paperclipai/paperclip · error · RunnerWorkflowInfrastructureError

live_provider_execution_failed

live_provider_execution_failed

Error message

live_provider_execution_failed

What it means

executeLiveRunnerWorkflow wraps a live (real provider) eval session; any error raised while creating, messaging, restoring, or shutting down the live session is captured, the runtime root is deleted, and rethrown as RunnerWorkflowInfrastructureError with code live_provider_execution_failed and retryable=true. This code marks the failure as infrastructure (provider/transport/process) rather than an assertion failure, so the failure message is the stringified underlying cause.

Source

Thrown at packages/paperclip-runner/src/eval/live-workflow-executor.ts:941

      unsubscribe();
      if (budgetInterrupt !== null) await budgetInterrupt;
      if (budgetInterruptError !== undefined) throw budgetInterruptError;
    }
  } catch (error) {
    infrastructureError = error;
  }

  const snapshot = session?.snapshot();
  if (session !== null) {
    try {
      await service.shutdown(session.id, "Runner live workflow eval complete");
    } catch (error) {
      infrastructureError ??= error;
    }
  }
  if (infrastructureError !== undefined) {
    await rm(runtimeRoot, { recursive: true, force: true });
    throw new RunnerWorkflowInfrastructureError(
      "live_provider_execution_failed",
      true,
      safeFailureMessage(infrastructureError),
    );
  }
  const calls = snapshot === undefined ? [] : observedCalls(snapshot);
  const expectedCalls = input.evalCase.assertions.requiredOperationIds ?? [];
  const taskState =
    snapshot === undefined
      ? null
      : ((
          JSON.parse(snapshot.mockState) as {
            tasks?: Array<{ id: string; status: string }>;
          }
        ).tasks?.find((task) => task.id === "task-1") ?? null);
  const terminalStatuses = turns.map((turn) => turn.status);
  const cancellationExpected = input.evalCase.id === "cancellation-permissions";
  const terminalOkay = cancellationExpected

View on GitHub (pinned to 01ad858492)

Solutions

1. Read the wrapped cause via safeFailureMessage in the error's message/detail to identify the underlying provider or transport failure.
2. Check provider credentials and environment variables required by the candidate's qualification.requiredEnvironment/profile before running the campaign.
3. Confirm network egress to the provider endpoint and that the requested model is enabled for the account.
4. Re-run: the error is flagged retryable=true, so transient provider outages and rate limits can be retried within maxAttempts.
5. If it reproduces, raise budget.maxLatencyMs/maxCostUsd for the candidate or run a single eval case to isolate whether startup, messaging, or shutdown fails.

// Before the live run, check provider reachability and env:
for (const envName of candidate.qualification.requiredEnvironment) {
  if (!process.env[envName]) throw new Error(`missing ${envName} for live provider ${candidate.provider}`);
}
// Optionally probe the provider endpoint health before starting the campaign.

function isInfrastructureError(e: unknown): e is RunnerWorkflowInfrastructureError {
  return e instanceof RunnerWorkflowInfrastructureError && e.code === "live_provider_execution_failed";
}

try {
  await executeLiveRunnerWorkflow(input);
} catch (error) {
  if (error instanceof RunnerWorkflowInfrastructureError && error.code === "live_provider_execution_failed" && error.retryable) {
    // back off and retry within candidate.budget.maxAttempts;
    // inspect error message for the wrapped provider/transport cause
  } else {
    throw error;
  }
}

When it happens

Trigger: Any throw inside the session lifecycle during the live run: CapabilityLiveSessionService failing to start the adapter process, sendMessage rejecting on provider API error/timeout (turnTimeoutMs exceeded), session.restore failing during the restart-recovery case, budget-interrupt errors thrown in the finally block, or service.shutdown rejecting after the run.

Common situations: Expired or missing provider API credentials; provider rate limits or 5xx; model name not available to the account; network egress blocked in CI; turn latency exceeding maxLatencyMs budget; adapter binary failing to launch (e.g. the trusted opencode binding errors); shutdown racing an already-dead session.

Related errors

• provider turn ended with status ${turn.status}
• PAPERCLIP_EVAL_MAX_CAMPAIGN_COST_USD must be a positive fini
• Runner live rotation requires a valid generated-at time
• schedule references unknown candidate ${entry.candidateId}
• live schedule coverage failed: ${JSON.stringify(coverage)}

AI-assisted analysis of paperclipai/paperclip@01ad858492 (2026-09-02). Data as JSON: /api/errors/023fa7a867eb62f4. Report an issue: GitHub.