@workos/quickstudy 0.2.0 → 0.2.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@workos/quickstudy",
-  "version": "0.2.0",
+  "version": "0.2.1",
   "description": "An open-source, company-agnostic eval harness for measuring whether coding agents can complete real integration tasks, across an agent x feature x framework x surface matrix.",
   "license": "MIT",
   "type": "module",

package/src/build-info.generated.ts

@@ -9,7 +9,7 @@ export const EMBEDDED_BUILD_INFO: {
   // (the `x-release-please-version` marker anchors the replacement). It is
   // restored as-is after a binary build (scripts/build-binary.ts), so the
   // committed source `--version` stays in lockstep with package.json.
-  version: "0.2.0", // x-release-please-version
+  version: "0.2.1", // x-release-please-version
   revision: null,
   dirty: null,
   sourceHash: null,

package/src/isolation/docker.ts

@@ -246,8 +246,13 @@ export async function createAttemptContainer(opts: CreateAttemptContainerOptions
     execInWorkspace: async (cmd, execOpts = {}) => {
       const result = await runDocker(["exec", "-w", CONTAINER_WORKSPACE, id, ...cmd], execOpts);
       // Killing only the Docker client leaves the exec process alive in the
-      // daemon. Stop the whole attempt on timeout, including descendants.
-      if (result.timedOut) await runDocker(["kill", id], { timeoutMs: 10_000 });
+      // daemon. Stop the whole attempt on timeout, including descendants —
+      // then restart PID 1 (`sleep infinity`) so the dead attempt's workspace
+      // stays exec-able for the diff. Descendants do not come back with it.
+      if (result.timedOut) {
+        await runDocker(["kill", id], { timeoutMs: 10_000 });
+        await runDocker(["start", id], { timeoutMs: 10_000 });
+      }
       return result;
     },
     writeFile: async (path, content) => {

package/src/runner/execute.ts

@@ -826,8 +826,26 @@ async function runContainerAttemptV3(
             );
           }
         }
+        // `git add -A` stages new files so the diff captures them; diffing
+        // the recorded baseline SHA (not HEAD) also captures any commits the
+        // agent made itself.
+        const captureDiff = async (): Promise<void> => {
+          if (baselineSha === undefined) return;
+          await mustExec(container, ["git", "add", "-A"]);
+          const diff = (await mustExec(container, ["git", "diff", baselineSha])).stdout;
+          persisted.diffRef = await artifacts.writeDiff(runId, attemptId, redactSecrets(diff, secretEnv));
+        };
+        if (agentResult.timedOut || agentResult.exitCode !== 0) {
+          // The workspace is evidence of what the agent got done before it
+          // died; a failed capture must not mask the attempt's own outcome.
+          try {
+            await captureDiff();
+          } catch (err) {
+            (options.log ?? console.log)(`attempt ${attemptId}: diff unavailable after failure: ${err instanceof Error ? err.message : String(err)}`);
+          }
+        }
         if (agentResult.timedOut) {
-          await artifacts.writeError(runId, attemptId, `attempt timed out after ${args.attemptTimeoutMs}ms — container removed`);
+          await artifacts.writeError(runId, attemptId, `attempt timed out after ${args.attemptTimeoutMs}ms — container killed`);
           return { status: "incomplete", error: `attempt timed out after ${args.attemptTimeoutMs}ms` };
         }
         if (agentResult.exitCode !== 0) {
@@ -836,14 +854,7 @@ async function runContainerAttemptV3(
           return { status: "error", error: `agent command exited ${agentResult.exitCode}` };
         }
 
-        if (baselineSha !== undefined) {
-          // `git add -A` stages new files so the diff captures them; diffing
-          // the recorded baseline SHA (not HEAD) also captures any commits
-          // the agent made itself.
-          await mustExec(container, ["git", "add", "-A"]);
-          const diff = (await mustExec(container, ["git", "diff", baselineSha])).stdout;
-          persisted.diffRef = await artifacts.writeDiff(runId, attemptId, redactSecrets(diff, secretEnv));
-        }
+        await captureDiff();
 
         // Export the workspace BEFORE scoring, with the container still alive:
         // file checks read the export, ctx.exec reaches into the live sandbox.