haansi 0.1.56 → 0.1.57

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: "0.1.57",
       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",
@@ -19315,7 +19333,11 @@ 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 +26605,12 @@ 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", {
@@ -83604,7 +83632,7 @@ function resolveToken3() {
   return null;
 }
 function triggerCollectInBackground() {
-  if (isCollecting) return;
+  if (isCollecting || REPLAY_SESSION) return;
   isCollecting = true;
   collect({}).then(() => {
     lastCollectTime = Date.now();
@@ -83621,7 +83649,8 @@ 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 +83665,8 @@ 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${path2}`, {
     method: "POST",
@@ -83681,7 +83711,7 @@ async function main3() {
   }
   warmupInBackground();
 }
-var import_dotenv3, import_path5, import_fs6, import_os5, DEFAULT_API_URL3, CREDENTIALS_FILE5, API_URL, AGENT_SOURCE, DEFAULT_SOURCE_TOOL, TOKEN, contextSchema, isCollecting, lastCollectTime, COLLECT_COOLDOWN_MS, mcpServer, server;
+var import_dotenv3, import_path5, import_fs6, import_os5, DEFAULT_API_URL3, CREDENTIALS_FILE5, API_URL, REPLAY_SESSION, REPLAY_HEADERS, AGENT_SOURCE, DEFAULT_SOURCE_TOOL, TOKEN, contextSchema, isCollecting, lastCollectTime, COLLECT_COOLDOWN_MS, mcpServer, server;
 var init_mcp_server2 = __esm({
   "../service-capture/agent-sessions/mcp-server.ts"() {
     "use strict";
@@ -83700,6 +83730,10 @@ var init_mcp_server2 = __esm({
     DEFAULT_API_URL3 = SERVICE_URL_DEFAULTS.haansiApiBaseUrl;
     CREDENTIALS_FILE5 = (0, import_path5.join)((0, import_os5.homedir)(), ".haansi", "credentials.json");
     API_URL = process.env.HAANSI_API_URL ?? DEFAULT_API_URL3;
+    REPLAY_SESSION = /^[0-9a-f-]{36}$/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 +83954,13 @@ 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,
@@ -85565,13 +85606,17 @@ them all.`;
 // 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 +85627,40 @@ 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 ${session.token}`,
+            "X-Haansi-Replay-Session": session.sessionId
+          }
+        }
+      }
+    },
+    null,
+    2
+  );
+}
 function areaOf(path2) {
   const dirs = path2.split("/").slice(0, -1);
   return dirs.length ? dirs.slice(0, 2).join("/") : ".";
@@ -85644,13 +85719,7 @@ async function runStoredReplays(options, deps) {
     (t) => states.get(t.taskId) === "valid" && t.testFiles.length > 0 && t.testFiles.every(safeTestFile) && SHA.test(t.baseSha) && SHA.test(t.targetSha)
   );
   const notValid = set2.tasks.length - valid.length;
-  let arms = options.arms;
-  if (arms.includes("haansi")) {
-    deps.log(
-      "The with-Haansi arm on stored tasks needs Haansi answering as of each task's start; running the baseline arm only."
-    );
-    arms = arms.filter((a) => a !== "haansi");
-  }
+  const arms = options.arms;
   const chosen = options.limit ? valid.slice(0, options.limit) : valid;
   const results = [];
   for (const task of chosen) {
@@ -85670,17 +85739,27 @@ async function runStoredReplays(options, deps) {
     for (const arm of arms) {
       let changed = [];
       try {
+        const session = arm === "haansi" ? {
+          ...await deps.openSession(task.taskId),
+          apiUrl: deps.apiUrl,
+          token: deps.token
+        } : null;
+        const configPath = `${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);
-          }
+          },
+          { HAANSI_REPLAY_MCP_CONFIG: configPath }
         );
+        deps.removeFile(configPath);
+        const haansi = session ? haansiUseOf((await deps.getCalls(session.sessionId)).calls) : null;
         const overlap = overlapOf(changed, task);
         results.push({
           taskId: task.taskId,
@@ -85691,10 +85770,11 @@ 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(
@@ -85721,7 +85801,16 @@ 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) {
+  return use.calls === 0 ? "did not call Haansi" : `called Haansi ${use.calls}\xD7, ${use.returned} item${use.returned === 1 ? "" : "s"} returned`;
+}
+var FEEDBACK_HEADER, REPORT_CHUNK2, HAANSI_LINE, share, pct;
 var init_replay_tasks = __esm({
   "src/commands/replay-tasks.ts"() {
     "use strict";
@@ -85729,6 +85818,7 @@ var init_replay_tasks = __esm({
     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)}%`;
   }
@@ -85902,7 +85992,7 @@ 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)(
     deps.workDir(),
@@ -85931,6 +86021,11 @@ async function replayArm(deps, options, task, arm, playbook, promptOf = (feedbac
       deps.log(`\u2026 ${label}: set up in ${since(at)}`);
     }
     const promptFile = (0, import_node_path10.join)(dir, ".haansi-replay", "prompt.md");
+    let mcpConfigFile = agentEnv.HAANSI_REPLAY_MCP_CONFIG;
+    if (!mcpConfigFile) {
+      mcpConfigFile = `${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"),
@@ -85947,7 +86042,12 @@ 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))]));
@@ -86001,6 +86101,7 @@ async function replayArm(deps, options, task, arm, playbook, promptOf = (feedbac
     };
   } finally {
     deps.removeDir(dir);
+    if (!agentEnv.HAANSI_REPLAY_MCP_CONFIG) deps.removeFile(`${dir}.mcp.json`);
   }
 }
 async function runReplays(options, deps) {
@@ -86222,12 +86323,28 @@ async function replayStored(options, token, apiUrl2) {
         apiUrl2
       ),
       env: process.env,
-      cliVersion: cliVersion2()
+      cliVersion: cliVersion2(),
+      openSession: async (taskId) => await apiPost2(
+        `/memory/replays/tasks/${encodeURIComponent(taskId)}/sessions`,
+        {},
+        token,
+        apiUrl2
+      ),
+      getCalls: (sessionId) => apiGet3(
+        `/memory/replays/sessions/${encodeURIComponent(sessionId)}/calls`,
+        token,
+        apiUrl2
+      ),
+      apiUrl: apiUrl2,
+      token
     });
-    const passed = results.filter((r) => r.passed).length;
+    const arm = (a) => {
+      const mine = results.filter((r) => r.arm === a);
+      return `${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.`
@@ -86272,7 +86389,7 @@ 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": "0.1.57",
   "description": "Haansi CLI - Session collector and MCP server for Claude Code & OpenAI Codex",
   "bin": {
     "haansi": "./dist/haansi.js"