haansi 0.1.51 → 0.1.53

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: "0.1.53",
       description: "Haansi CLI - Session collector and MCP server for Claude Code & OpenAI Codex",
       bin: {
         haansi: "./dist/haansi.js"
@@ -19000,6 +19000,14 @@ var init_exclusions = __esm({
           "join-key",
           "A canonical object id (github:<repo>#<n>), never text."
         ],
+        revert_evidence: [
+          "not-personal",
+          'Revert signal kinds with ledger-written details ("Revert #12", "branch revert-12-\u2026"); no PR text.'
+        ],
+        possible_revert: [
+          "not-personal",
+          "A candidate revert's canonical id, time and ledger-written signals; no PR text."
+        ],
         survival_14d: [
           "operational",
           "Closed CHECK vocabulary (held, failed, pending)."
@@ -19022,6 +19030,10 @@ var init_exclusions = __esm({
         ]
       }),
       ...exchangeExclusions(INTEL, "workstream_attempts", {
+        evidence: [
+          "not-personal",
+          "CI check names, commit ids, PR numbers and GitHub links a verdict rests on; no PR text."
+        ],
         process_id: [
           "join-key",
           "The workstream's process_registry.process_id, joined by equality."
@@ -19112,6 +19124,10 @@ var init_exclusions = __esm({
           "operational",
           "Closed CHECK vocabulary (held, reversed, revisited)."
         ],
+        outcome_evidence: [
+          "not-personal",
+          "PR keys (owner/repo#n), commit ids and timestamps the outcome rests on; no text."
+        ],
         change_keys: [
           "join-key",
           "owner/repo#n of the workstream's PRs, as in workstream_attempts."
@@ -19150,6 +19166,50 @@ var init_exclusions = __esm({
           "Counts by Actions job and step name, from CI config."
         ]
       }),
+      ...exchangeExclusions(INTEL, "repository_checks", {
+        repository: [
+          "join-key",
+          "owner/repo of the org's CI, as in change_outcomes."
+        ],
+        check_key: [
+          "join-key",
+          "Normalized CI workflow and job name, matched against captured jobs."
+        ],
+        workflow: ["not-personal", "A CI workflow's name, e.g. CI."],
+        job: ["not-personal", "A CI job's name, e.g. Unit tests."],
+        kind: [
+          "operational",
+          "Closed CHECK vocabulary (test, typecheck, build, lint, format, deploy, other)."
+        ]
+      }),
+      ...exchangeExclusions(INTEL, "revert_labels", {
+        repository: [
+          "join-key",
+          "owner/repo of the reverted PR, as in change_outcomes."
+        ],
+        reverted_by: [
+          "join-key",
+          "Canonical id of the reverting PR or commit (github:owner/repo#n, github:commit:<sha>)."
+        ],
+        label: [
+          "operational",
+          "Closed CHECK vocabulary (confirmed, rejected, cleared)."
+        ]
+      }),
+      ...exchangeExclusions(INTEL, "repository_replay_settings", {
+        repository: [
+          "join-key",
+          "owner/repo, as in replay_tasks and repository_checks."
+        ],
+        setup_command: [
+          "not-personal",
+          "A shell command an org admin typed for its CI (e.g. npm ci); run by the customer's own CI."
+        ],
+        test_command: [
+          "not-personal",
+          "A shell command an org admin typed for its CI (e.g. npm run test); run by the customer's own CI."
+        ]
+      }),
       ...exchangeExclusions(INTEL, "replay_results", {
         process_id: [
           "join-key",
@@ -85057,13 +85117,16 @@ __export(replay_exports, {
   runCommand: () => runCommand,
   runReplays: () => runReplays,
   safeTestFile: () => safeTestFile,
-  skipReason: () => skipReason
+  skipReason: () => skipReason,
+  taskCommands: () => taskCommands
 });
 function parseReplayArgs(argv) {
   const options = {
     familyId: "",
     agentCmd: "",
     testCmd: "",
+    setupCmd: null,
+    repoCommands: false,
     attempts: 3,
     arms: [...ARMS],
     limit: null,
@@ -85082,6 +85145,7 @@ 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();
@@ -85090,6 +85154,7 @@ function parseReplayArgs(argv) {
     else if (a === "--model") options.model = next();
     else if (a === "--run-ref") options.runRef = next();
     else if (a === "--dry-run") options.dryRun = true;
+    else if (a === "--repo-commands") options.repoCommands = true;
     else if (a === "--arms")
       options.arms = next().split(",").map((s) => s.trim()).filter(Boolean);
     else if (!sub && !a.startsWith("-")) sub = a;
@@ -85098,7 +85163,10 @@ function parseReplayArgs(argv) {
   if (!/^[0-9a-f-]{36}$/i.test(options.familyId))
     errors.push("--family <id> is required (the family's id).");
   if (!options.agentCmd) errors.push("--agent-cmd is required.");
-  if (!options.testCmd) errors.push("--test-cmd is required.");
+  if (!options.testCmd && !options.repoCommands)
+    errors.push(
+      "--test-cmd is required (or --repo-commands, to use the repository's from Haansi)."
+    );
   if (!Number.isInteger(options.attempts) || options.attempts < 1 || options.attempts > 20)
     errors.push("--attempts must be 1 to 20.");
   if (options.limit !== null && (!Number.isInteger(options.limit) || options.limit < 1))
@@ -85122,6 +85190,16 @@ function repositoryFromRemote(remote) {
 function safeTestFile(path2) {
   return path2.length > 0 && !path2.startsWith("/") && !path2.startsWith("-") && !path2.split("/").includes("..") && !/[\u0000-\u001f]/.test(path2);
 }
+function taskCommands(options, task) {
+  const repo = options.repoCommands ? task.repoCommands : null;
+  const setupCmd = options.setupCmd ?? repo?.setupCommand ?? null;
+  const testCmd = options.testCmd || repo?.testCommand || null;
+  return {
+    setupCmd,
+    testCmd,
+    fromHaansi: Boolean(repo) && (!options.setupCmd && setupCmd !== null || !options.testCmd && testCmd !== null)
+  };
+}
 function skipReason(task, repository) {
   if (!repository || task.repository.toLowerCase() !== repository)
     return `in ${task.repository}, not this repository`;
@@ -85173,6 +85251,25 @@ 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 +85280,10 @@ 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 +85297,21 @@ 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,
@@ -85260,6 +85370,24 @@ async function runReplays(options, deps) {
   const chosen = options.limit ? runnable.slice(0, options.limit) : runnable;
   const results = [];
   for (const task of chosen) {
+    const commands = taskCommands(options, task);
+    if (!commands.testCmd) {
+      skipped += 1;
+      deps.log(
+        `skip ${task.name ?? task.processId}: no test command (pass --test-cmd, or set one for ${task.repository} in Haansi)`
+      );
+      continue;
+    }
+    if (commands.fromHaansi) {
+      deps.log(
+        `\u2026 ${task.name ?? task.processId}: using ${task.repository}'s replay commands from Haansi (setup: ${commands.setupCmd ?? "none"}; test: ${commands.testCmd})`
+      );
+    }
+    const taskOptions = {
+      ...options,
+      setupCmd: commands.setupCmd,
+      testCmd: commands.testCmd
+    };
     if (!await ensureCommit(deps, task.baseSha) || !await ensureCommit(deps, task.targetSha)) {
       skipped += 1;
       deps.log(
@@ -85269,7 +85397,7 @@ async function runReplays(options, deps) {
     }
     for (const arm of arms) {
       try {
-        const result = await replayArm(deps, options, task, arm, playbook);
+        const result = await replayArm(deps, taskOptions, task, arm, playbook);
         results.push(result);
         deps.log(
           `${result.passed ? "pass" : "fail"} ${arm.padEnd(8)} ${task.name ?? task.processId} (${result.attempts}/${result.maxAttempts} attempts)`
@@ -85376,7 +85504,7 @@ Baseline: ${passed("baseline")}/${ran("baseline")} passed \xB7 With Haansi: ${pa
     (0, import_node_fs10.rmSync)(work, { recursive: true, force: true });
   }
 }
-var import_node_child_process4, import_node_fs10, import_node_os9, import_node_path9, SHA, ARMS, FEEDBACK_CHARS, USAGE;
+var import_node_child_process4, import_node_fs10, import_node_os9, import_node_path9, SHA, ARMS, FEEDBACK_CHARS, FAILURE_LINES, USAGE;
 var init_replay = __esm({
   "src/commands/replay.ts"() {
     "use strict";
@@ -85389,13 +85517,18 @@ var init_replay = __esm({
     SHA = /^[0-9a-f]{7,40}$/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]
+                         [--repo-commands]
+
+--repo-commands uses the setup and test commands an org admin set for the
+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 "$FAMILY_ID" \\
+  haansi replay run --family "$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": "0.1.53",
   "description": "Haansi CLI - Session collector and MCP server for Claude Code & OpenAI Codex",
   "bin": {
     "haansi": "./dist/haansi.js"