spacecode 2.14.4 → 2.14.6

package/dist/repl.mjs

@@ -4088,7 +4088,7 @@ var init_defaults_core_loop = __esm({
         body: `You are the core of a coding run. You direct; you never read files, search, run commands or write code. Workers do that.
 
 ## Tools (the only ones that exist)
-- edit_plan {tasks}: add tasks. Each task has id, role, goal (what the worker must do, concrete and self-contained: the worker sees only this, the files and the acceptance), files (a file belongs to one open task at a time), acceptance {kind: checks|rubric|none, detail} (a code task that the repo can check, with tests or a typecheck, uses kind checks so the repo's own typecheck and tests must pass; rubric and none skip them), budgetTokens, deadlineMs (milliseconds; a worker needs minutes to clone, work and check, so use 600000 for a code task; values are clamped to 600000-840000).
+- edit_plan {tasks}: add tasks. Each task has id, role, goal (what the worker must do, concrete and self-contained: the worker sees only this, the files and the acceptance), files (a file belongs to one open task at a time), dependsOn (optional list of task ids that must be accepted first; set it for any task that builds on another's work), acceptance {kind: checks|rubric|none, detail} (a code task that the repo can check, with tests or a typecheck, uses kind checks so the repo's own typecheck and tests must pass; rubric and none skip them), budgetTokens, deadlineMs (milliseconds; a worker needs minutes to clone, work and check, so use 600000 for a code task; values are clamped to 600000-840000).
 - dispatch_task {taskId}: start a pending task (at most 2 attempts per task).
 - accept_task {taskId}: accept a task in review. A checks task needs a passing verifier first.
 - drop_task {taskId, reason}: drop a task you no longer need.
@@ -4101,7 +4101,7 @@ var init_defaults_core_loop = __esm({
 - When the token budget is exhausted, package what is accepted and send the result.
 - Prefer small tasks with one owner per file. Stay inside the budget.
 - You are only woken by worker results and user messages, never by your own actions. So in the same answer as edit_plan, dispatch_task every task that can start now; a task left pending with nothing running stalls the run.
-- Tasks that must follow another (a test after its implementation) are dispatched after the earlier one is accepted; the later worker starts from the accepted work, so give it only its own files and do not repeat the earlier task's work.
+- Tasks that must follow another (a test after its implementation) name it in dependsOn and are dispatched only after the earlier one is accepted (a dispatch before that is refused, including while the earlier one is being retried); the later worker starts from the accepted work, so give it only its own files and do not repeat the earlier task's work. If a task it depends on is dropped or failed, the dependent can never start: drop it or replan it without that dependency.
 - Worker digests are already included in what you are shown; there is nothing more to read. Decide from them (accept_task, drop_task, or dispatch a retry).
 - Make every tool call of your answer at once; you do not see results between calls. Reply with tool calls only.`
       },
@@ -104840,6 +104840,32 @@ var init_worker_contract = __esm({
   }
 });
 
+// ../shared/src/core/check-baseline.ts
+function judgeAgainstBaseline(checks, baseline) {
+  const newChecks = {};
+  const baselineFailed = [];
+  for (const key of Object.keys(checks)) {
+    const code = checks[key];
+    if (code === 0) {
+      newChecks[key] = 0;
+      continue;
+    }
+    const baseCode = baseline[key];
+    if (typeof baseCode === "number" && baseCode !== 0) {
+      newChecks[key] = 0;
+      baselineFailed.push(key);
+    } else {
+      newChecks[key] = code;
+    }
+  }
+  return { checks: newChecks, baselineFailed };
+}
+var init_check_baseline = __esm({
+  "../shared/src/core/check-baseline.ts"() {
+    "use strict";
+  }
+});
+
 // ../shared/src/core/worker-pipeline.ts
 function agentFailure(run) {
   const tail = (run.stderr.trim() || run.stdout.trim()).slice(-300).replace(/\s+/g, " ");
@@ -104910,6 +104936,7 @@ async function runWorkerJob(deps, job) {
     const changedFiles = names.stdout.split("\n").map((l) => l.trim()).filter(Boolean);
     const diffLines = patch.stdout === "" ? 0 : patch.stdout.split("\n").length;
     const checks = {};
+    const baselineFailed = [];
     if (job.task.acceptance.kind === "checks") {
       const pkg = await deps.sandbox.exec(`cd ${shq(dir)} && cat package.json 2>/dev/null`);
       let scripts;
@@ -104931,6 +104958,31 @@ async function runWorkerJob(deps, job) {
           if (r.exitCode !== 0) deps.log("worker check failed", { ...ids, check: c.name, exitCode: r.exitCode, tail: redactSecrets(`${r.stdout}
 ${r.stderr}`.trim().slice(-300).replace(/\s+/g, " ")) });
         }
+        const failing = wanted.filter((c) => checks[c.name] !== 0);
+        if (deps.baseline && failing.length > 0) {
+          const bdir = `${dir}-baseline`;
+          try {
+            const added = await deps.sandbox.exec(`cd ${shq(dir)} && git worktree add --detach ${shq(bdir)} ${base}`);
+            if (added.exitCode !== 0) {
+              deps.log("baseline worktree failed", { ...ids, exitCode: added.exitCode });
+            } else {
+              await deps.baseline.prepare(bdir);
+              const baselineExit = {};
+              for (const c of failing) baselineExit[c.name] = (await deps.sandbox.exec(`cd ${shq(bdir)} && timeout ${slice()} ${c.cmd}`)).exitCode;
+              const judged = judgeAgainstBaseline(checks, baselineExit);
+              Object.assign(checks, judged.checks);
+              for (const name of judged.baselineFailed) {
+                baselineFailed.push(name);
+                deps.log("worker check also fails on the base revision", { ...ids, check: name, exitCode: baselineExit[name] });
+              }
+            }
+          } catch (e) {
+            if (e instanceof TypeError || e instanceof ReferenceError) throw e;
+            deps.log("baseline comparison failed", { ...ids, error: e instanceof Error ? e.message.slice(0, 120) : "unknown" });
+          } finally {
+            await deps.sandbox.exec(`cd ${shq(dir)} && git worktree remove --force ${shq(bdir)}`).catch((e) => deps.log("baseline worktree cleanup failed", { ...ids, error: e instanceof Error ? e.message.slice(0, 100) : "unknown" }));
+          }
+        }
       }
     }
     if (await deps.cancelled()) return "cancelled";
@@ -104947,7 +104999,7 @@ ${r.stderr}`.trim().slice(-300).replace(/\s+/g, " ")) });
       ...patch.stdout !== "" && !tooBig ? { patch: patch.stdout } : {},
       ...tooBig ? { reason: "patch exceeds the size limit" } : {},
       ...digest === "" ? { reason: run.exitCode !== 0 ? agentFailure(run) : noDigestReason(result?.text) } : {},
-      verify: { checks, changedFiles, diffLines }
+      verify: { checks, changedFiles, diffLines, ...baselineFailed.length > 0 ? { baselineFailed } : {} }
     });
     deps.log("worker reported", { ...ids, files: changedFiles.length, diffLines });
     return "reported";
@@ -104963,6 +105015,7 @@ var MAX_PATCH_BYTES, shq, clip;
 var init_worker_pipeline = __esm({
   "../shared/src/core/worker-pipeline.ts"() {
     "use strict";
+    init_check_baseline();
     init_worker_contract();
     MAX_PATCH_BYTES = 5e6;
     shq = (s) => `'${s.replace(/'/g, `'\\''`)}'`;
@@ -104982,6 +105035,7 @@ var init_core = __esm({
     init_verify();
     init_worker_contract();
     init_worker_pipeline();
+    init_check_baseline();
   }
 });
 
@@ -104991,6 +105045,15 @@ async function must(shell, cmd, what) {
   if (r.exitCode !== 0) throw new Error(`${what} failed: ${r.stderr.trim() || `exit ${r.exitCode}`}`);
   return r.stdout.trim();
 }
+async function linkDependencyFolders(shell, top, dir) {
+  const found = await must(shell, `cd ${q2(top)} && find . -maxdepth 4 -name node_modules -type d -prune -not -path './.git/*'`, "listing dependency folders");
+  for (const line of found.split("\n")) {
+    const rel = line.trim().replace(/^\.\//, "");
+    if (rel === "" || rel === ".") continue;
+    const parent = rel.includes("/") ? rel.slice(0, rel.lastIndexOf("/")) : ".";
+    await must(shell, `mkdir -p ${q2(`${dir}/${parent}`)} && { [ -e ${q2(`${dir}/${rel}`)} ] || ln -s ${q2(`${top}/${rel}`)} ${q2(`${dir}/${rel}`)}; }`, `linking ${rel}`);
+  }
+}
 async function createTaskWorktree(shell, cwd2, base, runId, taskId, attempt, warn = () => {
 }) {
   const top = await must(shell, `cd ${q2(cwd2)} && git rev-parse --show-toplevel`, "finding the repository root");
@@ -105000,13 +105063,7 @@ async function createTaskWorktree(shell, cwd2, base, runId, taskId, attempt, war
   if (sweep.exitCode !== 0) warn(`stale worktree sweep failed: ${sweep.stderr.trim().slice(0, 200) || `exit ${sweep.exitCode}`}`);
   const dir = `${root}/${safe(runId)}/wt-${safe(taskId)}-${attempt}`;
   await must(shell, `git -C ${q2(top)} worktree add --detach ${q2(dir)} ${q2(commit)}`, "creating the task worktree");
-  const found = await must(shell, `cd ${q2(top)} && find . -maxdepth 4 -name node_modules -type d -prune -not -path './.git/*'`, "listing dependency folders");
-  for (const line of found.split("\n")) {
-    const rel = line.trim().replace(/^\.\//, "");
-    if (rel === "" || rel === ".") continue;
-    const parent = rel.includes("/") ? rel.slice(0, rel.lastIndexOf("/")) : ".";
-    await must(shell, `mkdir -p ${q2(`${dir}/${parent}`)} && { [ -e ${q2(`${dir}/${rel}`)} ] || ln -s ${q2(`${top}/${rel}`)} ${q2(`${dir}/${rel}`)}; }`, `linking ${rel}`);
-  }
+  await linkDependencyFolders(shell, top, dir);
   return { top, dir };
 }
 async function removeTaskWorktree(shell, top, dir) {
@@ -105057,6 +105114,10 @@ function localPipelineDeps(d, job) {
     chown: () => "true",
     installDependencies: false,
     excludeFromDiff: ["**/node_modules"],
+    // A check that already fails on the base revision (this repo's `lint`: eslint is not installed) is re-run there, so it is not blamed on the task.
+    baseline: { prepare: async (bdir) => {
+      if (created) await linkDependencyFolders(shell, created.top, bdir);
+    } },
     basePatches: async () => {
       const res = await d.fetch(`${base}/base-patches`, { headers: auth, signal: d.signal });
       if (!res.ok) throw new Error(`base patches request failed (${res.status})`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "spacecode",
-  "version": "2.14.4",
+  "version": "2.14.6",
   "description": "SpaceCode - launch cloud dev environments with AI coding agents",
   "license": "MIT",
   "type": "commonjs",