fix(evals): coalesce null output for Braintrust reporter and guard log() throws

.changeset/strong-owls-hug.md

@@ -0,0 +1,11 @@
+---
+"eve": patch
+---
+
+fix(evals): coalesce null output for Braintrust reporter and guard against log() throws (#1405)
+
+A no-turn eval (e.g. schedule-dispatch + DB assertions) legitimately produces `result.output === null` per the eval API's own derivation. The Braintrust SDK rejects null/undefined output (`"output must be specified"`), and the throw escaped `onEvalComplete`, killing the entire `eve eval` run — remaining evals never executed and no artifacts were written.
+
+Two fixes:
+1. Coalesce `output: result.result.output ?? ""` so null is sent as an empty string.
+2. Wrap `experiment.log()` in try/catch so any reporter throw is logged but does not abort the run.

packages/eve/src/evals/runner/reporters/braintrust.test.ts

@@ -199,4 +199,60 @@ describe("Braintrust", () => {
       }),
     );
   });
+
+  it("coalesces null output to empty string for no-turn evals (#1405)", async () => {
+    const reporter = Braintrust(makeConfig());
+    await reporter.onRunStart([makeEval()], makeTarget());
+
+    reporter.onEvalComplete(
+      makeEvalResult({
+        result: {
+          // A no-turn eval (schedule-dispatch + DB assertions) produces
+          // output === null per the eval API's own derivation.
+          output: null,
+          finalMessage: null,
+          status: "completed",
+          events: [],
+          derived: {
+            toolCalls: [],
+            toolCallCount: 0,
+            subagentCalls: [],
+            subagentCallCount: 0,
+            inputRequests: [],
+            parked: false,
+            messageCount: 0,
+            reasoningBlockCount: 0,
+          },
+          sessionId: "session-456",
+        },
+        verdict: "passed",
+      }),
+    );
+
+    // Braintrust's SDK rejects null output ("output must be specified").
+    // The reporter coalesces to "" so the run doesn't crash.
+    expect(braintrustMocks.log).toHaveBeenCalledWith(
+      expect.objectContaining({
+        output: "",
+      }),
+    );
+  });
+
+  it("survives a log() throw without aborting the run (#1405)", async () => {
+    const reporter = Braintrust(makeConfig());
+    await reporter.onRunStart([makeEval()], makeTarget());
+
+    // Simulate Braintrust SDK throwing on log — the reporter must catch
+    // so the remaining evals still execute.
+    braintrustMocks.log.mockImplementationOnce(() => {
+      throw new Error("output must be specified");
+    });
+
+    // Should NOT throw
+    expect(() => reporter.onEvalComplete(makeEvalResult())).not.toThrow();
+
+    // A second call (the next eval) should still reach log
+    reporter.onEvalComplete(makeEvalResult());
+    expect(braintrustMocks.log).toHaveBeenCalledTimes(2);
+  });
 });

packages/eve/src/evals/runner/reporters/braintrust.ts

@@ -172,16 +172,27 @@ class BraintrustReporter implements EvalReporter {
       reasoningBlockCount: result.result.derived.reasoningBlockCount,
     };
 
-    this.#experiment.log({
-      id: result.id,
-      input: evaluation?.description ?? "",
-      output: result.result.output,
-      error: result.error ?? undefined,
-      scores,
-      metadata,
-      metrics,
-      tags: evaluation?.tags ? [...evaluation.tags] : undefined,
-    });
+    try {
+      this.#experiment.log({
+        id: result.id,
+        input: evaluation?.description ?? "",
+        // Braintrust's SDK rejects null/undefined output ("output must be
+        // specified"). A no-turn eval (e.g. schedule-dispatch + DB assertions)
+        // legitimately produces output === null per the eval API's own
+        // derivation. Coalesce to an empty string so the reporter doesn't
+        // crash the entire eval run. See #1405.
+        output: result.result.output ?? "",
+        error: result.error ?? undefined,
+        scores,
+        metadata,
+        metrics,
+        tags: evaluation?.tags ? [...evaluation.tags] : undefined,
+      });
+    } catch (error) {
+      // A reporter throw in onEvalComplete must not abort the run and
+      // drop the remaining evals + artifacts. Log and continue. See #1405.
+      console.error(`Braintrust reporter: failed to log eval "${result.id}":`, error);
+    }
   }
 
   async onRunComplete(_summary: EveEvalRunSummary): Promise<void> {