: "organization_id"
+      }),
+      entry("intelligence", INTELLIGENCE_DB, "replay_session_calls", "direct", {
+        scopingColumn: "organization_id"
+      }),
       // A person's revert verdicts, the checks that validate a change and the
       // replay commands, per repository (outcome evidence v2 plan).
       entry("intelligence", INTELLIGENCE_DB, "revert_labels", "direct", {
@@ -77480,7 +77516,7 @@ var require_view = __commonJS({
     var dirname5 = path2.dirname;
     var basename4 = path2.basename;
     var extname = path2.extname;
-    var join17 = path2.join;
+    var join18 = path2.join;
     var resolve5 = path2.resolve;
     module2.exports = View;
     function View(name, options) {
@@ -77542,12 +77578,12 @@ var require_view = __commonJS({
     };
     View.prototype.resolve = function resolve6(dir, file2) {
       var ext = this.ext;
-      var path3 = join17(dir, file2);
+      var path3 = join18(dir, file2);
       var stat = tryStat(path3);
       if (stat && stat.isFile()) {
         return path3;
       }
-      path3 = join17(dir, basename4(file2, ext), "index" + ext);
+      path3 = join18(dir, basename4(file2, ext), "index" + ext);
       stat = tryStat(path3);
       if (stat && stat.isFile()) {
         return path3;
@@ -81192,7 +81228,7 @@ var require_send = __commonJS({
     var Stream = require("stream");
     var util2 = require("util");
     var extname = path2.extname;
-    var join17 = path2.join;
+    var join18 = path2.join;
     var normalize = path2.normalize;
     var resolve5 = path2.resolve;
     var sep2 = path2.sep;
@@ -81364,7 +81400,7 @@ var require_send = __commonJS({
           return res;
         }
         parts = path3.split(sep2);
-        path3 = normalize(join17(root, path3));
+        path3 = normalize(join18(root, path3));
       } else {
         if (UP_PATH_REGEXP.test(path3)) {
           debug('malicious path "%s"', path3);
@@ -81497,7 +81533,7 @@ var require_send = __commonJS({
           if (err2) return self.onStatError(err2);
           return self.error(404);
         }
-        var p = join17(path3, self._index[i]);
+        var p = join18(path3, self._index[i]);
         debug('stat "%s"', p);
         fs.stat(p, function(err3, stat) {
           if (err3) return next(err3);
@@ -83604,7 +83640,7 @@ function resolveToken3() {
   return null;
 }
 function triggerCollectInBackground() {
-  if (isCollecting) return;
+  if (isCollecting || REPLAY_SESSION) return;
   isCollecting = true;
   collect({}).then(() => {
     lastCollectTime = Date.now();
@@ -83621,7 +83657,8 @@ function maybeCollect() {
 async function apiGet2(path2, options) {
   const headers = {
     Authorization: `Bearer ${TOKEN}`,
-    ...options?.extraHeaders
+    ...options?.extraHeaders,
+    ...REPLAY_HEADERS
   };
   const response = await fetchWithTimeout(`${API_URL}/api/v1${path2}`, {
     headers
@@ -83636,7 +83673,8 @@ async function apiPost(path2, body, options) {
   const headers = {
     Authorization: `Bearer ${TOKEN}`,
     "Content-Type": "application/json",
-    ...options?.extraHeaders
+    ...options?.extraHeaders,
+    ...REPLAY_HEADERS
   };
   const response = await fetchWithTimeout(`${API_URL}/api/v1${path2}`, {
     method: "POST",
@@ -83681,7 +83719,7 @@ async function main3() {
   }
   warmupInBackground();
 }
-var import_dotenv3, import_path5, import_fs6, import_os5, DEFAULT_API_URL3, CREDENTIALS_FILE5, API_URL, AGENT_SOURCE, DEFAULT_SOURCE_TOOL, TOKEN, contextSchema, isCollecting, lastCollectTime, COLLECT_COOLDOWN_MS, mcpServer, server;
+var import_dotenv3, import_path5, import_fs6, import_os5, DEFAULT_API_URL3, CREDENTIALS_FILE5, API_URL, REPLAY_SESSION, REPLAY_HEADERS, AGENT_SOURCE, DEFAULT_SOURCE_TOOL, TOKEN, contextSchema, isCollecting, lastCollectTime, COLLECT_COOLDOWN_MS, mcpServer, server;
 var init_mcp_server2 = __esm({
   "../service-capture/agent-sessions/mcp-server.ts"() {
     "use strict";
@@ -83700,6 +83738,10 @@ var init_mcp_server2 = __esm({
     DEFAULT_API_URL3 = SERVICE_URL_DEFAULTS.haansiApiBaseUrl;
     CREDENTIALS_FILE5 = (0, import_path5.join)((0, import_os5.homedir)(), ".haansi", "credentials.json");
     API_URL = process.env.HAANSI_API_URL ?? DEFAULT_API_URL3;
+    REPLAY_SESSION = /^[0-9a-f-]{36}$/i.test(
+      process.env.HAANSI_REPLAY_SESSION?.trim() ?? ""
+    ) ? process.env.HAANSI_REPLAY_SESSION.trim() : null;
+    REPLAY_HEADERS = REPLAY_SESSION ? { "X-Haansi-Replay-Session": REPLAY_SESSION } : {};
     AGENT_SOURCE = process.env.HAANSI_AGENT_SOURCE?.toLowerCase();
     DEFAULT_SOURCE_TOOL = AGENT_SOURCE === "codex" ? "codex-mcp" : "claude-code-mcp";
     TOKEN = resolveToken3();
@@ -83920,6 +83962,13 @@ var init_mcp_server2 = __esm({
             };
           }
           case "save_solution": {
+            if (REPLAY_SESSION) {
+              return {
+                content: [
+                  { type: "text", text: "Not saved: nothing is recorded here." }
+                ]
+              };
+            }
             const {
               problem_description,
               solution_summary,
@@ -85562,16 +85611,125 @@ them all.`;
   }
 });
 
+// src/commands/replay-judge.ts
+function renderTemplate(template, vars) {
+  return template.replace(
+    /\{\{(\w+)\}\}/g,
+    (all, key) => key in vars ? vars[key] : all
+  );
+}
+async function changeOf(deps, dir, baseSha) {
+  await deps.git(["add", "--all", "--intent-to-add"], dir);
+  const diff = await deps.git(
+    ["diff", "--no-color", baseSha, "--", ".", ":(exclude).haansi-replay"],
+    dir
+  );
+  return diff.code === 0 ? clip(diff.output, MAX_CHANGE_CHARS) : "";
+}
+function parseJudgement(text, returnedIds) {
+  const match = /\{[\s\S]*\}/.exec(text);
+  if (!match) return null;
+  let v;
+  try {
+    v = JSON.parse(match[0]);
+  } catch {
+    return null;
+  }
+  if (!VERDICTS.includes(v.verdict)) return null;
+  const items = Array.isArray(v.used_items) ? new Set(
+    v.used_items.filter(
+      (id) => typeof id === "string" && returnedIds.has(id)
+    )
+  ) : /* @__PURE__ */ new Set();
+  const reason = typeof v.reason === "string" && v.reason.trim() ? v.reason.trim().slice(0, MAX_REASON) : null;
+  return { used: v.verdict, reason, usedItems: items.size };
+}
+async function judgeUse(deps, input) {
+  const returnedIds = new Set(input.calls.flatMap((c) => c.returnedIds));
+  if (input.calls.length === 0)
+    return {
+      used: "not_used",
+      reason: "The agent did not call Haansi.",
+      usedItems: 0,
+      judgeTag: null
+    };
+  if (returnedIds.size === 0)
+    return {
+      used: "not_used",
+      reason: "Haansi returned nothing.",
+      usedItems: 0,
+      judgeTag: null
+    };
+  if (!input.prompt)
+    return { used: null, reason: null, usedItems: null, judgeTag: null };
+  const tag = input.prompt.edited ? `${input.prompt.tag}+edited` : input.prompt.tag;
+  const dir = (0, import_node_path10.join)(deps.workDir(), `${input.taskId}-judge`);
+  const verdictFile = (0, import_node_path10.join)(dir, "verdict.json");
+  const promptFile = (0, import_node_path10.join)(dir, "prompt.md");
+  const configFile = (0, import_node_path10.join)(dir, "mcp.json");
+  const returned = input.calls.map((c) => c.returnedText ?? "").filter(Boolean).join("\n\n");
+  const prompt2 = [
+    input.prompt.system,
+    "",
+    renderTemplate(input.prompt.user, {
+      task: input.card,
+      returned: clip(returned || "(only ids came back)", MAX_RETURNED_CHARS),
+      change: input.change || "(no change)",
+      verdict_file: verdictFile
+    })
+  ].join("\n");
+  try {
+    deps.writeFile(promptFile, prompt2);
+    deps.writeFile(configFile, JSON.stringify({ mcpServers: {} }));
+    const ran = await deps.shell(
+      input.agentCmd,
+      dir,
+      {
+        HAANSI_REPLAY_PROMPT: prompt2,
+        HAANSI_REPLAY_PROMPT_FILE: promptFile,
+        HAANSI_REPLAY_MCP_CONFIG: configFile,
+        HAANSI_REPLAY_VERDICT_FILE: verdictFile
+      },
+      input.timeoutMs
+    );
+    const judged = parseJudgement(deps.readFile(verdictFile) ?? "", returnedIds) ?? parseJudgement(ran.output, returnedIds);
+    if (!judged) {
+      deps.log(`\u2026 ${input.label} [haansi]: the judge gave no verdict`);
+      return { used: null, reason: null, usedItems: null, judgeTag: tag };
+    }
+    return { ...judged, judgeTag: tag };
+  } finally {
+    deps.removeDir(dir);
+  }
+}
+var import_node_path10, MAX_CHANGE_CHARS, MAX_RETURNED_CHARS, MAX_REASON, VERDICTS, clip;
+var init_replay_judge = __esm({
+  "src/commands/replay-judge.ts"() {
+    "use strict";
+    import_node_path10 = require("node:path");
+    MAX_CHANGE_CHARS = 6e4;
+    MAX_RETURNED_CHARS = 6e4;
+    MAX_REASON = 300;
+    VERDICTS = ["used", "partly", "not_used"];
+    clip = (text, max2) => text.length > max2 ? `${text.slice(0, max2)}
+\u2026 (trimmed)` : text;
+  }
+});
+
 // src/commands/replay-tasks.ts
 var replay_tasks_exports = {};
 __export(replay_tasks_exports, {
+  HAANSI_LINE: () => HAANSI_LINE,
   areaOf: () => areaOf,
   cardPrompt: () => cardPrompt,
   changedFiles: () => changedFiles,
+  haansiText: () => haansiText,
+  haansiUseOf: () => haansiUseOf,
+  mcpConfig: () => mcpConfig,
   overlapOf: () => overlapOf,
   runStoredReplays: () => runStoredReplays
 });
-function cardPrompt(task, feedback) {
+function cardPrompt(task, feedback, arm = "baseline") {
   return [
     "# Task",
     "",
@@ -85582,10 +85740,40 @@ function cardPrompt(task, feedback) {
     task.card.acceptance,
     ...task.card.area ? ["", "## Part of the code", "", task.card.area] : [],
     "",
+    ...arm === "haansi" ? [HAANSI_LINE, ""] : [],
     ...feedback ? [FEEDBACK_HEADER, "", "```", feedback, "```", ""] : [],
     "Work in this repository as it is now. Do not look at later commits or other branches."
   ].join("\n");
 }
+function mcpConfig(arm, session) {
+  if (arm !== "haansi" || !session) return JSON.stringify({ mcpServers: {} });
+  const base = session.apiUrl.replace(/\/+$/, "");
+  return JSON.stringify(
+    {
+      mcpServers: {
+        "haansi-solutions": {
+          command: "haansi",
+          args: ["mcp-server"],
+          env: {
+            HAANSI_API_URL: base,
+            HAANSI_TOKEN: session.token,
+            HAANSI_REPLAY_SESSION: session.sessionId
+          }
+        },
+        "haansi-skills": {
+          type: "http",
+          url: `${base}/api/v1/memory/skills/mcp`,
+          headers: {
+            Authorization: `Bearer ${session.token}`,
+            "X-Haansi-Replay-Session": session.sessionId
+          }
+        }
+      }
+    },
+    null,
+    2
+  );
+}
 function areaOf(path2) {
   const dirs = path2.split("/").slice(0, -1);
   return dirs.length ? dirs.slice(0, 2).join("/") : ".";
@@ -85644,14 +85832,14 @@ async function runStoredReplays(options, deps) {
     (t) => states.get(t.taskId) === "valid" && t.testFiles.length > 0 && t.testFiles.every(safeTestFile) && SHA.test(t.baseSha) && SHA.test(t.targetSha)
   );
   const notValid = set2.tasks.length - valid.length;
-  let arms = options.arms;
-  if (arms.includes("haansi")) {
-    deps.log(
-      "The with-Haansi arm on stored tasks needs Haansi answering as of each task's start; running the baseline arm only."
-    );
-    arms = arms.filter((a) => a !== "haansi");
-  }
+  const arms = options.arms;
   const chosen = options.limit ? valid.slice(0, options.limit) : valid;
+  let judgePrompt;
+  const promptForJudge = async () => {
+    if (judgePrompt === void 0)
+      judgePrompt = await deps.getJudgePrompt().catch(() => null);
+    return judgePrompt;
+  };
   const results = [];
   for (const task of chosen) {
     const label = task.name ?? task.taskKey;
@@ -85669,18 +85857,47 @@ async function runStoredReplays(options, deps) {
     };
     for (const arm of arms) {
       let changed = [];
+      let change = "";
       try {
+        const session = arm === "haansi" ? {
+          ...await deps.openSession(task.taskId),
+          apiUrl: deps.apiUrl,
+          token: deps.token
+        } : null;
+        const configPath = `${deps.workDir()}/${task.taskId}-${arm}.mcp.json`;
+        deps.writeFile(configPath, mcpConfig(arm, session));
         const r = await replayArm(
           deps,
           { ...options, setupCmd, testCmd },
           asTask,
           arm,
           null,
-          (feedback) => cardPrompt(task, feedback),
+          (feedback) => cardPrompt(task, feedback, arm),
           async (dir) => {
             changed = await changedFiles(deps, dir, task.baseSha);
-          }
+            if (arm === "haansi")
+              change = await changeOf(deps, dir, task.baseSha);
+          },
+          { HAANSI_REPLAY_MCP_CONFIG: configPath }
         );
+        deps.removeFile(configPath);
+        let haansi = null;
+        if (session) {
+          const { calls } = await deps.getCalls(session.sessionId);
+          haansi = {
+            ...haansiUseOf(calls),
+            ...await judgeUse(deps, {
+              label,
+              taskId: task.taskId,
+              agentCmd: options.agentCmd,
+              timeoutMs: options.timeoutMs,
+              card: cardPrompt(task, null),
+              calls,
+              change,
+              prompt: await promptForJudge()
+            })
+          };
+        }
         const overlap = overlapOf(changed, task);
         results.push({
           taskId: task.taskId,
@@ -85691,10 +85908,11 @@ async function runStoredReplays(options, deps) {
           attempts: r.attempts,
           maxAttempts: r.maxAttempts,
           durationMs: r.durationMs,
-          overlap
+          overlap,
+          haansi
         });
         deps.log(
-          `${r.passed ? "pass" : "fail"} ${arm.padEnd(8)} ${label} (${r.attempts}/${r.maxAttempts} attempts; files ${pct(overlap.fileRecall)} of the target's, ${pct(overlap.filePrecision)} of its own on target)`
+          `${r.passed ? "pass" : "fail"} ${arm.padEnd(8)} ${label} (${r.attempts}/${r.maxAttempts} attempts; files ${pct(overlap.fileRecall)} of the target's, ${pct(overlap.filePrecision)} of its own on target${haansi ? `; ${haansiText(haansi)}` : ""})`
         );
       } catch (err2) {
         deps.log(
@@ -85721,14 +85939,28 @@ async function runStoredReplays(options, deps) {
   }
   return { results, notValid, recorded };
 }
-var FEEDBACK_HEADER, REPORT_CHUNK2, share, pct;
+function haansiUseOf(calls) {
+  return {
+    calls: calls.length,
+    returned: calls.reduce((n, c) => n + c.returnedIds.length, 0)
+  };
+}
+function haansiText(use) {
+  if (use.calls === 0) return "did not call Haansi";
+  const base = `called Haansi ${use.calls}\xD7, ${use.returned} item${use.returned === 1 ? "" : "s"} returned`;
+  if (!use.used) return base;
+  return use.used === "not_used" ? `${base}, not used` : `${base}, ${use.used === "partly" ? "partly " : ""}used${use.usedItems ? ` (${use.usedItems})` : ""}`;
+}
+var FEEDBACK_HEADER, REPORT_CHUNK2, HAANSI_LINE, share, pct;
 var init_replay_tasks = __esm({
   "src/commands/replay-tasks.ts"() {
     "use strict";
     init_replay_calibrate();
+    init_replay_judge();
     init_replay();
     FEEDBACK_HEADER = "Your previous attempt did not pass the tests. Their output ends with:";
     REPORT_CHUNK2 = 200;
+    HAANSI_LINE = "Haansi's MCP servers are available to you: they give access to your organization's knowledge (solutions to past problems, team skills, digests, guides and Exchange). Use them if they help.";
     share = (part, whole) => whole === 0 ? null : Math.round(part / whole * 1e3) / 1e3;
     pct = (v) => v === null ? "n/a" : `${Math.round(v * 100)}%`;
   }
@@ -85902,9 +86134,9 @@ async function cleanClone(deps, dir, sha, unreachable) {
 function cleanCheckout(deps, dir, task) {
   return cleanClone(deps, dir, task.baseSha, task.targetSha);
 }
-async function replayArm(deps, options, task, arm, playbook, promptOf = (feedback) => replayPrompt(task, arm, feedback), inspect) {
+async function replayArm(deps, options, task, arm, playbook, promptOf = (feedback) => replayPrompt(task, arm, feedback), inspect, agentEnv = {}) {
   const started = deps.now();
-  const dir = (0, import_node_path10.join)(
+  const dir = (0, import_node_path11.join)(
     deps.workDir(),
     `${task.processId.replace(/[^a-zA-Z0-9_-]/g, "")}-${arm}`
   );
@@ -85930,10 +86162,15 @@ async function replayArm(deps, options, task, arm, playbook, promptOf = (feedbac
       }
       deps.log(`\u2026 ${label}: set up in ${since(at)}`);
     }
-    const promptFile = (0, import_node_path10.join)(dir, ".haansi-replay", "prompt.md");
+    const promptFile = (0, import_node_path11.join)(dir, ".haansi-replay", "prompt.md");
+    let mcpConfigFile = agentEnv.HAANSI_REPLAY_MCP_CONFIG;
+    if (!mcpConfigFile) {
+      mcpConfigFile = `${dir}.mcp.json`;
+      deps.writeFile(mcpConfigFile, JSON.stringify({ mcpServers: {} }));
+    }
     if (arm === "haansi" && playbook)
       deps.writeFile(
-        (0, import_node_path10.join)(dir, ".haansi-replay", "playbook.md"),
+        (0, import_node_path11.join)(dir, ".haansi-replay", "playbook.md"),
         playbook.skillMd
       );
     let feedback = null;
@@ -85947,14 +86184,19 @@ async function replayArm(deps, options, task, arm, playbook, promptOf = (feedbac
       await deps.shell(
         options.agentCmd,
         dir,
-        { HAANSI_REPLAY_PROMPT: prompt2, HAANSI_REPLAY_PROMPT_FILE: promptFile },
+        {
+          ...agentEnv,
+          HAANSI_REPLAY_MCP_CONFIG: mcpConfigFile,
+          HAANSI_REPLAY_PROMPT: prompt2,
+          HAANSI_REPLAY_PROMPT_FILE: promptFile
+        },
         options.timeoutMs
       );
-      const own = new Map(files.map((f) => [f, deps.readFile((0, import_node_path10.join)(dir, f))]));
+      const own = new Map(files.map((f) => [f, deps.readFile((0, import_node_path11.join)(dir, f))]));
       for (const f of files) {
         const content = await deps.showFile(task.targetSha, f);
-        if (content === null) deps.removeFile((0, import_node_path10.join)(dir, f));
-        else deps.writeFile((0, import_node_path10.join)(dir, f), content);
+        if (content === null) deps.removeFile((0, import_node_path11.join)(dir, f));
+        else deps.writeFile((0, import_node_path11.join)(dir, f), content);
       }
       const testAt = deps.now();
       deps.log(`\u2026 ${label}: agent done in ${since(agentAt)}, grading`);
@@ -85984,8 +86226,8 @@ async function replayArm(deps, options, task, arm, playbook, promptOf = (feedbac
         };
       }
       for (const [file2, content] of own) {
-        if (content === null) deps.removeFile((0, import_node_path10.join)(dir, file2));
-        else deps.writeFile((0, import_node_path10.join)(dir, file2), content);
+        if (content === null) deps.removeFile((0, import_node_path11.join)(dir, file2));
+        else deps.writeFile((0, import_node_path11.join)(dir, file2), content);
       }
       feedback = test.output.slice(-FEEDBACK_CHARS);
     }
@@ -86001,6 +86243,7 @@ async function replayArm(deps, options, task, arm, playbook, promptOf = (feedbac
     };
   } finally {
     deps.removeDir(dir);
+    if (!agentEnv.HAANSI_REPLAY_MCP_CONFIG) deps.removeFile(`${dir}.mcp.json`);
   }
 }
 async function runReplays(options, deps) {
@@ -86134,7 +86377,7 @@ function localDeps(work) {
     repoDir: process.cwd(),
     workDir: () => work,
     writeFile: (path2, content) => {
-      (0, import_node_fs11.mkdirSync)((0, import_node_path10.dirname)(path2), { recursive: true });
+      (0, import_node_fs11.mkdirSync)((0, import_node_path11.dirname)(path2), { recursive: true });
       (0, import_node_fs11.writeFileSync)(path2, content);
     },
     readFile: (path2) => (0, import_node_fs11.existsSync)(path2) ? (0, import_node_fs11.readFileSync)(path2, "utf-8") : null,
@@ -86163,7 +86406,7 @@ async function replay(argv) {
     await replayStored(options, token, apiUrl2);
     return;
   }
-  const work = (0, import_node_fs11.mkdtempSync)((0, import_node_path10.join)((0, import_node_os10.tmpdir)(), "haansi-replay-"));
+  const work = (0, import_node_fs11.mkdtempSync)((0, import_node_path11.join)((0, import_node_os10.tmpdir)(), "haansi-replay-"));
   try {
     const { results, skipped, recorded } = await runReplays(options, {
       getTasks: async (familyId, cutoff) => (await apiGet3(
@@ -86200,7 +86443,7 @@ Baseline: ${passed("baseline")}/${ran("baseline")} passed \xB7 With Haansi: ${pa
 async function replayStored(options, token, apiUrl2) {
   const { runStoredReplays: runStoredReplays2 } = await Promise.resolve().then(() => (init_replay_tasks(), replay_tasks_exports));
   const { cliVersion: cliVersion2 } = await Promise.resolve().then(() => (init_replay_calibrate(), replay_calibrate_exports));
-  const work = (0, import_node_fs11.mkdtempSync)((0, import_node_path10.join)((0, import_node_os10.tmpdir)(), "haansi-replay-"));
+  const work = (0, import_node_fs11.mkdtempSync)((0, import_node_path11.join)((0, import_node_os10.tmpdir)(), "haansi-replay-"));
   try {
     const { results, notValid, recorded } = await runStoredReplays2(options, {
       ...localDeps(work),
@@ -86222,12 +86465,29 @@ async function replayStored(options, token, apiUrl2) {
         apiUrl2
       ),
       env: process.env,
-      cliVersion: cliVersion2()
+      cliVersion: cliVersion2(),
+      openSession: async (taskId) => await apiPost2(
+        `/memory/replays/tasks/${encodeURIComponent(taskId)}/sessions`,
+        {},
+        token,
+        apiUrl2
+      ),
+      getCalls: (sessionId) => apiGet3(
+        `/memory/replays/sessions/${encodeURIComponent(sessionId)}/calls?include=text`,
+        token,
+        apiUrl2
+      ),
+      getJudgePrompt: () => apiGet3("/memory/replays/prompts/use-judge", token, apiUrl2),
+      apiUrl: apiUrl2,
+      token
     });
-    const passed = results.filter((r) => r.passed).length;
+    const arm = (a) => {
+      const mine = results.filter((r) => r.arm === a);
+      return `${mine.filter((r) => r.passed).length}/${mine.length}`;
+    };
     console.log(
       `
-Baseline: ${passed}/${results.length} passed \xB7 ${notValid} task(s) not valid to replay`
+Baseline: ${arm("baseline")} passed \xB7 With Haansi: ${arm("haansi")} passed \xB7 ${notValid} task(s) not valid to replay`
     );
     console.log(
       recorded === null ? options.dryRun ? "Dry run: nothing reported." : "Nothing to report." : `Reported ${recorded} result(s) to Haansi.`
@@ -86239,14 +86499,14 @@ Baseline: ${passed}/${results.length} passed \xB7 ${notValid} task(s) not valid
     (0, import_node_fs11.rmSync)(work, { recursive: true, force: true });
   }
 }
-var import_node_child_process4, import_node_fs11, import_node_os10, import_node_path10, SHA, ARMS, FEEDBACK_CHARS, FAILURE_LINES2, REPLAY_BRANCH, USAGE2;
+var import_node_child_process4, import_node_fs11, import_node_os10, import_node_path11, SHA, ARMS, FEEDBACK_CHARS, FAILURE_LINES2, REPLAY_BRANCH, USAGE2;
 var init_replay = __esm({
   "src/commands/replay.ts"() {
     "use strict";
     import_node_child_process4 = require("node:child_process");
     import_node_fs11 = require("node:fs");
     import_node_os10 = require("node:os");
-    import_node_path10 = require("node:path");
+    import_node_path11 = require("node:path");
     init_paths();
     init_api2();
     SHA = /^[0-9a-f]{7,40}$/i;
@@ -86272,7 +86532,7 @@ repository in Haansi (Settings \u2192 Data sources) for any not given here.
 
 Run from a clone of the repository. Example (GitHub Actions):
   haansi replay run --family "$FAMILY_ID" --setup-cmd 'npm ci' \\
-    --agent-cmd 'claude -p "$HAANSI_REPLAY_PROMPT" --permission-mode acceptEdits' \\
+    --agent-cmd 'claude -p "$HAANSI_REPLAY_PROMPT" --permission-mode acceptEdits --mcp-config "$HAANSI_REPLAY_MCP_CONFIG"' \\
     --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": "0.1.56",
+  "version": "0.1.58",
   "description": "Haansi CLI - Session collector and MCP server for Claude Code & OpenAI Codex",
   "bin": {
     "haansi": "./dist/haansi.js"